An Overview of Selected Prognosis Technologies with Reference to

Document Sample
An Overview of Selected Prognosis Technologies with Reference to Powered By Docstoc
					An Overview of Selected Prognosis
Technologies with Reference to an
   Integrated PHM Architecture




  Roemer, M.J., Byington, C.S., Kacprzynski, G.J.
             and Vachtsevanos, G.
Prognosis Technologies at ALL Levels
           Prognosis Technical Approaches
                          Applicability

                                                              Algorithms

                  Reliability and Usage Based             Usage-Based
                            Algorithms                     Fleet-Wide
Cost & Accuracy
  Increasing




                      Pattern Recognition/
                         Fuzzy Logic -             Evolutionary or
                         Evolutionary/           Data-Driven Models
                       Estimation Models
                                           Reduced-Order or
                                           Estimation Models
                           Physical
                            Model

                                   Physics of Failure-based
Application of Prognosis Approaches
PHM Priority Areas                  Approach & Algorithms
                                                 Component Degradation
1. System                                        •Physics-based
   Components                                    •Real-time life prediction
                      Performance
  •Gears/Bearings      Evaluation
  •Shafts/Couplings


2. Performance
  •Degradation
  •Wear

3. Drive Train                                   Signal Analysis
                                                 • Anomaly Detection
  •Vibration
                                                 • Sensor Fault Isolation
  •Oil Analysis
                                                 • Feature Prediction
                          Gray-Scale Trending
4. Actuation
                       Pressure 1
  •Hydraulic
  •Electrical
                         Oil 2
Prognosis Integration Across Vehicle


                                                                   PTO/AMAD
                                                                     Health




Valve Diagnostic Health    Turbine Degradation           APU Performance
• Anomaly Detection         •Physics-based                  Evaluation
• Sensor Fault Isolation    •Real-time life prediction
• Virtual Sensing
 Reliability and Usage Based
          Prognostics
                                    •Weibull Formulation
       Legacy-Based                  •Update Capability
      Maintenance Action              New Data
             PDF                      Legacy Data


              In-field Inspection
                 Results PDF




       In-Field MTBF
Legacy MTBF
     Evolutionary-based Prognostics
                                                              Degradation2
    Multidimensional                                            4% Fault
     Feature Space



                          Feature n
                                      Degradation 1
                                       2% Fault
   Diagnostics:
   Which fault is
   closest?                                                  Projected
                                                               Path      Prognostics:
    Fault Detection                                                      How long will it
      Threshold                                           Time2          take to get there?
                         Zone of
                        Uncertainty                                Feature 1
                                                  Time1

           F C
 
          f 2   s2
Degree Fi  2 (  )
Compressor Wash Prediction Using
    Evolutionary Technique

                   Uses PR, Fuel
                    Flow, and CDT to
                    predict efficiency
                     CIP assumed
                     CDT measurement
                      or pseudo-sensor
                      needed
                   Evolutionary
                    classifier
                    implemented for
                    prediction
Gas Turbine Implementation
     Data-Driven Prognostic Modeling

Creates an accurate data-driven model requiring no prior
      system knowledge, which can be used for fault                                                                 Prediction Accuracy
              detection and signal estimation
Raw data


       Estimate Sensor           Raw
      Values Using Model         data                     Model                                         Estimates



                                            3




                                           2.5




                                            2




      Compare to Actual                    1.5




                                            1




                                           0.5
                                                 0   10   20   30   40   50   60   70   80   90   100




                            No           Diagnose FAULT and
           Within limits?
                                        Predict model estimate
                   Yes
                                          Proceed as NORMAL
Neural Network/Data Driven Approaches
                              Weights


         Input features
                                        
                          •
                          •
                          •
Parameter Estimation Based Approach
                                           Measurement
  Input load/ Asset                           Noise                          comparison
    Environment                            x(k)                       z(k)
                            System                         Sensor                 +
  u(k)
                                                                                  _
                                                            Measurement
                        Model Error                         estimate    ˆ
                                                                        z (k )
                                       Current &
                                      future state
                                       estimates             Output                   Error e(k)
                        Model(s)                               H
                                      x (k ) | x (k  1)
                                      ˆ        ˆ

   Updated internal                          Feedback
 parameters or states                          Gain

              xk 1  xk  K k ( zk  H k xk ) | K k  Kalman Gain
              ˆ       ˆ                   ˆ
               xk 1  ( I  H k K k ) xk  K k zk
               ˆ                       ˆ
Kalman Filter Example Results




                                           Model Noise
                                           Toggle OFF




                               Kalman IC




              Measurement IC
   Model/Physics-Based Approaches

                     Physical Stochastic Model
  Diagnostic
   Results



 Expected
   Future
Conditions
 (based on
                                                 Current &
 historical
conditions)                                       Future
                                                  Failure
                                                 Prediction

Experience - Based
   Information
Model-Based Prognostic Example
Model-based Prognostic Results




                 Damage PDF at end
                    of Projection
             Integrated Model/Data Driven
                Prognostic Approaches
                     Raw Data
                                                                                Spall
                     Features                                                  Initiated



System Level                                                                                       Component Level
                 State Awareness
Model/Features                                                                                        Features



Initiation                Progression                                                                  Probabilistic   Current/Future
                                                                                                                         Capability
                                                                                                         Fusion          Prediction
                                                                                                         Process

                                                                                                                               Will it last the
Material/Physics Stochastic Models                                                                                             mission(s)?
                                                                                                         Reliability
                                                                                                                               What
                               Material                                                                 Information            Operational
              Expected         Resume                                                                                          changes
                                       Air Crew Diagnostics Demonstration Description
                                                Overall Project Objectives:
                            (1) Develop advanced machinery systems diagnostic processing




             Operating
                                  (2) Support the investigation of operator information
                                                      requirements




                                                                                                                               must take
                                          Approximate Duration of Demo: 5 min.
                               Equipment: 2 PCs with at least 19” color monitors + Internet
                                                     connections.
                                              Operating System / Software:




             Conditions      Executive Summary: The Air Crew Diagnostics demonstration
                                will highlight diagnostic algorithm capability over a variety of
                             helicopter mission and fault scenarios. The demonstration itself
                              will utilize available visual and aural commercial software tools
                              such as the Visual C++ developer’s studio. For each scenario
                               excerpt, the fault diagnosis is carried out using an automated
                                  intelligent agent / blackboard – based reasoning system.
                                Viewers of the demonstration observe the scenario re-plays
                                                                                                                               place?
                              from both ground-truth and pilot/air crew - centered viewpoints.
Adaptation Iteratively Improves
          Prognosis
     Current time   Current time       Current time
Damage




                                                    a3
                                                   Corrected
                                                    Models


                      a2
                                Measurement
         a1
                                  Update                 Time
          L1               L2                 L3
Integrated Model-Based Reasoning for
             Prognosis


                              Upstream
                             from F/FM
                             or Sym/Eff

                            Downstream
                            from F/FM
                            or Sym/Eff

 Illustrates causal dependencies and
   fault paths across the entire system
 Reasoning algorithms traverse the
  topology
                Prognostic Reasoning within a
                 Distributed PHM Architecture
                                                                                          A/D/P
                                                                                           Anomaly Detection /
    Logistics and Decision Support
                                                                                           Diagnostics / Prognostics
      Readiness & Availability,                        Mission Objectives                  Software
      Maintenance                                                                FPSI
                                                                                          FPSI
                                                                                            Failure Propagation &
                      Vehicle Level Reasoner                                               Subsystem Interactions
                                                                                          DSF
                             Publish - Subscribe                                           Domain Specific FPSI
                                                                                          LRU
Data Pipe                                                          A                        Line Replaceable Unit
                                               Subsystem
                                    DSF                             D
                                                Reasoner
                                                                    P
                                                                                          Readiness, Availability,
                                                                             Publish          Maintenance
                                    A
                    Subsystem                                                Only
    DSF              Reasoner
                                    D                                                                    Knowledge
                                    P
                                                   A       D                  Safety
                                                        LRU                                       Vehicle
                                                                              Critical
                A      D        P                                           Contingency
                       LRU                                                  Management         Subsystem
                                                            LRU              Functions
                                                                                                   LRU               Data
A         D     P                       PHM algorithms reside at the lowest                00010110     11010001
          LRU
                                         possible level (LRU or Subsystem)
     Prognosis Reasoning

          Reasoning Algorithm                      PosE

                                                    (1  FARj ) 
                                                                      NegE

                                                                        RFP   k
          • Evidence-Based     RankFM (i )  1 
                                                   j 1                k 1

          • Temporal-Based                                ( PosE)  ( NegE )



                             “Acts On”
   Integrated
     Model
   Graphical XML
     Database

• PHM Technologies
• Mission Requirements
• Interconnected Sub
Systems
• Symptoms/Effects
• Failure Propagation