Soft sensor application in the pulp and paper industry_ass…

Document Sample
Soft sensor application in the pulp and paper industry_ass… Powered By Docstoc
					Report – CETC 2005-100 (TR)
   THE APPLICATION OF SOFT SENSORS IN THE PULP AND PAPER
    AND CEMENT MANUFACTURING SECTORS FOR PROCESS AND
             ENERGY PERFORMANCE IMPROVEMENT

      OPPORTUNITY ANALYSIS AND TECHNOLOGY ASSESSMENT




                                             Prepared by:

                                 Marc Champagne, Effective Assets Inc

                                  Mouloud Amazouz and Radu Platon
                              Canmet Energy Technology Centre - Varennes




                                              June 2005




Report – CETC 2005-100 (TR)
                                 ACKNOWLEDGEMENT

The authors acknowledge the assistance from end-user companies, technology providers,
universities and R&D organisations. The experts from these organisations as outlined in the
annex are also credited for theirs expert opinions.




                                        DISCLAMER

This report is distributed for informational purposes and does not necessarily reflect the views of
the Government of Canada nor constitute an endorsement of any commercial product or person.
Neither Canada nor its ministers, officers, employees or agents makes any warranty in respect to
this report or assumes any liability arising out of this report.




Report – CETC 2005-100 (TR)                    i                                        June 2005
Executive Summary
A soft sensor is the correlation from various raw data sources to create a new source of useful
information. It is an empirical model that infers process state and product quality variables
that are difficult to measure on-line (composition, melt index, molecular distribution, etc.)
from readily available process measurements (temperature, pressure, flow, etc.)

Canadian industry has begun to accept soft sensor technology as a useful tool in reducing
their energy consumption, operating costs, environmental impact and in improving the final
product quality.

The manufacturing industry usually faces a problem of lack of real time measurement of
product and process variables. This on-line unavailability of critical process variables can lead
to out of specifications production. A soft sensor provides on-line, accurate estimates of these
variables, eliminating additional energy and production cost associated with out of
specifications production. Soft sensors can also play a significant role in more complex
systems used for process optimization, such as fault and diagnosis systems, and control
systems.

In many cases an infrequent process sample (once per hour, once per 8 hour shift, once per
day) which may not be representative of the process is taken to the laboratory and depending
on the reference standard being used in the laboratory, the final result will have typical delays
of between 20 minutes to 1 week . In many cases because of the whether the sample is
representative as well as reference standard being used, many companies regardless of type of
the industry will need a minimum of 3 laboratory results before taking action. For example, in
the pulp and paper industry, the Kappa number represents the hallmark of the quality of the
paper, being related to the lignin content remaining in the pulp. For a Kappa number which is
taken once an hour, with a laboratory delay of 20 minutes, the pulp mill could easily produce
3 hours off quality product before action can be taken. This increases the energy usage as well
as creates a greater environmental impact, since more waste is created which needs to be
disposed. If the final product cannot be sold due to its marginal performance, then this creates
the hidden factory, since often both industries will have to reprocess the material which
creates an additional cost due to more energy and chemicals consumption and additional labor
costs.

The major impact of soft sensors will be increasing the productivity of Canadian industries,
which will result in important energy savings. A good example of this is the reduction of the
specific energy in a thermo-mechanical pulp (TMP) refiner (the pulp is produced in a thermo-
mechanical process where wood particles are softened by steam before entering a pressurized
refiner). Using an Ontario based TMP operation with current electricity prices, a 5%
reduction in the specific energy consumption to produce TMP pulp can achieve an electrical
savings approximately $1,200,000/year. In a typical 600 ton per day operation, a 1% yield

Report – CETC 2005-100 (TR)                    ii                                        June 2005
increase provides annual energy savings of approximately of $60,000/year; however due to
this higher yield, the wood costs are reduced by over $1,000,000/year.

For pulp produced by cooking wood chips in pressure vessels in the presence of a soda liquor
– also known as Kraft pulping - if we assume a 50% acceptance rate on 30 Kraft and paper
mills in Canada, the potential energy reduction of purchased green house gases based fuels
will be between 500,000 and 1,300,000 GJ/year. The potential electrical energy reduction for
15 BCTMP pulp (bleached chemi-thermomechanical pulp) and TMP based newsprint mills at
a 50% acceptance rate is between 270,000 and 675,000 GJ/year. For the Canadian cement
industry, the potential energy reduction based on a 20% acceptance rate is between 26,000
and 63,000 GJ/year.

The potential economic impact of applying a soft sensor based control package for both Pulp
and Paper and Cement Industry is a cost reduction between $75 and 135 million dollars per
year if it was 100% successful for every mill.

In Canada there are two universities who have a strong background in research and applying
in industrial data analysis and applying it to industrial problems, McMaster University and
University of Alberta. Dr. John McGregor leads this research at McMaster University and Dr.
Sirish Shah leads this research at University of Alberta.

Soft sensor technology is now being considered or being adopted with a number of pulp and
paper companies in Canada; e.g. Tembec Inc. and Abitibi-Consolidated Inc. These types of
soft sensors include the near infra-red based soft sensors that the Pulp and Paper Research
Institute of Canada are developing to predict Kappa and wood moisture. The cement industry
is much more resistant in applying this technology in their processes.

The University of Toronto Pulp and Paper Centre and l’École Polytechnique in Montreal have
recently begun work in applying soft sensors for the pulp and paper industry. Pulp and Paper
Research Institute of Canada has also have spent years in developing new soft sensors for
chemical pulping segment. A key Canadian player of soft sensors is Matrikon Inc. based in
Edmonton, Alberta. Other players of soft sensor technology include Pacific Simulation, a
division of Metso Automation based out Moscow, Idaho, USA, Umetrics Inc. based out of
Kinnelon, New Jersey, USA, and GENSYM based out of Burlington, Massachusetts, USA,
and IETek based out of Tacoma, Washington, USA.

The interest of applying soft sensors to Pulp and Paper and Cement industries has to be
further accelerated. Demonstration projects should be developed to confirm and document the
economic impact of the soft sensors as well as supporting universities and research centres to
perform more research in developing easy off the shelf applications for all Canadian
Industries.



Report – CETC 2005-100 (TR)                   iii                                     June 2005
Sommaire exécutif
Un ‘’soft sensor’’ est un modèle qui utilise des mesures de procédé disponibles (telles que
température, débit, etc.) afin d’estimer la valeur des différentes données de procédé et de la
qualité du produit final qui sont difficiles à obtenir en temps réel (telle que la composition,
distribution moléculaire, etc).

L’industrie canadienne a commencé à accepter la technologie ‘’soft sensor’’ comme outil de
réduction de la consommation énergétique, des coûts d’opération, de l’impact
environnemental et d’amélioration de la qualité du produit final.

Un des plus grands problèmes des industries manufacturières, incluant celles de pâte et
papier, et ciment, est l’absence des mesures en temps réel des paramètres critiques de procédé
et qualité du produit. Cette déficience peut entraîner une production de moindre qualité ou de
qualité inacceptable. Un ‘’soft sensor’’ peut estimer, en temps réel, la valeur de ces
paramètres critiques, éliminant ainsi les coûts énergétiques et de production associés à une
production hors grade. Un ‘’soft sensor’’ peut aussi jouer un rôle significatif à titre de
composante individuelle d’un autre système plus complexe d’optimisation de procédé.

Dans la plupart des cas, le résultat d’un test de laboratoire afin de déterminer la qualité du
produit final peut durer de 20 minutes à une semaine. Ceci est dû principalement à une
fréquence d’échantillonnage inadéquate et au délai requis afin d’effectuer le test et obtenir le
résultat. Dans plusieurs cas, cet échantillon peut ne pas être représentatif de la production en
cours. En conséquence, dans la plupart des procédés industriels, au moins trois analyses de
qualité sont effectuées avant qu’une une décision concernant la production en cours soit prise.
Par exemple, dans l’industrie de pâtes et papiers, la qualité du papier peut être estimée à l’aide
de l’indice Kappa, une mesure des composantes non cellulosiques de la pâte. Si cet indice est
déterminé sur une base horaire, avec un délai de laboratoire de 20 minutes, l’usine peut
facilement produire pendant trois heures de la pâte hors normes. Ceci entraîne une
augmentation de la consommation énergétique, de la consommation des matières premières,
des déversements des effluents des usines et de coûts de production.

L’impact majeur des ‘’soft sensors’’ sera l’augmentation de la productivité des industries
canadiennes, entraînant ainsi des importantes diminutions des coûts de fabrication, incluant
ceux de l’énergie consommée et des effets sur l’environnement. Un bon exemple est la
réduction de la consommation énergétique d’une usine de pâte TMP – pâte thermomécanique,
un procédé de fabrication de pâte à partir des copeaux de bois traités à la vapeur avant d’être
envoyés au raffineur. Dans une usine TMP typique d’Ontario, en considérant les coûts actuels
de l’électricité, une réduction de 5% de la consommation énergétique peut entraîner des
économies annuelles de l’ordre de $1,200,000. Pour une capacité de production de 600 tonnes
métriques par jour de pâte thermomécanique, une augmentation de 1% de la productivité



Report – CETC 2005-100 (TR)                     iv                                        June 2005
entraîne des économies énergétiques annuelles de $60,000. Les économies au niveau des
matières premières seront encore plus significatives, dépassant $1,000,000.

Dans les cas des usines qui produisent des papiers à partir de pâte de bois fabriquée selon un
procédé au sulfate – procédé Kraft – le potentiel de réduction des émissions des gaz à effet de
serre se situe entre 500 000 et 1 300 000 GJ/année, si la technologie ‘’soft sensor’’ est
implantée dans la moitie des 30 usines Kraft situées au Canada.

Le potentiel de réduction de la consommation électrique des usines de pâte thermomécanique
et pâte chimique se situent entre 270 000 et 675 000 GJ/année si cette technologie est
implémentée dans la moitie des usines. Pour l’industrie du ciment, en considérant un taux
d’implantation de 20%, le potentiel d’économie énergétique se situe entre 26 000 et 63 000
GJ/année.

L’implémentation d’un système de contrôle de procédé basé sur la technologie ‘’soft sensor’’
dans toutes les usines des pâtes et papiers et ciment au Canada aura un effet économique
significatif, entraînant une réduction de coûts de production de 75 à 130 millions de dollars
par année.

Présentement, au Canada, il y a deux universités qui se spécialisent dans l’analyse des
données de procédé et son application industrielle, l’université McMaster, sous la direction de
Dr. John McGregor et l’université d’Alberta, sous la direction de DR. Sirish Shah. Le Centre
de pâtes et papiers de l’Université de Toronto et L’École Polytechnique de Montréal ont
récemment commencé à concentrer leurs efforts sur l’application de SS dans l’industrie.
L’institut Paprican développe des nouvelles technologies ‘’soft sensor’’ applicables au
procédé de fabrication de la pâte chimique.

La technologie ‘’soft sensor’’ est présentement implémentée ou à l’étude d’implémentation
dans quelques usines canadiennes de pâtes et papiers comme celles de Tembec Inc. et
d’Abitibi-Consolidated, par exemple. Un de ces ‘’soft sensors’’ est le système optique
développé par l’Institut de recherche sur les pâtes et papiers Papricarn, utilisé à déterminer
l’indice Kappa et la teneur en humidité du bois. L’industrie du ciment est cependant moins
réceptive à l’idée de l’utilisation de cette technologie dans leurs procédés.

Dans le secteur de développement et commercialisation de la technologie, la compagnie
Matrikon, située à Edmonton (Alberta) occupe une importante part du marché canadien.
Parmi les autres importantes compagnies nord-américaines dans ce domaine on retrouve
Pacific Simulation (une division de Metso Automation, situé à Moscow, Idaho, aux Etats-
Unis), Umetrics Inc. (située à Kinnelon, New Jersey, aux Etats-Unis), Gensym (située à
Burlington, Massachusetts, aux Etats-Unis) et IETek (située à Tacom, Washington, aux Etats-
Unis).



Report – CETC 2005-100 (TR)                   v                                        June 2005
Des efforts plus grands doivent être réalisés afin d’accélérer l’implémentation des
technologies ‘’soft sensors’’ dans les industries de pâtes et papiers et ciment. Des projets de
démonstrations doivent être réalisés afin de démontrer et documenter l’impact économique de
cette technologie. Les centres de recherches et les universités doivent être aidées dans leurs
efforts de conception et développement des outils industriels ‘’soft sensor’’ standards.




Report – CETC 2005-100 (TR)                   vi                                       June 2005
                                                                 Table of Content

EXECUTIVE SUMMARY................................................................................................................................................... II

SOMMAIRE EXÉCUTIF...................................................................................................................................................IV

1          PROJECT OBJECTIVES .......................................................................................................................................1

2          AN INTRODUCTION TO SOFT SENSOR TECHNOLOGY.....................................................................3
           2.1         THE NEED FOR SOFT SENSORS ................................................................................. 3
           2.2         FIRST PRINCIPAL MODELS ...................................................................................... 4
           2.3         STATISTICAL BASED M ODELS................................................................................. 5
                       2.3.1 Classical Statistical Regression Models.................................................... 5
                       2.3.2 Factor Based Statistical Models................................................................ 5
           2.4         BLACK BOX MODELS ............................................................................................. 6
           2.5         THE KEY SUCCESS FACTORS FOR A SOFT SENSOR PROJECT ...................................... 7
           2.6         THE BENEFITS OF SOFT SENSOR TECHNOLOGY ...................................................... 9
           2.7         THE DRAWBACKS OF SOFT SENSOR TECHNOLOGY............................................... 10
           2.8         THE BARRIERS FOR IMPLEMENTATION OF SOFT SENSORS ...................................... 10
3          STATE OF THE ART IN SOFT SENSOR R&D AND INDUSTRIAL APPLICATIONS ................12
           3.1         INTRODUCTION ..................................................................................................... 12
           3.2         STATE OF THE ART IN SOFT SENSOR DESIGN TECHNIQUES ..................................... 12
           3.3         RELEVANT COMMERCIAL SOFTWARE TOOLS ......................................................... 16
           3.4         STATE OF THE ART IN SOFT SENSORS R&D ........................................................... 18
           3.5         CONCLUSION ........................................................................................................ 19
4          POTENTIAL APPLICATIONS OF SOFT SENSORS IN THE PULP AND PAPER INDUSTRY20

           4.1         CHEMICAL PULPING.............................................................................................. 20
           4.2         MECHANICAL PULPING ......................................................................................... 30
           4.3         NEWSPRINT AND PAPERMAKING........................................................................... 32
           4.4         CURRENT APPLICATIONS OF SENSORS IN PULP AND PAPER IN CANADA ................ 40
5          POTENTIAL APPLICATIONS OF SOFT SENSORS IN THE CEMENT INDUSTRY...................41

           5.1         PROCESS DESCRIPTION .......................................................................................... 41
           5.2         POTENTIAL APPLICATIONS .................................................................................... 42
6          INDUSTRY SURVEY............................................................................................................................................46

7          KEY AREAS OF FUTURE R&D.......................................................................................................................47
           7.1         DEVELOPMENT OF MULTI- GRADE MODELS ............................................................ 47
           7.2         DEVELOPMENT OF ADAPTIVE MODELS ABLE TO COPE WITH PRODUCTION CHANGES
                       AND SHUTDOWNS ................................................................................................. 47
           7.3         INCREASING THE AWARENESS OF PLANT MANAGERS AND OPERATORS ................. 47
           7.4         DEVELOPMENT OF USER- FRIENDLY SOFT SENSOR SOFTWARE ............................... 47
           7.5         RECOMMENDATIONS ON FUTURE WORK................................................................ 47
8          CONCLUSIONS ......................................................................................................................................................49

ANNEX A BIBLIOGRAPHICAL LIST.........................................................................................................................51


Report – CETC 2005-100 (TR)                                                          vii                                                                         June 2005
ANNEX B ARTICLE SUMMARIES ...............................................................................................................................55

ANNEX C INTERVIEW SUMMARIES ........................................................................................................................60


                                                             Table of Figures

FIGURE 1 - THE RELATIONSHIP BETWEEN INFORMATION AND SOFT SENSOR TECHNOLOGY ................ 9

FIGURE 2 – OVERVIEW OF THE KRAFT MILL PROCESS......................................................................................... 21

FIGURE 3 – OVERVIEW OF THE FINISHING STEP OF THE KRAFT PROCESS ................................................... 22

FIGURE 4 – OVERVIEW OF THE NEWSPRINT MILL PROCESS .............................................................................. 35

FIGURE 5 – OVERVIEW OF THE FINISHING STEP OF THE NEWSPRINT PROCESS......................................... 36

FIGURE 6 – OVERVIEW OF THE CEMENT MANUFACTURING PROCESS ........................................................... 41




                                                                List of Tables

TABLE 1 - SOFT SENSOR TECHNIQUES AND APPLICATION AREAS EXAMINED ............................................. 13

TABLE 2 - TYPICAL NEURAL-NETWORK BASED SOFT SENSORS INDUSTRIAL APPLICATIONS ............... 14

TABLE 3 – SOFT SENSOR COMMERCIAL SOFTWARE PACKAGES ....................................................................... 17

TABLE 4 – SUMMARIES OF ARTICLES REVIEWED.................................................................................................... 56




Report – CETC 2005-100 (TR)                                                viii                                                                June 2005
1        Project Objectives
This project constitutes the scoping phase of a possible longer-term research and development
program (R&D) that would ultimately aim at increasing the adoption of soft sensors or
"virtual analyzer" for process and energy performance in the pulp and paper and cement
industries.

It consists of the identification of the processes in pulp and paper and cement manufacturing
sectors, which are the most promising candidates to the development and implementation of
soft sensors. Also, this scoping study will help define a roadmap toward such long term R&D
program in this area. In other words, the objective of this scoping study is to get information
from process experts and operators, R&D experts and sensors and control systems developers
and providers to help with the development and implementation of this technology.

The specific objectives are:

    1. Perform a survey on the needs for soft sensors in the cement and pulp and paper
       manufacturing sectors;

    2. Identify the most promising processes and how soft sensors may increase productivity,
       product quality and energy efficiency in cement and pulp and paper manufacturing
       plants;

    3. If any, evaluate case studies and the degree of satisfaction from existing soft sensors
       and identify weak points;

    4. Identify key barriers to the development and implementation of soft sensors;

    5. Identify key barriers in the area of soft sensor development for industry (R&D,
       manufacturers, suppliers, etc.);

    6. Evaluate performance and benefits (economical and environmental) of soft sensors
       implementation and potential improvement;

    7. Get an expert point of view on whether these two sectors will adopt or not soft sensors
       and for which reasons.

    8. Make recommendations to industry and R&D institutions for soft sensors
       development and implementation.

The results of this scoping study will be used to direct future projects in soft sensors
development that will be performed in cooperation with industrial partners, universities,



Report – CETC 2005-100 (TR)                   1                                        June 2005
private and/or public laboratories, engineering companies and sensors and control systems
manufacturers.




Report – CETC 2005-100 (TR)                 2                                     June 2005
2        An introduction to soft sensor technology

2.1      The need for soft sensors

Soft sensor technology is an enabling technology for industrial users to improve their
productivity, to become more energy efficient, to reduce their environmental impact, to
improve their productivity and ultimately improve their business profitability.

A soft sensor is an empirical model that infers process state and product quality variables that
are difficult to measure on-line (composition, melt index, molecular distribution, etc.) from
readily available process measurements (temperature, pressure, flow, etc.)

The major problem in all industries including the cement and pulp and paper industries is the
lack of real-time measurement of product and process characteristics. This on-line
unavailability of critical process variables can lead to out of specifications production. A soft
sensor provides on-line, accurate estimates of these variables, eliminating additional energy
and production cost associated with out of specifications production.

In many cases an infrequent process sample ( once per hour, once per 8 hour shift, once per
day) which may not be representative of the process is taken to the laboratory and depending
on the reference standard being used in the laboratory, the final result will have typical delays
of between 20 minutes to 1 week . In many cases because of the whether the sample is
representative as well as reference standard being used, many companies regardless of type of
the industry will need a minimum of 3 laboratory results before taking action. For example, in
the pulp and paper industry, the Kappa number represents the hallmark of the quality of the
paper, being related to the lignin content remaining in the pulp. For a Kappa number which is
taken once an hour, with a laboratory delay of 20 minutes, the pulp mill could easily produce
3 hours off quality product before action can be taken. This increases the energy usage as well
as creates a greater environmental impact, since more waste is created which needs to be
disposed. If the final product cannot be sold due to its marginal performance, then this creates
the hidden factory, since often both industries will have to reprocess the material which
creates an additional cost due to more energy and chemicals consumption and additional labor
costs.

Soft sensors can also play a significant role in more complex systems used for process
optimization. In these systems, the soft sensor represents an individual component whose
output is used by another component of the system. The soft sensor can be used in a control
system, to provide the value of the variable being used in the feedback loop, for example. It
can also be used to identify a fault – a process disturbance, for example – and it transmits this
information to a system that analyzes and diagnosis the fault. Identifying, diagnosing and
predicting potential process disturbances can lead to significant production and energy
efficiency improvements Another application is sensor fault diagnostic, which is a sub-set of

Report – CETC 2005-100 (TR)                    3                                         June 2005
process fault detection: an algorithm (the soft sensor) infers a process variables, and it
compares it with the actual measured variable; a drift between those two values indicates a
deterioration of the sensor’s performance, and its potential failure. . Therefore, soft sensors
can play important multiple roles in optimizing industrial processes, as stand alone
applications or as individual components of more complex systems.

The goal of soft sensor technology is to relate physical properties or physical states to data
sets with a mathematical model. The data sets can be real-time process data from
programmable logic controllers and distributed control systems, manual entry data, and
specialized instruments; i.e. near-infrared sensors, Raman sensors, camera imaging systems,
etc.

The first applications of soft sensor technology began in the mid 80’s. Soft sensor technology
has matured during the 1990’s. Today there are three types of soft sensor technology available
to the users:

      1. First Principles Models

      2. Statistical Based Models

      3. Black Box Models

2.2       First Principal Models

First principle models are the least common soft sensor available today. The soft sensor
inferring algorithm is developed based on some understanding of the physics of the process,
such as mathematical functional relationships between the inputs and the outputs of the
system, or qualitative functions centered around different units in the process. These models
are typically more robust than either the statistical based or black box models, but they are
difficult to obtain since they require an accurate mathematical model of the process and
production data to calibrate the model coefficients. Currently two suppliers commercializing
products in the pulp and paper area provide first principle soft sensor or a mix of first
principles and statistical based soft sensors. Typically these models are used to predict
product quality parameters.

First principle models are ideal for use in feedback control applications and data
reconciliation applications. The user can extrapolate the results from the model. Unfortunately
the major drawback of first principle models is the complexity of building the model using
first principle equations, since often there is missing significant amount of critical information
to develop the model or the solution is difficult to obtain in a timely fashion.




Report – CETC 2005-100 (TR)                     4                                         June 2005
2.3       Statistical Based Models

Statistical based models use data fitting techniques such as regression methods in order to
estimate unknown process variables using available historic process data. They can be broken
into two sub-groups, classical statistical regression models and factor based models.

2.3.1      Classical Statistical Regression Models

Classical statistical regression models are based on multiple linear regression. The key
limitation of multiple linear regression is that the input variables must be statistically
independent of each other. Typically models of this type will have no more than 10 input
variables. If there are more than input variables, step wise regression techniques are used to
determine the best combination of input variables to predict physical characteristics.

The primary use of classical statistical regression models is in the area of near-infrared
sensors to predict physical properties. There is one industrial supplier which uses this
methodology to develop energy monitoring of processes.

2.3.2     Factor Based Statistical Models

Factor based Statistical Models which are used for soft sensors are principal components
analysis (PCA) and partial least squares also known as projection to latent structures (PLS)
and principal component regression (PCR). Factor based statistical models are designed to
work with industrial data sets. They are able to cope with the following issues of industrial
data:

      Ø   Noisy data sets

      Ø   Missing data in the data sets

      Ø   Correlated variables within the data sets

      Ø   Large data sets, both observations and variables

      Ø   Data sets with many variables and a small number of observations

      Ø   Data sets with many observations and a small number of variables

Factor based statistical models are used to transform a number of related process variables to
a smaller set of uncorrelated variable. These techniques are often called data reduction
methodologies since they identify factors that have a much lower dimension than the original
data set and still can properly describe the major trends in the original data set. Often 100
input variables can be reduced to 5 or 6 new variables explaining between 40 to 80% of the
process variability of industrial data sets.


Report – CETC 2005-100 (TR)                       5                                   June 2005
Principal component analysis models typically are used to develop multivariate statistical
process control charts, as well as on-line classification detection. Principal component
regression and partial least squares models are used for predicting product quality parameters.
These models are useful to process analysts in seeing the relationships between the input
variables and output variables respectively.

Typical applications for these types of soft sensors are:

      Ø   Multivariate statistical process control charts to detect abnormal situations and help
          locate the cause of the problem.

      Ø   Classifying products

      Ø   Predictions of product characteristics

These sensors can be used for feedback control; however, the model must be built using
designed data from a design of experiment, and not from data from a data historian. These
techniques show the correlation of the various inputs and outputs within the range of data
recorded. For the model to remain valid all the variables in the model must change
simultaneously in the right proportion to maintain the correlation structure. Also you cannot
extrapolate the results of this model outside the range of its original data range. A feature of
these techniques is that there is a built-in self diagnostic to detect when incoming data
correlation structure is different than the correlation structure of the original model to warn
the user that model predictions are no longer valid.

2.4       Black Box Models

Black box models are typically based on neural net and/or fuzzy logic techniques.

Artificial Neural Networks (ANN) is a type of Artificial Intelligence technique that mimics
the behavior of the human brain; it attempts to describe a nonlinear relationship between the
input and output of a complex system using historic process data. Similar to a human neural
system, an artificial neural network is an information processing structure that consists of a
number of input units and output units connected in a systematic fashion. Between the input
and output units, there may be one or more hidden layers, each consisting of a number of
units called neurons, nodes or cells. The connections between units lying on different layers
are assigned with varying weights. Input signals (or data sets) are fed in from the input layer,
and they follow all possible connection paths to reach the next layer.

Along each connection link, the signal suffers a transformation, eventually reaching the
output layer. Through this mechanism an ANN learns to identify patterns in the data set and
predict variables.



Report – CETC 2005-100 (TR)                        6                                        June 2005
Fuzzy logic is an Artificial Intelligence technique that mimics the human reasoning by using
gradients of true and false. A fuzzy system can convert a set of user-supplied human language
rules into mathematical equivalents. It can also use historic process to create a set of rules that
are use to estimate the unknown variables.

These types of models can provide the same type of applications as factor based statistical
models. Neural net models prefer data sets whose inputs are independent. Many neural net
packages have tools to determine if variables are correlated to allow the user to remove the
correlated variables. They are able to cope with the following issues of industrial data:

      Ø   Noisy data sets

      Ø   Missing data in the data sets

      Ø   Large data sets, both observations and variables

      Ø   Data sets with many observations and a small number of variables

These types of models are more difficult to interpret the relationship between the variables
and if the user changes the order of the dataset, a new model can be obtained.

These sensors can be used for feedback control; however, the model must be built using
designed data from a design of experiment for the same reasons as factor based models. The
results of this model cannot be extrapolated outside the range of its original data range.

Typical applications for these types of soft sensors are:

      Ø   Detecting abnormal situations and help locate the cause of the problem.

      Ø   Classifying products

      Ø   Predictions of product characteristics

2.5       The key success factors for a soft sensor project

A successful soft sensor project requires the following key components:

      Ø   A good business driver (i.e. reduce energy costs, eliminate out of specification
          product, etc)

      Ø   An expert in the soft sensor technology

      Ø   An process expert where the soft sensor technology will be used




Report – CETC 2005-100 (TR)                        7                                         June 2005
    Ø   An infrastructure to collect the data from the process (i.e. distributed control system or
        programmable logic controller connected to a data historian)

    Ø   A sampling rate appropriate to the process

    Ø   A good sample collection system to obtain a representative sample of the parameter to
        be predicted

    Ø   A dataset which adequately represents the process for normal operation to develop and
        validate the sensor

    Ø   A good response standard to develop the soft sensor

    Ø   A relatively stable production process in terms of design (i.e. processes not being
        constantly reengineered for process improvements)

If any of these components are missing, then implementing a soft sensor project will likely
fail.

A soft sensor project is an iterative process consisting of getting process data, cleaning up the
process data, developing a soft sensor model, reviewing the model performance and if the
performance is acceptable placing the sensor on-line. If the performance is not acceptable,
obtaining more data, redesigning the soft sensor model, in extreme cases changing the
reference standard and repeating the process until the soft sensor is acceptable or rejecting the
technology all together.

Soft Sensor technology will have an indirect effect on energy and environmental impact
reduction. Most of the soft sensor economic benefits to any industrial process is increased
productivity to an organization, thus increasing its efficiency in converting the raw material to
the final product.

An additional benefit of soft sensor technology is the opportunity for process personnel and
management to obtain a better understanding of their process. This process knowledge
coupled with additional process changes will allow them to improve their bottom line. This
fact is well understood in the pharmaceutical industry, where the US FDA is now promoting
the process analytical technology program. The goal of this program is for pharmaceutical
companies show to the FDA their understanding of their processes.




Report – CETC 2005-100 (TR)                     8                                         June 2005
               Figure 1 - The relationship between information and soft sensor technology


Soft sensor technology is now available as standard options in new distributed control
systems, as well as add-on options to existing systems.

Soft sensor technology is designed to recognize patterns in datasets. The soft sensor uses this
pattern recognition to predict future results.

Soft sensor technology is next logical step in today’s manufacturing environment of testing in
the final quality of the product to manufacturing in the final quality of the product.

2.6       The Benefits of Soft Sensor Technology

The ultimate benefit of soft sensor technology is allowing companies to improve their bottom
line through the reduction of operating costs and improvement in the product quality. The
users are able to achieve these results because of:

      1. understanding obtained during the process analysis phase in implementing this
         technology and apply process changes based on this analysis, and

      2. rapidly converting raw data into useful feedback information for operators and
         management to take rapid action to correct out of control scenarios

      3. eliminating additional energy and production cost associated with out of specifications
         production due the on-line unavailability of critical process variables



Report – CETC 2005-100 (TR)                        9                                        June 2005
      4. providing critical information to other process optimization systems, such as fault
         identification and diagnosis systems

Typical return of investments of 50% or higher are common in the application of soft sensors.

2.7       The Weaknesses of Soft Sensor Technology

There are a number of drawbacks with soft sensor technology. The drawbacks are related to
problem that soft sensor technology is data driven. The particular drawbacks are:

      1. the data requirements to build a model may require multiple years of data to accurately
         model the process,

      2. most soft sensor models are data driven and custom built. Processes which are
         modified frequently are not good candidates for a soft model; for example, if the
         process has a significant process design change, then a new model must be
         redeveloped, or each production Grade requires a separate model

      3. the soft sensor model prediction cannot be more accurate than the laboratory standard
         used to calibrate the model.

      4. the lack of qualified personnel to monitor and maintain the performance over time

2.8       The barriers for implementation of soft sensors

The potential of soft sensors combined with advanced control strategies has the potential of
saving the Canadian Pulp and Paper Industry between $40 to 100 Million//year and for the
Canadian Cement Industry approximately $35 million/year. For this technology to be
accepted, a major education program must be put into place at all levels of the Canadian
corporate system. Support for soft sensors must come from both the boardroom and shop
floor.

Currently, the majority of engineers, technicians, managers and executives do not know what
is a soft sensor and its economic implications. This is rapidly changing in the Pulp and Paper
through the excellent work of Dr. Tranh Trung.

Soft sensors are most efficient when combined with a basic or advanced control strategy to
achieve the economic goals that the end user is demanding. A basic control scheme can be the
manual adjustment, by the operator, of process settings as a result of new process or product
information provided by the soft sensor.

University programs should be adjusted to include multivariate data analysis at the bachelor
level. Universities in Sweden provide this training to their chemists and engineering students.



Report – CETC 2005-100 (TR)                     10                                       June 2005
In Canada there are only two universities that have courses in multivariate data at the master
level for engineering students.

Soft sensor technology is just a tool. Understanding the process relationships is the key for
successful energy management in today’s process industries.




Report – CETC 2005-100 (TR)                  11                                       June 2005
3        State of the art in soft sensor R&D and industrial applications

3.1      Introduction

A literary analysis was carried out in order to provide an overview of soft sensor design
techniques and their applications in the industrial process manufactory by analyzing the
development of related published work. Canadian academic research activities related to soft
sensor R&D were also examined.

3.2      State of the art in soft sensor design techniques

Different techniques used to develop soft sensor algorithms were identified, as well as their
suitable applications – for example, prediction vs. classification: a soft sensor used to estimate
(predict) a temperature value can use a different technique than a soft sensor used to classify
products according to their quality. The survey did not include only stand-alone soft sensor
applications. It was extended to more complex systems in which the soft sensor represents
one individual component whose output is used as an input by another component, such a
fault identification and diagnosis system, where the soft sensor that can identify a
instrument/process fault, and then an expert system uses this information in order to diagnose
(classify) the fault.

Some of the soft sensor applications described in these publications are already implemented
on-line in industrial installations and some of them have been tested off-line, using actual
process data.

The publications reviewed and analyzed in this work comprise papers that appeared in recent
issues of several well-established publications, industrial journals and presentations from
different international conferences related to soft sensor technology; material published from
universities and research centers was also reviewed and analyzed. A bibliographical list of the
examined publications is presented in Annex A, while brief summaries of these articles are
presented in Annex B.




Report – CETC 2005-100 (TR)                    12                                        June 2005
The table below shows the main applications and techniques described in the publications
reviewed:

                     Table 1 - Soft sensor techniques and application areas examined

                 Soft sensor techniques and application areas examined

      Application areas                                       A.I. technique

Value inference / prediction        Neural networks (NN): backpropagation, Bayesian,
                                    recurrent, multi-layer perceptrons
                                    Principal component analysis/projection to latent structures
                                    (PCA/PLS – statistical techniques (offer predictive
                                    modelling capabilities)
Single-class membership             Rho NN
classification: fault detection     PCA
Multiple-class membership           Radial Basis Function NN
classification: fault diagnosis     PCA
Signal filtering: noise             Auto-associative NN
reduction                           PCA

Faulty sensor detection and         Auto-associative NN
substitution                        PCA/PLS
System modeling                     Sugeno fuzzy logic model
                                    PCA
Data clustering (system             Kohonen self-organizing map NN (unsupervised clustering
identification)                     of input data)
                                    PCA




Report – CETC 2005-100 (TR)                        13                                     June 2005
It can be seen that a large number of these soft sensor applications are neural-network based.
Some of the typical neural-network based soft sensors applications implemented in the
industry are as follows1:

                Table 2 - Typical neural-network based soft sensors industrial applications

       Industry                 Predicted Variables                               Benefits
    Environmental :      Exhaust gas oxygen, CO,                 Less than ½ of initial investment
    Emissions            NOX, SO2 ,CO2 , Opacity.                compared to hardware analyzers;
    Monitoring &         Applications: Gas, steam                reduced maintenance costs
    Control              turbines, boilers, and furnaces.        (usually
                                                                 1/3); Enhanced production
                                                                 efficiency.
    Polymers             Melt flow index, density,               Improved process efficiency,
                         isotactic index, co-monomer             better quality control, less product
                         and monomer concentration,              give-away
                         catalyst flows, polymer
                         concentration and residence
                         time
    Chemical Plants      Product composition, impurity
                         estimation, reactor or
                         distillation
    Food and             Weight and moisture of food
    Beverage             products. Stack height. Solvent
                         in product, and powder
                         material properties.
    Pulp and Paper       Kappa numbers, viscosity,               Significant increase in on-grade
                         brightness, strength, stiffness,        production, reduction in transition
                         freeness, E modulus, fiber              times, increased production,
                         coarseness, opacity, porosity,          optimal energy usage, predictive
                         specific surface area, wet end          boiler emissions monitoring,
                         chemistry, and crush test.              sheet-break prediction
Refining                 Boiling points, octane number,
                         cloud point, distillation, flash
                         points, freeze point viscosity,
                         cut points, end points, overhead
                         quality, bottoms residues, tail
                         gas H2S (sulfur plant), PEMS
                         for utility boilers and cracking
                         furnaces, Intermediates, octane,
                         cetane, RVP, sensor values.


1     Achieving consistent product quality, increased throughput, and minimal use of resources with Gensym’s
      NN based soft analyzers (a paper published by the software maker Gensym); the information presented is
      based on the company’s industrial experience and implementation of its products in various industries



Report – CETC 2005-100 (TR)                           14                                             June 2005
Quite a few applications of neural networks-based soft sensors can be found in the pulp and
paper industry:

    Ø   in 1998, there were already 30 mills world-wide using soft sensors for estimating
        paper sheet properties, digester kappa numbers, and for emissions monitoring
        applications (Finchem K., 1998)

    Ø   a new prototype soft sensor for measuring pulp brightness was being tested in a mill in
        Germany (Bartos, F., 2003)

Some soft sensor applications are based solely on statistical methods. One such typical
application is a PCA/PLS combination. The PCA analysis provides an overview of the data
and can model it, establishing correlations between the variables. Unlike black-box methods,
such as neural networks, this approach offers an understanding of the process, by revealing
inter-relationships among the different variables. The projection to latent structures (PLS)
provides predictive modeling capabilities used to infer the output (the value to be estimated).
This combination has been demonstrated to be effective for variable prediction, selection of
significant process variables, improving process diagnostics and knowledge (Zamprogna and
Seboprg, 2005; Wold S. et al., 2003). Multivariate analysis and neural network based soft
sensors are currently implemented in the industry, mostly for predict various process
variables, as reported by the software makers (Umetric’s SIMCA-P and Pavillon
Technologies’ Property Predictor, among others).

This literature assessment indicated a growing trend of combining statistical methods, such as
principal component analysis with artificial intelligence methods, such as neural networks.
When dealing with a process containing a large number of variables, PCA can be used in
order to determine the available process measurements that have the most significant
influence on the variable to be inferred. This allows for a reduction in the number of
variables, without losing significant information. It also allows for a selection of the proper
inputs, the variables that really have an effect on the output – the value to be estimated. These
variables are then used as the inputs for the soft sensor, which will infer the value to be
estimated. By combining these data analysis techniques with a soft sensor technique, the
inputs that have the greatest impact on the output are selected, and the configuration of the
soft sensor algorithm is greatly simplified, since the number of the input variables is reduced.

Publications were found describing the same combination: PCA and NN. The conclusion of
the work presented in theses papers is the same: the soft sensor using the inputs determined
by the PCA analysis performs much better that the soft sensor using conventionally
determined inputs (Zamprogna and Seboprg, 2005; Edwards and al, 1999; Qin and al, 1997).




Report – CETC 2005-100 (TR)                    15                                        June 2005
Proper input identification leads to an increase of the predictive performance of a neural
network, and comparative studies prove that this method yields superior results than the
traditional methods.

The successful performance of implemented soft sensors in industry led to their incorporation
in more complex systems, such as control systems and fault and diagnosis systems (Fortuna
and al., 2005; Patan and Parisi, 2005).

The soft sensor can be used in a control system, to provide the value of the variable being
used in the feedback loop, for example. It can also be used to identify a fault – a process
disturbance, for example – and it transmits this information to a system that analyzes and
diagnosis the fault. Identifying, diagnosing and predicting potential process disturbances can
lead to significant production and energy efficiency improvements. Therefore, soft sensors
can play important multiple roles in optimizing industrial processes.

A three-part series paper presenting a systematic and comparative study of various process
diagnostic methods was examined (Venkatasubramanian and al., 2003). The application of
statistical methods – such as PCA – and artificial intelligence methods – such as neural
networks – was examined in these papers; the industrial implementation of history based
methods is highlighted.

The application of soft sensors for prediction of energy consumption was also examined. In
this area, almost exclusively, the publications examined dealt with the prediction of energy
consumption for buildings, and the preferred approach was neural network-based
(Devogelaere and al., 2002; Olofsson and Anderson, 2001; Kaloogiru and Bojic, 2000)

However, this approach for energy consumption estimation can be generalized to
manufacturing processes, since this application consists of value prediction using measured,
historical process data – a typical use for soft sensor application in predicting the energy
consumption of industrial processes.

3.3      Relevant commercial software tools

Key features of some selected commercial software used to design and/or implement soft
sensor applications were examined. There are a multitude of commercial tools offering
predictive capabilities based on statistical analysis or on artificial intelligence approaches.
Some of them have a considerable market penetration in the industrial manufacturing sector.

Some of the commercial available statistical and AI-based tools are presented below. The list
is not exhaustive for there are a large number of computer-based tools offering a variety of
artificial intelligence and statistical add-ons. The products reviewed focused either on soft
sensor industrial applications or on general suitable tools for designing (programming) the
soft sensor algorithm.

Report – CETC 2005-100 (TR)                   16                                       June 2005
                              Table 3 – Soft sensor commercial software packages

                                                                                       Industrial
 Product                Vendor                             Features
                                                                                   implementations
FactNet        Pacific Simulation             statistical analysis package         on-line applications
                                              based on the factor analysis         in the pulp and paper
                                              technique                            and power industry
               www.pacsim.com/FN



NeurOn-        Gensym                         object-oriented software for         on-line applications
Line                                          building neural networks for         in a manufacturing
                                              real-time applications.              industries, such as
               www.gensym.com                                                      chemical, cement
                                                                                   and automotive
AspenIQ        Aspen Tech                     inferential package (value
                                              prediction);models based on
                                              neural networks, fuzzy logic,
               www.aspentech.com              linear and non linear PLS.



Simca-P        Umetrics               statistical data analysis package,           chemical,
                                      uses principal components                    pharmaceutical, pulp
               www.umetrics.com
                                      analysis and partial least                   and paper production
                                      squares; predictive models that
                                      can be used as soft sensors can
                                      be created.
Property
               Pavillion Technologies multivariate data analysis tools             pulp and paper,
Predictor                             to build predictive models                   cement
               www.pavtech.com        acting as soft sensors for paper
                                      machine sheet properties
Pegasus
               Pegasus Technologies neural networks models to                      power industry
OS2003                                optimize boiler operation. It acts
               www.pavtech.com        as a soft sensor analyzing
                                      emissions and using this
                                      information to optimize
                                      combustion, by computing
                                      optimal air and fuel flows.




Report – CETC 2005-100 (TR)                           17                                        June 2005
                                                                                  Industrial
 Product                Vendor                       Features
                                                                             implementations
                                                                             widely used in
Matlab’s        Mathworks                toolboxes for the Matlab
                                                                             universities, R&D
Neural                                   programming language. They
                www.mathworks.com                                            centers and other
                                         offer a variety of functions that
Network                                                                      institutions for soft
                                         are used for neural network
and                                                                          sensor development
                                         design and statistical analysis
Statistical                                                                  (programming)
                                         purposes.
toolboxes

Most soft sensor software makers also offer control and expert system modules that
incorporates the soft sensor module. Here are just a few examples, representatives of the soft
sensor software packages:

      Ø   Gensym’s G2 expert system that can incorporate soft sensors for optimizing
          operations and detecting, diagnosing, and resolving process faults and/or equipment
          failures

      Ø   Pavilion’s ProcessPerfecter can incorporates soft sensor created with Pavilion’s
          Property Predictor package in order to provide a multivariable predictive advanced
          process control solution.

While these tools are quite powerful of designing and implementing on-line either statistical-
based predictive models or AI-based inferring mechanisms, they fail to offer the possibility of
combining both methods into a single hybrid system. This is a major drawback, since such a
hybrid system is more powerful than a system based on only one method, as it was revealed
in the literature examined for this assessment study. Design a software tool that would allow
the design, testing and on-line implementation of this hybrid system would represent a major
advancement in the filed of industrial soft sensor development.

Also, the relative high cost for acquiring a license of some of the popular tools represents an
important barrier in their application in industry.

3.4       State of the art in soft sensors R&D

Currently there are 2 universities in Canada devoted to Soft Sensor development in general:

      Ø   Mc Master University, Department of Chemical Engineering, under the direction Dr.
          John McGregor

      Ø   University of Alberta, Department of Chemical Engineering, under the direction Dr.
          Sirish Shaw



Report – CETC 2005-100 (TR)                     18                                         June 2005
Both the University of Toronto Pulp and Paper Centre and L’Ecole Polytechnique have begun
work in trying to apply soft sensors in the pulp and paper industry.

Some of the key areas of research is now multivariate imaging sensors to quality and energy
as well as applying more NIR technology. Much of NIR technology research for Pulp and
Paper is being developed by Dr. Tranh Trung at Pulp and Paper Research Institute of Canada.
No work is being done in Canada on the use vibrations to estimate product characteristics.
There is a fair amount of work that is being developed in this area in Sweden.

3.5      Conclusion

The analysis of published work regarding soft sensor technology applications in the industrial
process manufactory provided information about current trends of soft sensor development.
Current applications were identified, along with suitable methods – the soft sensor algorithm
– to perform the required task. It was seen that different methods are used for different tasks:
for example, a soft sensor used to predict a process variable value will have a different
configuration than one used to classify product quality. Soft sensors are used to optimize
industrial processes and they can be used as stand-alone applications, or they can be
incorporated into other optimization systems, such as control and fault and diagnosis systems.

Different approaches can be employed to build the predictive model used by the soft sensor;
they can be based on statistical analysis, on artificial intelligence techniques, or on a
combination of both approaches. For value prediction, most of the published work reviewed
dealt with neural network-based approaches. The industrial implementation of statistical
methods such as PCA is relatively well established. This approach offers the advantage of
providing correlations among process variables, therefore offering an understanding of the
process.

However, the literature review clearly showed that a hybrid configuration using both these
methods performs better than soft sensors based solely on one of the either methods.




Report – CETC 2005-100 (TR)                   19                                        June 2005
4         Potential applications of soft sensors in the Pulp and Paper
          industry

4.1       Chemical pulping

There are two types of chemical pulping in Canada:

      1. the sulfate process, also known as the Kraft process

      2. and the sulfite process.

The Kraft pulping process consists of producing pulp by cooking wood chips in pressure
vessels in the presence of an alkaline solution (it has a high pH) during which time the
chemicals attack the lignin in the wood.

In the sulphite pulp process the cooking liquor is acidic (it has a low pH) or neutral. The
kappa number of the pulp is used as a measure of the pulp quality, by estimating the residual
lignin content.

The Kraft process is typically an energy neutral process, since it is able to reclaim and reuse
most of its energy requirements with the process. Most Kraft pulp mills in Canada today use a
biomass fuel source (hog fuel) as its primary fuel for its power boilers and the recovery
boilers generate steam energy from the recuperation of necessary chemicals required for the
breakdown of wood chips to turn into wood pulp. The main users of greenhouse gas fuels are
typically the lime kiln and the power boilers when the hog fuel is not available or is too wet.
Most energy savings in Kraft pulp mills processes outside of the power house and lime kiln is
due to increased productivity by obtaining better wood yields at the digester house. Most
Kraft mills have a wood yield of between 45 and 48 percent. The typical Kraft cooking value
is a Kappa number of 20 to 25. An upward shift in the Kappa number target of 1 kappa will
typically result in a wood yield increase of .5 to 1 percent. For a typical 600 metric tons per
day (mtpd) Kraft pulp mill, a one percent wood yield increase means that approximately 25
mtpd less of wood chips is required. It takes approximately 1.9 GJ of energy to produce 1
ADMT wood pulp during digester process. The annual energy savings would be
approximately $56,000. The annual wood chip savings would be approximately $1,135,000.
With Eastern Canada’s tight wood supply, the goal is to increase the pulp yield.

The sulfite process is an energy consumer process, since it cannot reclaim and reuse most of
its energy requirements. There are 3 sulfite mills in Canada, and only 2 mills are still in
operation. Most of the steam energy produced is with a biomass fuel source (hog fuel). The
main user of greenhouse gas fuels are the power boilers when the hog fuel is not available or
the hog fuel is too wet to burn. The sulfite process has similar problems as Kraft pulping



Report – CETC 2005-100 (TR)                     20                                     June 2005
process, except that its wood yield is between 43 and 47 percent. The savings on wood yield
would be similar to the Kraft process.

An overview of the Kraft pulp mill process is shown below:




1 Chip Pile
5 Blow Tank
9 Brown Stock Thickener
13 Bleach Washer
17 Seal Tank


2 Chip Screening
6 Knotters
10 Brown Stock H.D. Storage
14 High Density Pump
18 Cleaners


3 Digester
7 Washer
11 D 100 Tower
15 Caustic Tower
19 Thickener


4 Atmospheric Diffuser
8 Screens
12 Seal Tank
16 Bleach Washer
20 High Density Storage




                                       Figure 2 – Overview of the Kraft mill process


Please Note:              At the end of the above process, a small percentage of the output is available as additional
                          furnish for the newsprint mill.

The finishing process of the Kraft Pulp process is shown below:




Report – CETC 2005-100 (TR)                                  21                                              June 2005
21 Machine Chest
25 Dryer
29 Wrapper Dispenser


22 Head Box
26 Sheet Cooler
30 Tying Machines


23 Forming Wiring
27 Cutter Layboy
31 Stenciller


24 Press Section
28 Bale Press
32 Stacker




33 Transport




                       Figure 3 – Overview of the finishing step of the Kraft process


The following areas could benefit form soft sensor technology:

1)           Recaustization Area, Steam Plant, Evaporators Area, Brown Stock Washing,
             Bleach Plant, Pulp Machine, Paper Machine: On-line Fault Detection and
             Property Prediction System

Developer:                        Effective Assets, Matrikon, Gensym, Pavilion, Umetrics,
                                  BestWood

Description:


Report – CETC 2005-100 (TR)                         22                                      June 2005
On-line Fault Detection System allows the operator and management to rapidly detect the
location of the cause of the process shutdown or process upsets. This would speed up down
times from hours and days to minutes. Tembec has implemented this type of fault detection
system in one area of their operations and have it has saved them at least $80,000/year due to
higher productivity.

On-line property prediction system allows operators a new sensor to predict the laboratory
results in real-time, which help mill personnel to rapidly react to changing process conditions
which are impacting on key properties characteristics. Rapid response to changing
characteristics will reduce the off-specification final product which in turn will save industry
money in energy, raw materials and waste treatment.

This type of sensor can be used in any part of mill operation.

Cost To Implement:                            $30,000 to $150,000 per area

Potential Industry Wide Energy Savings*:      250,000 to 500,000 GJ/year
                                              if the entire mill uses this technology

Current Status:

This is still at the demonstration stage. More education and skilled personnel are required at
the mills and in some locations a new data collection infrastructure is needed to make this
work. This technology has a high potential of making a major impact to the industrial market.

Potential Mill Savings:                       $20,000 to 120,000/year per area

Other Requirements:                           Real Time Data Network
                                              Good Database
                                              Multivariate Data Analysis Package

Hardware Requirements:                        PCs

Other Industries:                             Cement, Oil and Gas, Steel, Petro-Chemicals,
                                              Chemicals, Mining

Contact:                                      Effective Assets, Makrikon, Umetrics, Pavilion,
                                              Gensym, BestWood

2)      Recaustization Area: On-line measurement of the recaust system liquor strength

Developer:                                    Pulp and Paper Research Institute and Modo-
                                              Chemetics



Report – CETC 2005-100 (TR)                    23                                       June 2005
Description:

ABB has purchased the patent rights from Modo-Chemetics and this patent is also jointly held
by Pulp and Paper Research Institute of Canada. This technology uses the spectrum from a
Near Infrared Analyzer together with a PLS algorithm to create a soft sensor to predict the
liquor strength from 3 to 5 points in the recaustization process.

Better control of liquor strength allows the user to increase his productivity at Kamyr
Digester, which means lower overall energy consumption.

Cost To Implement:                          $350,000 to $500,000

Potential Industry Wide Energy Savings*:   25,000 to 65,000 GJ/year

Current Status:

This is at the implementation stage. It is a commercial product available to pulp and paper
companies worldwide. This technology has a high impact to pulp and paper industry.

Potential Mill Savings:                    $300,000 to 600,000/year

Other Requirements:                        Distributed Control System

Hardware Requirements:                     ABB Liquor Analyzer

Contact:                                   ABB Canada




Report – CETC 2005-100 (TR)                 24                                      June 2005
3)      Recaustization Area: Prediction of Lime Kiln Ringing issues

Developer:                                 University of Toronto Pulp and Paper Centre

Description:

Teresa D’Souza, a master student, under the supervision of Dr. Joe Ripka and Dr. Honghi
Tran at the University of Toronto’s Pulp and Paper Centre is currently doing research in
predicting what the internal temperature conditions are which causes a lime kiln to lose its
production efficiency by creating ring formation within a lime kiln. This lack of efficiency
will cause mill operations to use more energy to compensate for this problem. If the ring
formation is serious, then the production must take downtime to correct the problem. If this
sensor can be commercialized, it would give operators advance notice of this condition, and
prevent ring formation.

Estimated Cost To Implement:               Not Known

Potential Industry Wide Energy Savings*: Not Known

Other Requirements:                        Real-time data network
                                           Process database
                                           Soft Sensor Server

Current Status:

This is at the R&D stage. The University of Toronto Pulp and Paper Center is working in this
area. This is part of a Master’s student Graduate thesis. This may have a moderate to high
impact to pulp and paper and cement industries.

Hardware Requirements:                     Server and Monitors for Soft Sensor Server

Other Industries:                          Cement

Contact :                                  Prof. Honghi Tran,
                                           University of Toronto, Pulp and Paper Centre
                                           Phone: (416) 978-8585
                                           Fax: (416) 971-2106
                                           Email: tranhn @ chem-eng.utoronto.ca




Report – CETC 2005-100 (TR)                  25                                     June 2005
4)      Recaustization Area, Utilities Area, Environment Control: Prediction of the final
        lime quality

Developer:                                   ProSensus, McMaster University

Description:

Dr. John McGregor at McMaster University has been working on this problem for the last 5
years. ProSensus is a new spin-off company from Dr. McGregor’s research at McMaster
University. This firm has been established to provide industrial solutions using multivariate
image analysis. One of the potential applications of this technology is applying multivariate
image analysis on the flame within the lime kiln to predict the lime quality and liquor
strength. An additional benefit of this technology is its ability to predict the emission
pollutants as well. This technology has the benefit of better control of lime kiln energy costs
and Kraft digester energy costs.

Another issue in Canadian pulp and paper industry is the use of biomass fuels to replace
natural gas. In a paper published in AiCHe, Dr. McGregor shows that this technique can also
be used to estimate the fuel BTU content. This information will also be very critical for
operator or control system to obtain the proper combustion parameters in the kiln burner.

Estimated Cost To Implement:                 $80,000 to 100,000

Potential Industry Wide Energy Savings*:     50,000 to 130,000 GJ/year

Potential Mill Savings:                      up to $50,000/year

Current Status:

This is at the demonstration stage. The R&D work has shown that this technology is feasible.
An industrial partner as well as a supplier is needed to implement this technology. This has a
high impact to all industries.

Other Requirements:                          Real Time Data Network

Hardware Requirements:                       Server and Monitors for Soft Sensor Server

Other Industries:                            Cement, Energy, Steel, Oil and Gas, Chemical

Contact:                                     Prof. John McGregor, McMaster University
                                             Phone:     (905)   525-9140   ext.24951
                                             Email: macgreg@mcmaster.ca




Report – CETC 2005-100 (TR)                   26                                       June 2005
5)       Kamyr Digester, Utilities Area: New on-line wood moisture soft sensor

Developer:                                  Pulp and Paper Institute of Canada

Description:

Wood moisture has a significant impact to a hog fuel boiler efficiency as well as having a
major impact to the final Kappa number of wood pulp.

Paprican has developed an on-line wood moisture soft sensor using a filter based NIR unit.
The proper measurement of wood chip moisture has been a major problem for the pulp and
paper industry. This unit will cost approximately $30k and has a payback of less than 3
months.

Estimated Cost To Implement:                $30,000 to $50,000

Potential Industry Wide Energy Savings*:    50,000 to 130,000 GJ/year

Potential Mill Savings:                     up to $100,000/year

Current Status:

This is at the demonstration stage. A number of Pulp and Paper mills in Canada are looking to
try this technology. This will have a moderate impact to pulp and paper industries.

Other Requirements:                         Real-time   data   network,   Process   database


Hardware Requirements:                      None

Other Industries:                                  Energy, Chemical

Contact:                                    Tranh Trung, Pulp and Paper Institute of Canada

Phone:                                      (604) 222-3259
                                            Email: ttrung@paprican.ca




Report – CETC 2005-100 (TR)                  27                                      June 2005
6)       Kamyr Digester: New on-line Kappa soft sensor

Developer:                                  Pulp and Paper Institute of Canada

Description:

Paprican has developed a new prototype kappa analyzer using Raman technology based on a
statistical multivariate regression. This analyzer has the advantage of being species
independent.

Better control of the Kamyr Digester will allow the mill to change its operating parameters to
increase its yield up by 1 percent which means less energy is required. On a typical 600 tpd
mill , the mill would need 25 tons per day of wood chips to make the same amount of pulp. It
is currently being trialed at a number of Kraft mills in Canada

Estimated Cost To Implement:                 $250,000 to 400,000

Potential Industry Wide Energy Savings*:    100,000 to 200,000 GJ/year

Potential Mill Savings:                      up to $1,200,000/year

Current Status:

This is at the demonstration stage. A number of Pulp and Paper mills in Canada are looking to
try this technology. This will have a high impact to pulp and paper industry.

Other Requirements:                         Real-time data network

Hardware Requirements:                      None

Other Industries:                           Not Applicable

Contact:                                    Tranh Trung, Pulp and Paper Institute of Canada

Phone:                                      (604) 222-3259
                                            Email: ttrung@paprican.ca




Report – CETC 2005-100 (TR)                  28                                       June 2005
7)       Kamyr Digester: Off-line Kamyr Digester First Principal Model

Developer:                                  IETek

Description:

IETek – Integrated Engineering Technologies of Tacoma has developed a first principle off-
line model of the Kamyr digester which was partially funded by the US department of energy.
This off-line soft sensor allows the end user to play with his Kamyr digester process to
optimize the operating conditions to improve productivity, reduce energy and improve the
overall pulp quality. It has been reported at a US pulp mill, the mill personnel were able to
improve the quality of pulp and reduce the Kamyr digester cooking temperature by 2 degrees
F. This temperature reduction would save approximately US$12K per year in fuel costs.

The next step is placing this model on-line and adapt automatically to changing process
conditions.

Estimated Cost To Implement:                $100,000

Potential Industry Wide Energy Savings*:    20,000-40,000 GJ/year

Potential Mill Savings:                     up to $200,000/year

Current Status:

This is at the demonstration stage in Canada. An American pulp mill has used this technology
and has shown it has improved their product while reducing their operating costs This will
have a moderate impact to the pulp and paper industry.

Other Requirements:                         Real-time data network

Hardware Requirements:                      None

Other Industries:                           Not Applicable

Contact:                                    Ferhan Kayihan, PhD

Phone:                                      253-925-2179
Fax:                                        253-925-5023
Email:                                      fkayihan@ietek.net




Report – CETC 2005-100 (TR)                  29                                      June 2005
4.2      Mechanical pulping

Mechanical pulping is an electrical energy intensive activity; it consists of mechanically
grinding logs or wood chips. Today’s most common mechanical pulping technique is either
bleached or unbleached Chemi-Thermal Mechanical pulping for hardwoods or Thermal
Mechanical pulping for softwoods. TMP (Thermo Mechanical Pulp) is made by steaming wood
chips under pressure prior to and during refining, while Chemo-Thermo-Mechanical Pulp (CTMP) is
produced by treating wood chips with chemicals (usually sodium sulfite) and steam before
mechanical defibration.

The goal for many operations is to provide sufficient energy to obtain the desired pulp
freeness and strength qualities. The following areas could benefit form soft sensor
technology:

1) TMP Area, BCTMP Pulp Bleaching: On-line Fault Detection and Property

Prediction System

Developer:                                   Effective Assets, Matrikon, Gensym, Pavilion,
                                             Umetrics, BestWood

Description:

On-line Fault Detection System allows the operator and management to rapidly detect the
location of the cause of the process shutdown. This would speed up down times from hours
and days to minutes. Tembec has implemented this type of fault detection system in one area
of their operations and have it has saved them at least $80,000/year due to higher
productivity.

On-line property prediction system allows operators a new sensor to predict the laboratory
results in real-time, which help mill personnel to rapidly react to changing process conditions
which are impacting on key properties characteristics. Rapid response to changing
characteristics will reduce the off-specification final product which in turn will save industry
money in energy, raw materials and waste treatment.

This type of sensor can be used in any part of mill operation

Cost To Implement:                            $30,000 to $350,000 per area

Potential Industry Wide Energy Savings*:     250,000 to 500,000 GJ/year

Potential Mill Savings:                       $20,000 to 120,000/year per area

Current Status:

Report – CETC 2005-100 (TR)                   30                                        June 2005
This is still at the demonstration stage. More education and skilled personnel are required at
the mills and in some locations a new data collection infrastructure is needed to make this
work. This technology has a high potential of making a major impact to the industrial market.

Other Requirements:                         Real Time Data Network
                                            Good Database
                                            Multivariate Data Analysis Package

Hardware Requirements:                      PCs

Other Industries:                           Cement, Oil and Gas, Steel, Petro-Chemicals,
                                            Chemicals, Mining

Contact:                                    Effective Assets, Makrikon, Umetrics, Pavilion,
                                            Gensym, BestWood

2)      TMP Plant: Advanced Quality Control Solution for TMP

Developer:                                  Pacific Simulation

Description:

Pacific Simulation, a subsidiary of Metso Automation, has a total soft sensor package which
combined both first principal models and statistical multivariate models and a multivariate
control package which will typically reduce the specific energy requirements between 2.5 and
7% while improving the final product quality. In Canada most of the installations are done for
the product quality improvements. The Tembec newsprint mill in Kapuskasing has purchased
the package for its energy savings. They currently estimate an electrical energy reduction of
approximately 2 to 3 %.

Estimated Cost To Implement:                 $800,000 to $1,200,000

Potential Industry Wide Energy Savings*:    270,000- 675,000 GJ/year

Potential Mill Savings:                     $600,000 to $1,200,000/year

Current Status:

This is at the implementation stage. It is a commercial product available to pulp and paper
companies worldwide. This technology has a moderate impact to pulp and paper industry.
Canadian companies which have used this technology include: Bowater, Thunder Bay,ON,



Report – CETC 2005-100 (TR)                   31                                      June 2005
Abitibi-Consolidated, Stephenville, NF, Tembec, Kapuskasing,ON,               Papier Masson,
Buckingham, QC.

Other Requirements:                          Real-time data network

Hardware Requirements:                       PCs for monitoring purposes

Other Industries:                                   Not Applicable

Contact:                                     Pacific Simulation




4.3      Newsprint and Papermaking

Newsprint and papermaking operations are energy consuming operations. Newsprint is a paper
manufactured mostly from mechanical pulps specifically for the printing of newspaper.
Papermaking involves cooking of wood chips cooked with chemicals to release cellulose fibers and
dissolve lignin, then washed to remove impurities. Pulp is mechanically and chemically
treated to impart certain desired characteristics such as strength, smoothness and sizing.

The area for energy savings using soft sensors is the boiler house area, specifically boiler
houses which have hog fuel plants. Optimizing the boiler performance with new soft sensors
will provide economic savings. An overview of the newsprint mill process is shown in the
following figure:




Report – CETC 2005-100 (TR)                   32                                        June 2005
Groundwood Furnish
3 Block Storage
6 Primary Screen Tank
9 Cleaners


1 Logs
4 Grinders
7 Screens
10 Deckers


2 Debarking Drum
5 Bull Screen
8 Cleaners Feed Tank
11 Groundwood Storage




TMP Furnish
14 Chip Screening
18 Preheater
22 Screens


12 Chip Pile
15 APS Bin
19 Refiner
23 Disc Saveall


13 Chip Silo
16 Chip Washing




Report – CETC 2005-100 (TR)   33   June 2005
20 Pressure Cyclone
24 H.D. Storage




17 Impregnation Chamber
21 Latency Chest
25 TMP Furnish Tank




Recycle Furnish
29 Detrasher
34 Deaeration Chest
39 Disc Saveall


26 Magazines
30 Retention Chest
35 Forward Cleaners
40 Twin Wire Press


27 Newspapers
31 H D Cleaners
36 Lightweight Cleaners
41 Disperger


28 Batch Pulper
32 Coarse Screening
37 Screen Feed Chest
42 H.D. Storage




33 Floatation Chest
38 Screens
43 Recycle Pulp Chest




Report – CETC 2005-100 (TR)   34   June 2005
                              Figure 4 – Overview of the newsprint mill process




Report – CETC 2005-100 (TR)                          35                           June 2005
The finishing step of the Newsprint process is shown below:




21 Machine Chest
25 Dryer
29 Wrapper Dispenser


22 Head Box
26 Sheet Cooler
30 Tying Machines


23 Forming Wiring
27 Cutter Layboy
31 Stenciller


24 Press Section
28 Bale Press
32 Stacker




33 Transport




                       Figure 5 – Overview of the finishing step of the newsprint process


The following areas could benefit from soft sensor technology:

1)           Recaustization Area, Steam Plant, Evaporators Area, Brown Stock Washing,
             Bleach Plant, Pulp Machine, Paper Machine: On-line Fault Detection and
             Property Prediction System




Report – CETC 2005-100 (TR)                            36                                   June 2005
Developer:                                   Effective Assets, Matrikon, Gensym, Pavilion,
                                             Umetrics, BestWood

Description:

On-line Fault Detection System allows the operator and management to rapidly detect the
location of the cause of the process shutdown. This would speed up down times from hours
and days to minutes. Tembec has implemented this type of fault detection system in one area
of their operations and have it has saved them at least $80,000/year due to higher
productivity.

On-line property prediction system allows operators a new sensor to predict the laboratory
results in real-time, which help mill personnel to rapidly react to changing process conditions
which are impacting on key properties characteristics. Rapid response to changing
characteristics will reduce the off-specification final product which in turn will save industry
money in energy, raw materials and waste treatment.

This type of sensor can be used in any part of mill operation. The energy savings are due to
production losses due to shut-downs and start-ups.

Cost To Implement:                           $30,000 to $150,000 per area

Potential Industry Wide Energy Savings*:     250,000 to 500,000 GJ/year if the entire mill
                                             uses this technology

Potential Mill Savings:                      $20,000 to 120,000/year per area

Current Status:

This is still at the demonstration stage. More education and skilled personnel are required at
the mills and in some locations a new data collection infrastructure is needed to make this
work. This technology has a high potential of making a major impact to the industrial market.

Other Requirements:                          Real Time Data Network
                                             Good Database
                                             Multivariate Data Analysis Package

Hardware Requirements:                       PCs

Other Industries:                            Cement, Oil and Gas, Steel, Petro-Chemicals,
                                             Chemicals, Mining


Report – CETC 2005-100 (TR)                   37                                        June 2005
Contact:                                    Effective Assets, Makrikon, Umetrics, Pavilion,
                                            Gensym, BestWood

2)      Paper Machine: On-line Break Prediction System

Developer:                                  Matrikon, Gensym, Pavilion

Description:

On-line Break Detection System allows the operator and management to rapidly to predict if
the process may have a failure due to sheet break. There has been claims by one supplier that
the operations can predict up to 20 minutes in advance.

Cost To Implement:                          $150,000 to $250,000

Potential Industry Wide Energy Savings*:    50,000 to 100,000 GJ/year

Current Status:

This is still at the demonstration stage. More education and skilled personnel are required at
the mills and in some locations a new data collection infrastructure is needed to make this
work. This technology has a high potential of making a major impact to the industrial market.

Potential Mill Savings:                     $800,000 to 1,200,000/year

Other Requirements:                         real time data network

Hardware Requirements:                      PCs

Other Industries:                           Steel

Contact:                                    Makrikon, Pavilion, Gensym

3)      Recaustization Area, Utilities Area, Environment Control: On-line Prediction of
        the fuel BTU

Developer:                                  ProSensus, McMaster University

Description:

Dr. John McGregor at McMaster University has been working on this problem for the last 5
years. ProSensus is a new spin-off company from Dr. McGregor’s research at McMaster


Report – CETC 2005-100 (TR)                   38                                      June 2005
University. This firm has been established to provide industrial solutions using multivariate
image analysis. One of the potential applications of this technology is applying multivariate
image analysis on the flame within the lime kiln to predict the lime quality and liquor
strength. An additional benefit of this technology is its ability to predict the emission
pollutants as well. This would help mills in better control of their mills boiler efficiencies.

Another issue in Canadian pulp and paper industry is the use of biomass fuels to replace
natural gas. In a paper published in AiCHe, Dr. McGregor shows that this technique can also
be used to estimate the fuel BTU content. This information will also be very critical for
operator or control system to obtain the proper combustion parameters in the kiln burner.

Estimated Cost To Implement:                 $80,000 to 100,000

Potential Industry Wide Energy Savings*:     50,000 to 130,000 GJ/year

Potential Mill Savings:                      up to $50,000/year

Current Status:

This is at the demonstration stage. The R&D work has shown that this technology is feasible.
An industrial partner as well as a supplier is needed to implement this technology. This has a
high impact to all industries.

Other Requirements:                          Real Time Data Network

Hardware Requirements:                       Server and Monitors for Soft Sensor Server

Other Industries:                            Cement, Energy, Steel, Oil and Gas, Chemical,
                                             Petro-Chemical

Contact:                                     Prof. John McGregor, McMaster University

Phone:                                       (905) 525-9140 ext.24951
Email:                                       macgreg@mcmaster.ca



4)       Entire Mill: Energy Monitor

Developer:                                   American Process Inc

Description:

American Process Inc has an on-line energy monitoring package called “Energy Monitor™”.
This is a three part package which includes a neural net package which predicts the typical

Report – CETC 2005-100 (TR)                   39                                       June 2005
energy consumption based on a one year energy model. The intent of this package is to assist
mills to locate where potential energy upsets occur.

Estimated Cost to Implement:                  $80,000 to 100,000

Potential Industry Wide Energy Savings*:     80,000 to 200,000 GJ/year

Potential Mill Savings:                      $40,000 - $100,000/year

Current Status:

This is at the demonstration stage. Tembec, St. Francisville, LA, USA mill is applying this
technology.

Other Requirements:                          Real Time Data Network

Hardware Requirements:                       Server and Monitors for Soft Sensor Server

Other Industries:                            Cement, Energy, Steel, Oil and Gas, Chemical

Contact:                                     American Process Inc.




4.4       Current applications of sensors in Pulp and Paper in Canada

Freeness control and pulp quality prediction are the main application of soft sensor
technology in the P&P industry. Currently, the following mills are using this technology:

      1. Tembec Inc., Kapuskasing, Ontario for Specific Energy Control, Freeness Control,
         and Pulp Property Strengths

      2. Abitibi-Consolated, Stephenville, Newfoundland, Freeness Control and Pulp Property
         Stengths

      3. Papier Masson, Buckingham, Quebec, Freeness Control

      4. Tembec, Temiscaming, Quebec - On-line Fault Diagnostics for Waste Water
         Treatment, On-line Prediction of Specialty Cellulose Properties

      5. Tembec, Skookumchuck, British Columbia - Carry-Over Index for the Recovery
         Boiler, Steam Production for Hog Fired Boilers

      6. Canfor, Prince George, British Columbia – On-line Sensor for Lime Kiln Production



Report – CETC 2005-100 (TR)                   40                                     June 2005
5        Potential applications of soft sensors in the Cement industry

5.1      Process description




                        Figure 6 – Overview of the cement manufacturing process



Extraction of the raw material (limestone, shale, silica, and pyrite) is done by blasting or by
scraping with the use of shearer loaders. The material is delivered to the crusher where it is
reduced to chunks by crushing or pounding. Crushed limestone and the other raw materials
are often stored under cover to protect them from the elements and to minimize dust. In the
raw mill, the material chunks are ground finer to allow high-quality blending. In this phase
vertical mills are used which grind the material through pressure exerted by rollers and
horizontal mills which pulverize the material by means of steel balls.
By far the greatest part of the electrical energy demanded for grinding is not exploited for
comminution but rather converted into lost heat. It is therefore an economical demand to
adjust the grinding plant so that energy losses are kept as low as possible.The raw meal is
finally transported to the homogenization silo for storage and further material blending.
Calcination is the core portion of this process. The raw meal is continuously weighed and fed
into the top most cyclone of the preheater. The material is heated by hot air rising from the
kiln. Inside of huge rotary kilns the raw material is transformed into clinker at 1,450 degrees
Celsius. From the kiln, the clinker goes to the clinker cooler for heat recuperation and cooling.

Report – CETC 2005-100 (TR)                       41                                     June 2005
The cooled clinker is then transported by a pan conveyor to the clinker silo for storage.
Because of the high energy consumption of the calcination process, automation and
optimization play an essential role in this production stage. The main requirement for low
emission and low energy consumption is a highly uniform kiln operation. Therefore, the
burning process must be monitored continuously using modern process control technology
and soft sensors.

Clinker is extracted from the clinker storage and sent to feed bins for further proportioning
with gypsum and additives before passing the finishing mill. Clinker enters the finishing mill
to be ground into cement. This is a high energy consuming process.

5.2      Potential applications

The following areas could benefit form soft sensor technology:

1)      Clinker Production: On-line Prediction of the fuel BTU

Developer:                                    ProSensus, McMaster University

Description:

Dr. John McGregor at McMaster University has been working on this problem for the last 5
years. ProSensus is a new spin-off company from Dr. McGregor’s research at McMaster
University. This firm has been established to provide industrial solutions using multivariate
image analysis. One of the potential applications of this technology is applying multivariate
image analysis on the flame within the cement kiln. An additional benefit of this technology
is its ability to predict the emission pollutants as well. This would help cement plants in better
control of their boiler efficiencies.

Another issue in Canadian cement industry is the use of biomass fuels to replace natural gas.
In a paper published in AiCHe, Dr. McGregor shows that this technique can also be used to
estimate the fuel BTU content. This information will also be very critical for operator or
control system to obtain the proper combustion parameters in the kiln burner.

Estimated Cost To Implement:                   $80,000 to 100,000

Potential Industry Wide Energy Savings*:      50,000 to 130,000 GJ/year

Current Status:

This is at the demonstration stage. The R&D work has shown that this technology is feasible.
An industrial partner as well as a supplier is needed to implement this technology. This has a
high impact to all industries.



Report – CETC 2005-100 (TR)                    42                                         June 2005
Potential Mill Savings:                      up to $50,000/year

Other Requirements:                          Real Time Data Network

Hardware Requirements:                       Server and Monitors for Soft Sensor Server

Other Industries:                            Cement, Energy, Steel, Oil and Gas, Chemical,
                                             Petro-Chemical

Contact:                                     Prof. John McGregor, McMaster University

Phone:                                       (905) 525-9140 ext.24951
Email:                                       macgreg@mcmaster.ca



2)       Clinker Production: On-line Prediction of 1 Day Strength

Developer:                                   Malvern Process Systems

Description:

A multivariable regression was made with fineness variables from a Insitec Blaine # and
chemical composition variables. They were able to show that if the Cement mill was at full
production, they could increase production capacity by 5%. If the mill was not production
limited, there was the potential of reducing the fineness of the clinker by about 100 cm2/g
which could have an energy reduction between of 2 to 3%.

Estimated Cost To Implement:                 Not Known

Potential Industry Wide Energy Savings*:     Not Known

Potential Mill Savings:

Current Status:

This is at the demonstration stage. This has a high impact to cement industry.

Other Requirements:                          Real Time Data Network

Hardware Requirements: Server and Monitors for Soft Sensor Server, Malvern Analyzer

Other Industries:                            Cement,

Contact:                                     Malvern Process Industries


Report – CETC 2005-100 (TR)                   43                                     June 2005
3)      All Areas: On-line Fault Detection and Property Prediction System

Developer:                                   Effective Assets, Matrikon, Gensym, Pavilion,
                                             Umetrics

Description:

On-line Fault Detection System allows the operator and management to rapidly detect the
location of the cause of the process shutdown. This would speed up down times from hours
and days to minutes. Tembec has implemented this type of fault detection system in one area
of their operations and have it has saved them at least $80,000/year due to higher
productivity.

On-line property prediction system allows operators a new sensor to predict the laboratory
results in real-time, which help mill personnel to rapidly react to changing process conditions
which are impacting on key properties characteristics. Rapid response to changing
characteristics will reduce the off-specification final product which in turn will save industry
money in energy, raw materials and waste treatment.

This type of sensor can be used in any part of mill operation

Cost To Implement:                            $30,000 to $150,000 per area

Potential Industry Wide Energy Savings*:     25,000 to 63,000 GJ/year if the entire mill uses
                                             this technology

Potential Mill Savings:                      $20,000 to 120,000/year per area

Current Status:

This is still at the demonstration stage. More education and skilled personnel are required at
the mills and in some locations a new data collection infrastructure is needed to make this
work. This technology has a high potential of making a major impact to the industrial market.

Other Requirements:                          Real Time Data Network
                                             Good Database
                                             Multivariate Data Analysis Package

Hardware Requirements:                       PCs

Other Industries:                            Cement, Oil and Gas, Steel, Petro-Chemicals,
                                             Chemicals, Mining

Contact:                                     Makrikon, Pavilion, Gensym,


Report – CETC 2005-100 (TR)                   44                                        June 2005
4)       Clinker Area: Prediction of Lime Kiln Ringing issues

Developer:                                 University of Toronto Pulp and Paper Centre

Description:

Teresa D’Souza, a master student, under the supervision of Dr. Joe Ripka and Dr. Honghi
Tran at the University of Toronto’s Pulp and Paper Centre is currently doing research in
predicting what the internal temperature conditions are which causes a lime kiln to lose its
production efficiency by creating ring formation within a lime kiln. This lack of efficiency
will cause mill operations to use more energy to compensate for this problem. If the ring
formation is serious, then the production must take downtime to correct the problem. If this
sensor can be commercialized, it would give operators advance notice of this condition, and
prevent ring formation.

Estimated Cost To Implement:                Not Known

Potential Industry Wide Energy Savings*:   Not Known

Other Requirements:                        Real-time data network
                                           Process database
                                           Soft Sensor Server

Hardware Requirements:                     Server and Monitors for Soft Sensor Server

Other Industries:                          Cement

Contact:                                   Prof. Honghi Tran , University of Toronto, Pulp
                                           and Paper Centre

Phone:                                     (416) 978-8585
Fax:                                       (416) 971-2106
Email:                                     tranhn @ chem-eng.utoronto.ca




Report – CETC 2005-100 (TR)                 45                                      June 2005
6        Industry survey
A survey was performed to obtain the opinion of different experts from the industry, such as
developers, suppliers and end-users of soft sensor applications. The following persons have
been interviewed:

    Ø   Pulp and Paper Industry

        Blair Rydberg, Mill Manager, Tolko Industries, Kraft Papers, La Pas, Manitoba

        Richard Adderly, Process Engineer, Tembec Industries, Skookumchuck,BC

        Bernard Lupien, Process Control Engineer, Papier Masson, Buckingham, QC

    Ø   Cement Industry

        Michel Dufault, area controller, Canada Lafarge Cement, Concord, ON


    Ø   Suppliers

        Warren Mitchell, Matrikon Inc, Edmonton, AB

        Dr. Svante Wold, Umetrics Inc., Kinnelon, NJ

    Ø   Research & Development

        Dr. Tranh Trung, Pulp and Paper Research Institute of Canada, Vancouver, BC

        Dr. John McGregor, McMaster University, Hamilton, ON

        Dr. Joe Kipka and Teresa D’Souza, Univerisity of Toronto, Toronto, ON

Appendix C summarizes the interviews with the various participants. The general sense is that
the pulp and paper industry in Canada are now actively looking at applying soft sensor
technology between now and 2006. In case of the cement industry, there is little demand for
soft sensors.




Report – CETC 2005-100 (TR)                  46                                         June 2005
7        Key Areas of future R&D

7.1      Development of multi-grade models

When a paper mill produces 80 different grades of papers, it is possible that 80 different
models are required. Better modeling strategies should be developed, so one model can
accommodate multiple product specifications, thus eliminating the need for individual models
for each product grade.

7.2      Development of adaptive models able to cope with production changes
         and shutdowns

Often a mill does a maintenance shutdown, equipment is replaced which will change the
production characteristics. Since the models are based on the previous process data, problems
will arise from the new production characteristics. Researchers have been working in the
adaptive models for product characteristic predictions, but more work is needed in this area.

7.3      Increasing the awareness of plant managers and operators

The key problem today is the lack of knowledge of what is soft sensor technology from the
typical engineering student to engineering suppliers to mill production personnel. The
development of a soft sensor algorithm is heavily based on identifying patterns and
relationships between process variables, thus proving an increased knowledge of the process.
Operators and managers should be educated to realize the full potential of applying this newly
gained knowledge toward improving their operations, by changing how they operate or build
their process.

7.4      Development of user-friendly soft sensor software

Soft sensor technology must be an easy to use technology. Not only the inferring algorithm
has to be accurate and reliable, but the on-line interface should allow operators to easily
understand the state of the process so rapid detection and correction of abnormal situations is
possible. Hybrid systems combining statistical and artificial intelligence techniques are a
promising avenue, worth investigating.

7.5      Recommendations on future work

The work on multivariate imaging sensors to predict product quality under the leadership of
Dr. John McGregor should be supported. This technology has the potential of being applied in
almost every industry to help reduce energy consumption and environmental emissions. The
R&D analysis work has been completed for the image analysis of flames to predict product
characteristics. This technology needs financial support as well as industrial test site to
demonstrate that this technology is feasible

Report – CETC 2005-100 (TR)                   47                                       June 2005
The NIR soft sensor technology that Paprican is developing should be given further support to
develop new applications to improve process productivity which will reduce energy
consumption and environmental emissions. Paprican has done extensive laboratory testing of
their NIR soft sensor models, and are now in the process of demonstrating that these
technologies can work in the industrial process. Paprican needs additional financing to
develop more NIR soft sensor technology. Currently Paprican has hired Dr. Lars Wallback
from Sweden to assist their NIR soft sensor technology in developing more NIR based
sensors to improve the Canadian industry.




Report – CETC 2005-100 (TR)                  48                                      June 2005
8        Conclusions
The major impact of soft sensors will be increasing the productivity of Canadian industries.
This translates into significant energy consumption and green house gas emission reductions,
since less energy will be required for a unit of production. Also, a soft sensor can estimate,
on-line, critical process data, thus allowing rapid detection and correction of abnormal
situations. This also translates into important energy savings associated with a decreased
production of out of specification product. Soft sensors are cross-cutting industrial
applications, since they can be use to optimize the operation of equipments used in multiple
industries: for example, soft sensors could have a significant direct impact on energy
reduction for processes such as boilers, present in a variety of industrial processes.

The potential economic impact of applying a soft sensor based control package for both Pulp
and Paper and Cement Industry is a cost reduction between $75 and 135 million dollars per
year if it was 100% successful for every mill.

Soft sensors’ greatest initial impact will be in productivity gains for both pulp and paper and
the cement industries.

The pulp and paper industry is starting to take advantage of soft sensors, while the cement
industry appears to be resisting. In the case of the cement industry part of the resistance is due
to the lack of qualified manpower at their various small cement operations.

Currently the production personnel must control a process based on various flows, pressures,
concentrations and we hope that we will get the desired results at the end of the process. The
goal should be obtaining more process understanding and redesign the process so that our
production would simply consist in dialing in the desired product characteristics and the
computer control system will produce the desired product.

The literary review, as well as the industry survey showed that most soft sensor applications
in process industries are based on process historical data-based approaches. This is due to the
fact these approaches are easy to implement, requiring very less modeling effort and a priori
knowledge than first principle models.

The literary analysis also showed a clear trend toward combining statistical techniques (such
as PCA/PLS) with artificial intelligence techniques (such as neural networks). The results of
the examined studies showed the superior performance of such a hybrid system versus soft
sensors based solely on one of the either methods. User-friendly commercial software
packages offering such a combination are not readily available. Therefore, the development of
a soft sensor based on a PCA/PLS/NN combination, as well as of a software tool allowing its
design and implementation is a promising area worth investigating.



Report – CETC 2005-100 (TR)                     49                                        June 2005
The resistance of applying soft sensor technology must be reduced by providing more
demonstration sites in Canada. We need to educate our future scientists as well as engineers to
the potential of soft sensors, so that they will demand the use of them in industry. Our
industrial leaders need to be educated why soft sensor technology will improve their bottom
line in this highly global competitive world.

Taking full advantage of the Soft Sensor technology would result in:


    1. improved process control

    2. improved process knowledge

    3. improved operational modifications towards optimizing the process

    4. improved product quality

    5. creating new products




Report – CETC 2005-100 (TR)                   50                                       June 2005
                                   Annex A
                              Bibliographical list




Report – CETC 2005-100 (TR)            51            June 2005
Artificial Intelligence … within
Bartos, F., 2003 Control Engineering, September issue (also available at
www.manufacturing.net)

Neural Network technology moves toward mainstream
Finchem, K., 1998 Paperloop, vol. 72, issue 4 (www.paperloop.com)

Improving pulp and paper process diagnostics and knowledge by means of multivariate
analytical techniques
Wold S. et al., 2003 Pulp & Paper Canada, vol. 104, issue 5

Bayesan neural networks on the inference of distillation product quality
Hall Barbosa C., Vellasco M., Melo B., Pacheco M. and Vasconcellos L., 2002 VII Brazilian
symposium on neural networks

Application of feedforward neural networks for soft sensors in the sugar industry
Devogelaere, D., Rijckaert M., Osvaldo G. and Lemus G. , 2002 IEEE, September issue

An adaptive neuro-fuzzy inference system as a soft sensor for viscosity in rubber mixing
process
Merikoski, S et al. (Automation and Control Institute, Tampere University of Technology)

Long-term energy demand predictions based on short-term measured data
Olofsson T., Anderson S., 2001 Energy and Buildings, vol. 33, pp. 85-91

Artificial neural networks for the prediction of the energy consumption of a passive solar
building
Kaloogiru, S. and Bojic M., 200, Energy, vol. 25, pp. 479-491

Intelligent lime kiln control system
Jarvensivu, M., Saari K. and Jamsa-Jounela S., 2001 Control Engineering Practice, vol. 9, pp.
589-606

Achieving Consistent product quality, increased throughput and minimal use of resources
From the software maker Gensym, based on their experience and industrial implementations
of their products (www.gensym.com)

Soft sensor development for on-line bioreactor estimation
Jose de Assis, A. and Filho, R., 200 Computers and chemical Engineering, vol. 24, pp. 1099-
1103

Pattern recognition of gases of petroleum based on RBF model
Barbosa M., Ludermir T. and Santos M., 2002, VII Brazilian Symposium on Neural Networks


Report – CETC 2005-100 (TR)                   52                                       June 2005
Identification of neural dynamic models for fault detection and isolation: the case of a real
sugar evaporation process
Patan K. and Parisi T., 2005 Journal of Process Control, vol. 15, pp. 67-69

Optimal selection of soft sensor inputs for batch distillation columns using principal
component analysis
Zamprogna E., Barolo M. and Seboprg D., 2005 Journal of Process Control, vol. 15, pp. 39-
52

Use of artificial neural networks process analyzers: a case study
Duwaish A., Halawani L. and Mohandes M., 2002, European Symposium on Artificial Neural
Networks 2002, proceedings pp. 465-470

The application of neural networks to the paper-making industry
Edwards P., Murray G., Papadoppoulous G. and Wallace A., 1999 European Symposium on
Artificial Neural Networks 2002, proceedings pp. 69-74

A review of process fault detection and diagnosis, Part I: Quantitative model-based methods
Venkatasubramanian V., Rengaswamy R., Yin K. and Navuri S., 2003 Computers and
Chemical Engineering, vol. 27, pp. 293-311

A review of process fault detection and diagnosis, Part II: Qualitative models and search
strategies
Venkatasubramanian V., Rengaswamy R., Yin K. and Navuri S., 2003 Computers and
Chemical Engineering, vol. 27, pp. 313-326

A review of process fault detection and diagnosis, Part III: Process history based methods
Venkatasubramanian V., Rengaswamy R., Yin K. and Navuri S., 2003 Computers and
Chemical Engineering, vol. 27, pp. 327-346

Modelling energy consumption in manufacturing industry
Newton J. K., 1985 European Journal of Operational Research, vol. 19, pp. 163-169

Adaptive neural networks for intelligent operation of the activated sludge process
Barnett W., 1997 Water Environment Federation Specialty Conference: Computer
Technologies for the Competitive Utility, proceedings

Soft sensors for product quality monitoring in debutanizer distillation columns
Fortuna L., Graziani S. and Xibilia M., 2005 Control Engineering Practice, vol. 13, pp. 499-
508

A self-validating inferential sensor for emission monitoring
Qin S., Yue H. and Dunia R., 1997 AACC vol. 9


Report – CETC 2005-100 (TR)                    53                                        June 2005
Applications of artificial neural networks in energy systems: a review
Kalogiru, S, 1998 Energy Conversion and Management, vol. 40, pp. 1073-1087

Process control’s latest tools: Soft sensors
Harrold D., 2001 Control Engineering, June issue

Application of Artificial Intelligence technology to Increase Productivity, Quality and Energy
Efficiency in Heavy Industry
Szladow A., Lobbe Technologies ltd., 1995 report prepared for CETC-Ottawa




Report – CETC 2005-100 (TR)                   54                                       June 2005
                                   Annex B
                              Article summaries




Report – CETC 2005-100 (TR)          55           June 2005
                               Table 4 – Summaries of articles reviewed



Title                   Application                Method                 Implementation
                                                                          status

Use of artificial       Value prediction: O2      NN: feedfoward, back propagation
neural networks         analyzer (soft sensor) to
process analyzers: a    predict O2 contents in
case study              the boiler flue gas from
                        other measured
                        variables
Bayesian neural         Value prediction: infer NN: back propagation multi-layer
networks on the         the quality of distillation perceptron and Bayesian neural networks
inference of            products, for the
distillation product    REPAR refinery in
quality                 Brazil.
Application of          Value prediction: infer    NN: feedfoward,        NN models have been
feedforward neural      the electrical             back propagation       already implemented
networks for soft       conductivity of the                               in the cane sugar
sensors in the sugar    mixture of crystals and                           manufacturing
industry
                        molasses during the in                            industry, namely
                        the crystallization                               evaporation and
                        process                                           crystallization.
Pattern recognition   multiple-class               NN: Radial Based Function (RBF)
of gases of petroleum classification (pattern
based on RBF model recognition): recognize
                        and classify ethane,
                        methane, propane,
                        butane and carbon
                        monoxide
Soft sensor             Value prediction:          NN: feedfoward,        feedforward
development for on-     prediction of certain      back propagation       backpropagation NNs
line bioreactor state   biomass measurements                              are already applied in
estimation                                                                the biotechnical
                                                                          industry
An adaptive neuro-      Value prediction: infer Neuro-fuzzy               Tested off-line using
fuzzy inference         rubber viscosity using 3 inference system         data from a Finish tire
system as a soft        process parameters.      (ANFIS)                  manufacturer
sensor for viscosity in
rubber mixing
process




Report – CETC 2005-100 (TR)                       56                                         June 2005
Title                    Application                  Method                Implementation
                                                                            status

Modeling energy          Predictive modeling:         multivariate statistical analysis
consumption in           model energy
manufacturing            consumption
industry
Soft sensor for          Value prediction:            NN: multi-layer       Implemented in a
product quality          estimate the gasoline        perceptron (MLP)      refinery in Italy; on-
monitoring in            and butane                                         line performance
debutanizer              concentration of a                                 judged satisfactory;
distillation column
                         debutanizer column.                                main advantage is
                                                                            overcoming the
                                                                            relatively long time
                                                                            delayed introduced by
                                                                            the gas
                                                                            chromatograph
Optimal selection of     Value prediction:            NNs, linear PLs and non-linear PLS;
soft sensor inputs for   estimate product
batch distillation       composition using            PCA analysis used to select the most
columns using            temperature values as        significant inputs
principal component
                         inputs
analysis

Artificial neural        Value prediction:            NN: Jordan Elman recurrent network
networks for the         estimate the building’s
prediction of the        hourly energy
energy consumption       consumption
of a passive solar
building
Long-term energy         Value prediction:            NN: feedforward, back propagation
demand predictions       estimate the building’s
based on short-term      hourly energy
measured data            consumption
The application of       Classification (different    classification: NN, Tested off-line using
neural networks to       specifications, of the       multi-layer         data from a paper mill
the paper-making         curl level); value           perceptron (MLP)
industry                 prediction (estimate the     prediction: NN,
                         absolute level of the        MLP
                         curl)
                                                      PCA analysis used
                                                      to select the most
                                                      significant inputs




Report – CETC 2005-100 (TR)                          57                                       June 2005
Title                   Application                  Method               Implementation
                                                                          status

Intelligent lime kiln   Value prediction:       NN: feedfoward,           Implemented at a
control system          estimate residual CaCO3 back propagation          pulp & paper mill in
                        content of the burned                             Finland. Recorded
                        lime.                                             benefits are:
                                                                          improved lime kiln
                        The soft sensor is also
                                                                          efficiency (increased
                        used as a part of the fan
                                                                          throughput), decrease
                        speed feedforward
                                                                          of variations in lime
                        control and feedback
                                                                          quality, and decrease
                        adjustment
                                                                          energy consumption
Adaptive neural         Data monitoring: signal Data monitoring : NN, auto-associative
network for             filtering, sensor         and back propagation
intelligent operation   validation.
of the activated                                  Sensor validation: auto-associative NNs.
sludge process          Data filtering and sensor Value prediction: feedforward back
                        validation.               propagation NNs. Single-class
                        Single-class                 membership classification: Rho neural
                        membership                   networks
                        classification.              Multiple-class membership classification:
                        Multiple-class               Radial Basis Function (RBF) networks
                        membership                   State assessment: detect faults or detect
                        classification               and diagnose faults: Rho network in
                        (sensor/process fault        conjunction with a RBF network
                        diagnosis).
                                                 Control: back propagation feedforward
                        State assessment: detect NN
                        faults or detect and
                        diagnose faults. 2
Improving pulp and      Value prediction:       PCA/PLS
paper process           predictive modeling,
diagnostics and         inferring properties of
knowledge by means      pulp and paper, such as
of MVA analytical
                        kappa number, viscosity
techniques
                        and strength
A self-validating      Value prediction: boiler PCA/NN
inferential sensor for emissions
emission monitoring


2
         A case study of NN configurations to identify appropriate applications area for NNs in
         the development of an advanced control system/automation of the activated sludge
         processes.


Report – CETC 2005-100 (TR)                         58                                       June 2005
Title                   Application           Method           Implementation
                                                               status

Applications of         Value prediction:     Backpropagation NN
artificial neural       building energy
networks in energy      consumption, steam
systems: a review       production




Report – CETC 2005-100 (TR)                  59                                 June 2005
                                     Annex C
                              Interview Summaries




Report – CETC 2005-100 (TR)           60            June 2005
Interviewee:              Blair Rydberg

Location:                 Tolko Industries Manitoba Kraft Papers, La Pas, MB

    1) Are you or do you intend to use Soft Sensors?

        Currently, we are not using soft sensors. We have the intention of implementing
        Paprician’s NIR based soft sensors for predicting Kappa Number, liquor quality, wood
        chip moisture and bark moisture beginning in the latter half of 2005 or early 2006.

    2) Why did you wait so long to get these sensors?

        The key difficulty was obtaining the necessary funding to purchase this equipment.




Report – CETC 2005-100 (TR)                    61                                       June 2005
Interviewee:              Richard Adderly

Location:                 Tembec Industries, Skookumchuck, BC

    1) Are you or do you intend to use Soft Sensors?

        I have developed an on-line carry-over sensor for the recovery boiler and I am
        currently developing a soft-sensor to predict steam flow of our hog fuel boiler to get a
        better understanding to how this boiler operates.

        We are hoping to purchase a lime kiln MPC package from Emerson which will
        include a neural net based soft sensor to predict the calcium carbonate content. Also
        we looking to purchase in fiscal 2006 the Paprican developed NIR based soft sensor to
        predict and help control white liquor quality and the NIR based moisture sensor for
        the wood chip moisture and hog fuel.

    2) What are you estimating the economic returns on the NIR based soft sensor to predict
        and help control the white liquor quality?

        We are estimating a 1 year return of investment, where we think 25% of the savings
        will be on energy reduction in the recovery cycle, while the remaining 75% in higher
        productivity gains.

    3) What are you estimating the energy savings on the lime kiln MPC package with soft
        sensor?

        We are estimating a reduction of between 5 and 7% of our Natural Gas Fuel
        consumption.

    4) Would like to make any other comments regarding Soft Sensors and their
        implementation?

        Tembec does believe in using soft sensor technology, however there is still a great
        need to educate both the operators and senior management in applying these tools. We
        still do not have the necessary manpower to support this technology.



Report – CETC 2005-100 (TR)                    62                                       June 2005
Interviewee:              Bernard Lupien

Location:                 Papier Masson, Buckingham, QC

    1) Are you or do you intend to use Soft Sensors?

        We are currently installing the AQC product from Pacific Simulation to better control
        and predict our pulp characteristics. We are talking with Paprican to see if we can
        develop an on-line sensor to measure further paper properties using a NIR instrument
        we currently which is measuring the consistency of the pulp.

    2) What were the issues that are facing you in applying soft sensors?

        The lack of qualified manpower from both the supplier side and from the mill
        personnel in applying these techniques.

    3) What is your rate of return?

        We are looking for a 2 year return on investment




Report – CETC 2005-100 (TR)                   63                                     June 2005
Interviewee:              Michel Dufault

Location:                 Eastern Canada Region Head Office

    1) Do you intend or use Soft Sensors in your organization?

        Currently we do not use this technology.

    2) Why are there difficulties in applying soft sensors?

        The primary reason is the lack of skilled local mill manpower to implement or
        maintain these types of systems. Also we do not the necessary infrastructure in our
        operations, since they are relatively small in size.

    3) What is your desired rate of return?

        We use an internal rate of return of 9%.




Report – CETC 2005-100 (TR)                      64                                 June 2005
Interviewee:              Warren Mitchell

Location:                 Matrikon Inc, Edmonton, AB

    1) Do you think Pulp and Paper and Cement Industries are accepting Soft Sensor
        Technology?

        Pulp and paper are rapidly looking this technology and we have helped Abitibi-
        Consolidated in Canada in applying this technology. There is no interest from Cement
        Industry.

    2) Does Process Doctor has soft sensor capability?

        Yes, it does. Many industries are now using this product to improve their bottom line.




Report – CETC 2005-100 (TR)                   65                                       June 2005
Interviewee:              Dr. Svante Wold

Location:                 Umetrics Inc., Kinnelon, NJ

    1) What is the acceptance of Pulp and Paper Industry to soft sensors?

        Tembec is a prime example of accepting and applying soft sensors. There are a
        number of pulp and paper companies from around the world who are now using
        multivariate techniques for prediction.

    2) What is the acceptance of the Cement Industry to soft sensor?

        We did some work with a Danish cement firm looking at NIR spectra to better control
        the kiln, but nothing came about it. There has been no demand from the cement
        industry in applying this technology.




Report – CETC 2005-100 (TR)                       66                               June 2005
Interviewee:              Dr. Tranh Trung

Location:                 Paprican, Vancouver, BC

    1) Where do you see the application of Soft Sensors in the Pulp and Paper Industry?

        We are doing a large amount of work in developing NIR based soft sensors for Kraft
        Liquor Analysis, Kappa Number Prediction, and Moisture Content for wood chips and
        bark.

    2) Is there a demand from Paprican Members?

        Yes, there is a strong demand from our members. We are looking at assisting Tembec
        Industries in Skookumchuck, BC, Smooth Rock Falls, ON as well as Marathon Pulp
        and Weyerhauser in applying our moisture soft sensor.




Report – CETC 2005-100 (TR)                   67                                     June 2005
Interviewee:              Dr. John McGregor

Location:                 McMaster University, Hamilton, ON

    1) What is the state of Soft Sensor Implementation for Industry in general?

        Soft sensors were beginning to see a major surge just prior to the internet bubble
        crash. I am now seeing that we are starting to demand beginning to come back from
        the crash in all industries.

    2) Has there been any interest in Canadian industry in applying your vision based soft
        sensor to predict product characteristics?

        The initial project was to be done with Dupont Canada, until they restructured their
        organization and the project has put on hold. What I noticed that I need to team up
        with either a burner supplier or instrumentation manufacturer to help this technology
        to the end user. Most end users are not interested. We are now in discussions with a
        Danish company to use this technology.

    3) Can this technology be use to also detect emissions such as NOx, SO2, etc?

        This technology has been shown to predict NOx and SO2 from looking at the flame in
        the latent variable space.




Report – CETC 2005-100 (TR)                     68                                   June 2005
Interviewees:             Dr. Joe Ripka and Teresa D’Souza

Location:                 University of Toronto, Toronto, ON

    1) Has your research shown that you can detect the conditions between a good lime kiln
        vs a lime kiln which has a ringing problem?

        Yes, we have developed a PLS-DA analysis which classifies a good lime kiln vs a
        lime kiln which has ringing. We have shown with the data sets that we have, the kiln
        will shift from the good mode to the bad mode.

    2) What are problems are you facing in creating your model?

The key problem is the lack of process data from the mills. Also we are noticing that we are
getting shifts between different kilns and we need to find a way to eliminate this problem. We
are having problems in trying to determine the root cause of the problem. However we see
that this tool can used as an on-line sensor to warn mill personnel of potential problems.




Report – CETC 2005-100 (TR)                    69                                        June 2005

				
DOCUMENT INFO
Shared By:
Stats:
views:114
posted:9/1/2010
language:English
pages:79