detecting, diagnosing and controlling process variations

Detecting, Diagnosing and Controlling Process Variations: A Review of Modeling Options

Venkat Venkatasubramanian

Laboratory for Intelligent Process SystemsSchool of Chemical Engineering

Purdue UniversityW.Lafayette, IN47907

© V.Venkatasubramanian

Outline

IntroductionExceptional Events (Abnormal Events) Management (AEM/EEM) Intelligent ControlChallenges and Issues

Model based and Process History-based methodsEmerging TrendsFuture Directions


Lecture Philosophy

Broad overviewNot a detailed, in-depth reviewIdentify key concepts, issues, challengesCompare and contrast different approaches


IntroductionIntelligent Control deals with Exceptional Events in Process PlantsExceptional Events are deviations in process behavior from normal operating regime

Safety problemsEnvironmental concernsQuality problems and Economic losses

Why do exceptional events occur?Human errors Equipment degradation and failures

This is really a Process Control Problem


Process Fault Detection and DiagnosisFirst Step in Exceptional Events Management

Find out what is wrong

Interpreting current status of process Utilizing sensor data & process knowledge, isolate abnormal situations

DiagnosticSystem

Sensor Readings

Process KnowledgeProcess

Infer Process State - Detection: Normal/Abnormal& Isolation of abnormality

Problem Size and ComplexityAs many as 1500 variables monitored every few secondsIntegrated Process Operations

Broad scope of the task: Variety of Process Failures Parametric Faults, Structural faults, Malfunctioning Sensors/Actuators

Process Measurements may beInsufficient, Incomplete, Unreliable

Diagnostic Approaches – Brief Review

Process Fault Diagnosis: First Step in Intelligent ControlDiagnostic Philosophies

Source of Process knowledge- Process Model - Process History

Form of Process knowledge- Qualitative - Quantitative

Process Model : Deep, Causal or Model-Based knowledgeProcess History : Shallow, Compiled, Evidential knowledge

V. Venkatasubramanian, R. Rengaswamy, K. Yin and S. N. Kavuri, “Review of Process Fault Diagnosis - Part I, II and III”, Computers and Chemical Engineering, 27(3), 293-346, 2003.


TransformationsDiagnostic Decision Process

View as a series of transformations or mappings on process measurements

Important Components in the process:a priori process knowledge and search technique

Measurement Space

Feature Space

Decision Space

Class Space

Sets of measurements input to the diagnostic system

Useful features extracted as a function of measurementsusing a priori problem knowledge

Map feature space to decision space to meet some objective - minimizing misclassification e.g. using discriminant function/threshold function

Class space: fault classes, final interpretation of the system deliveredto the user. By threshold functions, template matching, symbolic reasoning


Example

x1

xn

ynn

f1

f1

f2

f2

f3

Hidden layer Output layerInput layer

wi

... ... ... ...

Measurement Space Feature Space Decision Space

Class Space

Neural Network based classifierInput Nodes: Measurement SpaceHidden Nodes: Feature SpaceOutput Nodes: Decision SpaceInterpretation of Outputs: Class Space


Classification of Diagnostic Methods: Options

DiagnosticMethods

Structural

Functional

Causal Models

AbstractionHierarchy

Fault Trees

Digraphs

Quantitative Qualitative

EKF

ParitySpace

Observers

QualitativePhysics

Process History Based

Qualitative Quantitative

Expert systems

QTA Statistical

StatisticalClassifiers

PCA/PLS

Neural Networks

Model-Based

QUANTITATIVE MODEL-BASED


Model-based Detection and Diagnosis

Consistency Checking (Analytical Redundancy): Compare actual behavior with a nominal fault-free model driven by same inputs, using residuals.Residuals: Functions accentuated by faults representing this inconsistency

ProcessActual Operation

Model ofNormal Operation

MeasuredSituation

CalculatedSituation

Knowninputs (u)

Unknowninputs (d) Faults (f) Known

inputs (u)

Comparison

ResidualGeneration

DecisionMaking

Statistical testing of residuals to arrive at a diagnostic conclusion

Unknown fault modes, uncertain nominal model,system/measurement noise

ResidualAnalysis

FaultsResiduals


Quantitative ModelsKey idea is to check the parity (consistency) of plant models by using actual measurements - sensor outputs (measurements) and process inputs

Observer-based MethodsDevelop a set of observers each sensitive to set of faults while being insensitive to the other faults/unknown inputs.Observers for nonlinear systems: Elegant approach for fault-affine (Frank,1990)Bilinear nonlinearity (Ding et. al. 1995); Differential geometric methods (Yang and Saif, 1995)Good Reviews by Frank (1990,1994); Patton & Chen (1997)

Parity-Space ApproachesRearranged/transformed variants of the input-output/state-space modelsChow and Willsky (1984) provided a general scheme for both direct and temporal redundancyOnce residual properties have been selected all parity and observer based designs are fundamentally equivalent (Gertler, 1991) Reviews of parity space by Patton and Chen (1991); Gertler (1997)


Quantitative ModelsKalman Filters

Stochastic version of observers: Estimating x based on u and y; with gaussian white noisesKalman filter in state space model is equivalent to an optimal predictor for a linear stochastic system in the input-output model.Single Filter (Mehra and Peshon, 1971, Clark 1978)Bank of filters (Wilbers and Speyer, 1989)Extended Kalman Filter (Chang and Huang, 1998; Huang et. al., 2000)

Parameter Estimation TechniquesAlternative Approach: Not based on state estimationFaults of a dynamic system are reflected in the physical parametersDetect faults via estimation of the parameters (Isermann, 1984)Different estimation techniques: least squares, nonlinear optimizationGood Reviews: Young (1981), Isermann (1984), Patton et al., (1989), Frank (1990)


Issues in Model-Based Diagnosis

Modeling uncertainty: Availability of a ‘decent’ modelPerfectly accurate/complete model of a physical system is rarelyavailableUncertainty: Mismatch between actual process and its math modelEffect of modeling uncertainties is very crucial

Sensitizing/Desensitizing & RobustnessDecoupling fault effects, system/measurement noiseInsensitive to modeling uncertainty, noise, disturbances with increased sensitivity to faults

Online speed: Computations Vs PerformanceTheory

Well developed for linear systems, nonlinear systems’ still not mature


Quantitative Model-Based

MeritsMathematical approach based on first principlesEstimation of fault magnitude, multiple fault identifiabilityHandling novel situations

DemeritsRequirement of a good modelEffect of modeling uncertainties is very crucialNeed robustness to shifts in normal regimes, noiseLinear Vs Nonlinear systemsProcedural nature: No explanation facility

Online speed: Computations Vs Performance


Classification of Diagnostic Methods

DiagnosticMethods

Structural

Functional

Causal Models

AbstractionHierarchy

Fault Trees

Digraphs

Quantitative Qualitative

EKF

ParitySpace

Observers

QualitativePhysics

Process History Based

Qualitative Quantitative

Expert systems

QTA Statistical

StatisticalClassifiers

PCA/PLS

Neural Networks

Model-Based


QUALITATIVE MODEL-BASED


Qualitative Models

Fault Trees Qualitative cause-effect relationships represented as a treeFault Trees (Lapp & Powers, 1977)

Top-down, symptom-driven approachQualitative ambiguities

Signed Digraphs and Causal ModelsQualitative cause-effect relationships as a directed graph with nodes and arcsBottom-up, cause-driven approachQualitative ambiguitiesSingle fault diagnosis (Iri et.al., 1979; Kramer and Palowitch, 1987;Wilcox and Himmelblau, 1994)

Multiple fault diagnosis (Morales and Garcia, 1990; Vedam and Venkatasubramanian, 1999)

Abstraction Hierarchy


Qualitative Model Based

MeritsQualitative nature facilitates their development without exact model equationsCompleteness: Enumerates all possible root causes

DemeritsNeed to develop and maintain SDGQualitative nature results in poor resolutionNumber of spurious solutions


Model based Framework: Challenges

Appropriate level of modelingToo Coarse: Not very useful resultsToo Fine: Too complex and buried in details Quantitative vs Qualitative

ALL MODELS ARE WRONG, SOME ARE USEFUL

-- George Box

PROCESS HISTORY-BASED


Process History Based Techniques

Process-History Data from normal/abnormal operationQuantitative Methods

Neural Networks : Easy learning & interpolation capabilities leading to numerous diagnostic applications (Venkatasubramanian et. al. 1989, 93, 94)

Statistical Techniques :Univariate SPC based on limit-checkingMultivariate PCA/PLS based monitoring (MacGregor et. al., 1994)

Qualitative MethodsQualitative Trend Analysis : Sensor Trend Information

Trend Language, Trend Identification and Mapping to faults (Cheung & Stephanopoulos, 1990; Janusz & Venkatasubramanian, 1991; Rengaswamy & Venkatasubramanian, 1995)

Rule-Based Expert-Systems : Mapping of known symptoms to root causes (Kramer, 1987; Rich & Venkatasubramanian, 1987)

Process History enters the antecedent and consequent of rulesLack of generality; Poor handling of novel situations

Qualitative Trend Analysis: Representation

Primitives RepresentationEach segment is represented by a quadratic polynomial equation and is characterized by its first and second derivatives

23Janusz, M. et al. (1991). Automatic generation of qualitative description of process trends for fault detection and diagnosis. Engineering Applications of

Artificial Intelligence 4(5), 329‐339

D B C

G E F

A

(+,-) (+,+)(+,0)

(-,-) (-,+) (-,0)

(0, )

Fundamental Language: Primitives

D G

E B

D G

All trends can be represented by these seven basic shapes (primitives)

This sequence of basic shapes (primitives) represents the trend


Desirable Features of an Intelligent Control System

Early detection & diagnosis

Isolability : discriminate between failures

Robustness : to noise & uncertainities

Novelty Identifiability : novel malfunction

Explanation facility : Fault propagation

Adaptability : Processes change & evolve

Reasonable storage & computational requirement

Multiple Fault Identifiability : Difficult requirement


Comparison of Different Diagnostic Methods

No single method achieves all !!!


Hybrid FrameworkNo single method meets all the criteria of

a ‘good’ diagnostic methodA Hybrid Framework

Involving different methodologiesBased on a collective and synergistic approach to problem solving seems most promising (Mylaraswamy & Venkatasubramanian, 1997)Compensate one method’s weakness with the strengths of another’s

DKit implemented in G2Effectiveness demonstrated on Model IV FCCU by successfully diagnosing wide varieties of faultsCombined causal model-based diagnosis with statistical classifiersBasis for the prototype of the ASM ConsortiumLicensed to Honeywell by Purdue University

Hybrid Methods

27Maurya, M.R. et al (2007). A signed directed graph and qualitative trend analysis‐based framework for incipient fault diagnosis. Chemical Engineering

Research and Design 85(A10), 1407‐1422

Obtain SDG based on equations governing the processObtain Measured IRT

Obtain Set of Possible Candidate Faults

Compare IRT

Simple UnimodalsRepresent faults as sequences of basic shapes

Compare sequences

Knowledge Base of Fault

Patterns

Knowledge Base of Fault

Patterns

Most Probable Fault Determined

Qualitative Trend Analysis (QTA)

Fault Detection

& Diagnosis

(Level 1)

Fault Detection

& Diagnosis

(Level 1)

Fault Detection

& Diagnosis

(Level 2)

Fault Detection

& Diagnosis

(Level 2)

Data from ProcessSigned Directed Graphs (SDG)

Hybrid Methods: All methods have advantages and disadvantages. Combining methods allows us to capitalize on the advantages and make up for the disadvantages

28

EEM: Alexanderwerk Roller Compactor

Roll GapFeed Screw Speed

Hydraulic Pressure

Feedback Control

Roll Speed

Event: No powder entering roll regionCauses: No powder in hopper

Blockage in hopperJam in nip region

Event: No powder entering roll regionCauses: No powder in hopper

Blockage in hopperJam in nip region

Roller Compactor

29

Feedback Controller

Governing Equations of Roller Compaction Process

Qualitative Simplification of Mathematical Model

Signed Directed Graph and Initial Response Table

In the absence of models that capture abnormal behavior, expert and/or operator knowledge of exceptional events can be incorporated into initial response table

Qualitative Trend Analysis: Hybrid Methods

30

Fault Detection & Diagnosis: SDGFirst Level of Diagnosis

Fault Detection & Diagnosis: QTASecond Level of Diagnosis

Roll Gap Feed Speed Hydraulic Pressure

Roll Speed

Confidence Index

0.592 0.503 0.653 1.000 0.503

Step 1 Calculate Similarity Match Step 2 Calculate Similarity Measure

B curve under AreaA curve under Area

Based-Shaped =

( )⎥⎥⎦

⎤

⎢⎢⎣

⎡ −−= 2

2

21

expmr

ab

σBased-Size

∑Δ∑ Δ

=i

iittS

MeasureSimilarity

Maurya, M.R. et al (2007). A signed directed graph and qualitative trend analysis‐based framework for incipient fault diagnosis. Chemical Engineering

Research and Design 85(A10), 1407‐1422

31

Feedback Controller

32

EEM detected and diagnosed exceptional within 10 seconds of its inception

EEM detected and diagnosed before feedback controller reacted to decrease in roll gap by increasing feed screw speed

Incipient detection & diagnosis

Currently, exceptional event and mitigation strategy appear as warning

0 5 10 15 202200

2400

2600

2800Roll Gap

0 5 10 15 2029

30

31

32Hydraulic Pressure

0 5 10 15 2018

18.5

19

19.5

20Feed Screw Speed

0 5 10 15 203

3.5

4

4.5

5Roll Speed

Continuous Manufacturing Line

33

Feeder A

Feeder B

Blender Roller Compactor

API

Blend

Excipients

Ribbons

Partially Continuous Dry Granulation Line

Powder Weight (kg)

Powder Weight (kg)

Feed Rate (kg/hr)

Feed Rate (kg/hr)

Roll Gap (microns)

Feed Screw Speed (rpm)

Hydraulic Pressure (bar)

Roll Speed (rpm)

Blender Mix Speed (rpm)


Integrated Intelligent Control Systems

Intelligent MonitoringIntelligent Monitoring

Fault DiagnosisFault Diagnosis

Regulatory Control(MPC)

Regulatory Control(MPC)

ProcessProcess

Data AcquisitionData Acquisition

OperatorOperator

Data Reconciliation Parameter EstimationData Reconciliation Parameter Estimation

Supervisory ControlFault Administration

and RTOSupervisory ControlFault Administration

and RTO

Data

Data

Data Data

Manipulated Variables

Reconciled Data

Sens

or

Tren

ds

Faults

Faults FDDDesign

Setpoints

Cont

rolle

r Se

ttin

gs

Information Specifications

Parameter Estimates


Summary

Reviewed the approaches, challenges, and emerging trends in Intelligent Control Systems

Model based and Process history based approachesHybrid Systems

Intelligent control systems can make a substantial improvement to current practices

Considerable challenges remain but good progress has been made in recent years

Future potential is enormous and exciting


Acknolwedgements

CollaboratorsNSF ERC C-SOPS Team

Funding AgenciesHoneywell ASM ConsortiumExxonMobil, ICI, Air Products, Mitsubishi, A D Little, Searle, Eli Lilly, Nova ChemicalsNSF, NIOSH, DOE, INLNSF ERC C-SOPSPurdue Trask Fund

detecting, diagnosing and controlling process variations

Documents