Chiller Fault Detection
and Diagnosis (FDD)
Paul Riemer
June 20, 2000
ECE/CS/ME 539 Semester Project
Instructor: Prof. Y.H. Hu
(a little piece of my MS research project)
What is FDD?
A process of comparing quantities that
characterize a system’s actual
performance against their baseline values
to determine deviation from accepted
behavior and to identify which of the
system’s components are responsible and
how so.
Chiller Basics
Large Scale Air Conditioning Equipment
Cools Water To Be Piped Around Building
Vapor Compression Cycle
Uses Refrigerant such as CFC, HCFC, NH3
Large Energy Demands: Mostly Electrical
Compressor (Centrifugal, Reciprocal, Screw, Scroll)
Water Pumps
Chiller Schematic
Cooling Tower
Condenser
(Shell and Tube HX) 2
3
Expansion Centrifugal
Device Compressor
4 Evaporator 1
(Shell and Tube HX)
Air Handlers
&
GPMCW
TCWS Condenser TCWR
TCON
D T2
POWER
Expansion Centrifugal
Device Compressor
Evaporator
TEVAP
TCHWS TCHWR
GPMCHW
Independent Monitored Quantities
GPMCHW - Chilled Water Flow Rate
TCHWS - Chilled Water Supply Temp
TCHWR - Chilled Water Return Temp
GPMCW - Condenser Water Flow Rate
TCWS - Condenser Water Supply Temp
(a.k.a. Forcing Inputs)
Dependent Monitored Quantities
TCWR - Condenser Water Return Temp
TCOND - Condenser Saturation Temp
TEVAP - Evaporator Saturation Temp
T2 - Compressor Exiting Temp
Power - Electric Power Draw
Characteristic Quantities (CQs)
Evaporator Condenser Others
Heat Transfer Heat Transfer Isentropic
UA UA Efficiency
Approach Approach Motor
TCHWS-TEVAP TCOND-TCWR
Efficiency
CHWDT CWDT COP
TCHWR - TCHWS TCWR - TCWS
FDD Process 1
Neural Data
Network Predicted
Reduction CQs
Predictor Code No Fault!
Forcing Dependent Fault
Inputs Quantities Classifier
Data Fault X!
Physical
Reduction Actual
Chiller
Code CQs
Remedy?
FDD Process 2
Neural
Predicted
Network
CQs
Predictor
No Fault!
Forcing Dependent Fault
Inputs Quantities Classifier
Data Fault X!
Physical
Reduction Actual
Chiller
Code CQs
Remedy?
Fault Classifier & End Goal
Compare actual and predicted CQs
Comparison criteria from a detail
thermodynamic model of a chiller
Actual CQ1 2% Above Predicted CQ1
Actual CQ2 5% Below Predicted CQ2 } =No Fault
Actual CQ1 15% Above Predicted CQ1
Actual CQ2 10% Below Predicted CQ2 } =Fault X
Neural Network Predictor
All Approaches
5 Independent Quantities as Inputs
Feed Forward Multi-layer Perceptron
Created and Trained using Matlab Toolbox
Linear Activation Function
Fault Free Data Set - April
Neural Network Predictor
FDD Process 1
5 Dependent Quantities as Outputs
Approach 1 - 1 Network w/5 Outputs
Approach 2 - 5 Networks each w/1 Output
FDD Process 2
Approach 3
1 Network w/11 CQ’s as Outputs
Data, Valuable Data
Available
4 identical chillers for cooling season
10 monitored quantities on 1-minute interval
Utilization
Trimmed non-operating data
Trimmed to expand interval between points
April as fault free training and testing data
July as potential faulty data for FDD
Results
Approach 1 = Run 34
Approach 2 = Runs 51-55
Approach 3 = Run 74
Part 1
Approaches 1 & 2 and Actual Values
Plots of 5 Dependent Quantities
Condenser Water Return Temp
89
87
85
april
83 april34
F
april51-55
81
79
77
88
87
86
85
july
84 july34
F
july51-55
83
82
81
80
Condenser Saturation Temp
91
90
89
88
87
april
86 april34
F
april51-55
85
84
83
82
81
91
90
89
88
87
july
86 july34
F
july51-55
85
84
83
82
81
Evaporator Saturation Temp
44
42
40
april
38 april34
F
april51-55
36
34
32
46
44
42
july
40 july34
F
july51-55
38
36
34
Compressor Exiting Temp
139
137
135
april
133 april34
F
april51-55
131
129
127
139
137
135
july
133 july34
F
july51-55
131
129
127
Electric Power Draw
1300
1200
1100
1000 april
KW
april34
900 april51-55
800
700
600
1300
1200
1100
1000
900 july
KW
july34
800 july51-55
700
600
500
400
Results Continued
Approaches 1 & 2 not significantly
different
Approach 1 results converted to CQ’s by
EES data reduction code
Part 2
Approaches 1 & 3’s CQ’s vs actual values
Plots of 11 CQ’s
Evaporator Heat Transfer Rate
2000
1800
1600
1400 april
Tons
april34
1200 april74
1000
800
600
1800
1600
1400
july
Tons
1200 july34
july74
1000
800
600
Evap. Conductance Area Product
4.40E+06
3.90E+06
3.40E+06
Btu/hr-F
april
2.90E+06 april34
april74
2.40E+06
1.90E+06
1.40E+06
4.00E+06
3.50E+06
3.00E+06
2.50E+06
Btu/hr-F
july
2.00E+06 july34
july74
1.50E+06
1.00E+06
5.00E+05
0.00E+00
Evaporator Approach
12
10
8
april
6 april34
F
april74
4
2
0
8
7
6
5
july
4 july34
F
july74
3
2
1
0
Chilled Water Temp Difference
10
9
8
7 april
april34
F
6 april74
5
4
3
12
10
8
july
6 july34
F
july74
4
2
0
Condenser Heat Transfer Rate
2400
2200
2000
1800
april
Tons
1600 april34
april74
1400
1200
1000
800
2500
2000
1500
july
Tons
1000 july34
july74
500
0
-500
Cond. Conductance Area Product
7.00E+06
6.00E+06
5.00E+06
Btu/hr-F
april
4.00E+06 april34
april74
3.00E+06
2.00E+06
1.00E+06
7.00E+06
6.00E+06
5.00E+06
Btu/hr-F
july
4.00E+06 july34
july74
3.00E+06
2.00E+06
1.00E+06
Condenser Approach
6
5
4
april
3 april34
F
april74
2
1
0
6
5
4
july
3 july34
F
july74
2
1
0
Condenser Water Temp Difference
7.5
7
6.5
6
5.5
april
5 april34
F
april74
4.5
4
3.5
3
2.5
7
6
5
4 july
july34
F
3 july74
2
1
0
Compressor Isentropic Efficiency
0.7
0.68
0.66
0.64
0.62
Efficiency
april
0.6 april34
april74
0.58
0.56
0.54
0.52
0.5
0.7
0.68
0.66
0.64
0.62
Efficiency
july
0.6 july34
july74
0.58
0.56
0.54
0.52
0.5
Motor Efficiency
1.1
1
0.9
Efficiency
april
0.8 april34
april74
0.7
0.6
0.5
1.1
1
0.9
Efficiency
july
0.8 july34
july74
0.7
0.6
0.5
Overall Coefficient of Performance
6
5.5
Coefficient of Performance
5
april
april34
april74
4.5
4
3.5
7
6.5
6
Coefficient of Performance
5.5
5
july
4.5 july34
july74
4
3.5
3
2.5
2
Conclusions
Approaches 1 & 3 quite similar
Training Set Predictions (April Data)
Good Matches: QEVAP, DTCHW, QCOND,
DTCW
So-so Matches: APPREVAP, UAEVAP, NISEN,
NMOTOR, COP
Bad Matches: APPRCOND, UACOND
Conclusions Continued
Actual FDD Predictions (July Data)
Acceptable: QEVAP, DTCHW, QCOND,
DTCW, UAEVAP, APPREVAP,
Irrelevant: UACOND, APPREVAP
(recall training set prediction not acceptable)
Interesting and worth further study:
NISEN - decreased compressor efficiency?
NMOTOR - increased motor efficiency?
COP - increased overall performance?
End Notes and Beyond
Three approaches performed equally well
on April Training Data
Prediction worked on about half of CQ’s
Future work as part of thesis project
Modify Network Configurations
Utilize More Data
Training
FDD Prediction