pattern recognition for system monitoring - an...
TRANSCRIPT
![Page 1: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/1.jpg)
Nalco Seminar, October 2, 2002 1
Pattern Recognition for System Monitoring - An Overview
Thierry DenoeuxUniversity of Compiègne
CNRS (National Center for Scientific Research)Compiègne, France
![Page 2: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/2.jpg)
Nalco Seminar, October 2, 2002 2
Outline! Pattern recognition for process control and
monitoring: two generic applications! software sensors! system diagnosis
! The development cycle of a PR system! analysis (choice of sensors, data collection, …)! design (training, model selection and performance
assessment)! implementation (robustness, refinement, adaptation, …)
! Case study: prediction of optimal coagulant dosage in water treatment plant.
! Conclusions
![Page 3: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/3.jpg)
Nalco Seminar, October 2, 2002 3
Software Sensor: Example
softwaresensor
Ca, Malk, Cl-SO4, HSP,
PSO, PO4, Temp, Zn, pH,
TCP Saturation, CaCO3 Sat
corrosion rate
software sensor: procedure for estimating a quantity of interest (output, response variable) from observable quantities (input variables, features)
![Page 4: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/4.jpg)
Nalco Seminar, October 2, 2002 4
Buiding a software sensor
The methodology for building a software sensor depends on the available information:! domain knowledge (equations, physical laws,
expert rules, …) → deterministic or conceptual modeling approach (domain-specific)
! statistical knowledge (past observations of the input and output variables) → supervised-learning, pattern recognition approach (generic)
![Page 5: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/5.jpg)
Nalco Seminar, October 2, 2002 5
System Diagnosis
System Σ, e.g. machine, process, river, etc.se
nsor
sRaw
measurements
ξ1(t)
ξi(t)
……
ξr(t)
x1(t)
xd(t)
xi(t)
……
prep
roce
ssin
g
Features
predictedstate of Σ(normal, fault 1,
fault 2, …)diag
nosi
s
![Page 6: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/6.jpg)
Nalco Seminar, October 2, 2002 6
Design of a diagnosis systemThe methodology depends on the nature of the
available knowledge:! A (logical or numerical) model for some or all of
the states → model-based approach (AI, control engineering)
! No model, but historical data of past measurements and observations of the system state → feature-based, pattern recognition-based diagnosis
![Page 7: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/7.jpg)
Nalco Seminar, October 2, 2002 7
Common framework: Supervised learning
! Two groups of variables:! inputs, features, attributes (x1,…,xd)=x
(measurements, or functions of measurements)! output y
! quantitative (regression), ! qualitative y ∈ G={1,…,K} (classification)
! Learning set: L={(xi,yi), i=1,…,n}
! Goal : predict y for a new case, based on observed input vector x.
![Page 8: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/8.jpg)
Nalco Seminar, October 2, 2002 8
The development cycle
Analysis• choosing sensors• defining the feature space
and the output space• collecting the data
Design• fitting learning model(s)• selecting or combining the
best models• assessing the performances
Implementation• robustness, novelty detection• missing data• model refinement and adaptation
need more information
(examples, features, sensors)
need to retrain model, or
build new modeldomain knowledge
statistical knowledge+ operational constraints
statisticalknowledge
![Page 9: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/9.jpg)
Nalco Seminar, October 2, 2002 9
Analysis
Design• fitting learning model(s)• selecting or combining the
best models• assessing the performances
Implementation• robustness, novelty detection• missing data• model refinement and adaptation
need more information
(examples, features, sensors)
need to retrain model, or
build new model
Analysis• choosing sensors• defining the feature space
and the output space• collecting the data
![Page 10: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/10.jpg)
Nalco Seminar, October 2, 2002 10
Analysis Phase: some guidelines
! Application specific, the choice of relevant sensor information can only be guided by domain knowledge.
! Typical questions:! How many features ?! How many data examples ?
![Page 11: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/11.jpg)
Nalco Seminar, October 2, 2002 11
How many features ?
Selected features should be limited to possibly relevantones: Too many features may be harmul !
number offeatures
error
lowest achievable
error
infinite sample
finite sample
best number of features
![Page 12: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/12.jpg)
Nalco Seminar, October 2, 2002 12
How many examples ?
number ofcases in
learning set
error
lowest achievable
error
highestacceptable
error
required number of data
![Page 13: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/13.jpg)
Nalco Seminar, October 2, 2002 13
Exploratory data analysis! A preliminary step to validate the data, and
help selecting relevant features, using! Elementary techniques: visualize one or two
variables at a time (histograms, boxplots, scatter plots, …)
! Multidimensional techniques: analyze the correlations between multiple features
! Examples of multidimensional techniques:! principal component analysis (PCA)! self-organizing feature maps
![Page 14: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/14.jpg)
Nalco Seminar, October 2, 2002 14
Example: classification of waste water for reuse
! Five classes of water quality, each class corresponding to a different possible usage
! 11 input features describing the chemical and bacteriological characteristics of water: suspended solids, TOC, conductivity, nitrate, etc.
! Problem: which features are relevant for classifying a water sample into one of the 5 categories ?
![Page 15: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/15.jpg)
Nalco Seminar, October 2, 2002 15
Analysis of a single input variable
![Page 16: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/16.jpg)
Nalco Seminar, October 2, 2002 16
Principal component analysis
• Objective: summarize multi-dimensional data by defining a small number of “informative” features
• Approach: Find the directions in input space that maximize the variance (scatter) of projected data
• Each direction = linear combinationof original features → new feature.
![Page 17: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/17.jpg)
Nalco Seminar, October 2, 2002 17
PCA: example (1)
1st axis: 76 % of the total variance (initial information)
2nd axis: 13 % of the total variance
![Page 18: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/18.jpg)
Nalco Seminar, October 2, 2002 18
PCA: example (2)
group of correlated variables → 1st axis
the 2nd axisis mostly explained
by this variable
![Page 19: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/19.jpg)
Nalco Seminar, October 2, 2002 19
Self-organizing feature maps
• A connectionist (artificial neural network) model.
• Goal : map high-dimensional data to a 2-D grid of neurons, in such a way that similar input vectors are mapped to neighboring nodes.
• This « topology preservation » property is obtained by a simple learning algorithm.
![Page 20: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/20.jpg)
Nalco Seminar, October 2, 2002 20
Learning algorithm
![Page 21: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/21.jpg)
Nalco Seminar, October 2, 2002 21
Correlation Analysis Using SOM’s
![Page 22: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/22.jpg)
Nalco Seminar, October 2, 2002 22
Design
Analysis• choosing sensors• defining the feature space
and the output space• collecting the data
Design• fitting learning model(s)• selecting or combining the
best models• assessing the performances
Implementation• robustness, novelty detection• missing data• model refinement and adaptation
need more information
(examples, features, sensors)
need to retrain model, or
build new model
![Page 23: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/23.jpg)
Nalco Seminar, October 2, 2002 23
Design - Statistical decision theory
! Let X be a random vector, Y real-valued random variable, joint distr. Pr(X,Y).
! We seek a function f(X) for predicting Y given X.! We need to quantify errors using a loss function
L(Y,f(X)).! Regression: L(Y,f(X))=(Y-f(X))2
! Classification: L(Y,f(X))=1 if f(X)≠ Y, 0 otherwise.
! The optimal f should maximize the expected prediction error:
EPE(f) = E(L(Y,f(X)))
![Page 24: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/24.jpg)
Nalco Seminar, October 2, 2002 24
Stat. decision theory (cont.)! The optimal solution:
! regression: the regression function f(x)=E(Y | X=x)
! classification: the Bayes rule
f(X)=class gk with highest posterior probability P(gk|x)(the Bayes rule has minimal EPE= error probability)
! Most learning methods aim at approximating the regression function or the Bayes rule.
![Page 25: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/25.jpg)
Nalco Seminar, October 2, 2002 25
Learning models! Hundreds of methods for classification and
regression. ! Some of the most popular models:
! linear methods (linear/logistic regression, LDA, …)! non parametric methods (k-NN, Kernel methods)! neural network techniques (multilayer
perceptrons, LVQ,…)! Support vector machines, ! decision trees, ! fuzzy systems, …
![Page 26: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/26.jpg)
Nalco Seminar, October 2, 2002 26
Linear discriminant analysis• Gaussian distribution (parametric method)• equal covariance matrices
linear decision boundary
![Page 27: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/27.jpg)
Nalco Seminar, October 2, 2002 27
Nearest neighbor method
?
• The k-NN method for classification: approximate theBayes classifier by classifying to the majority class among the k nearest neighbors of x.
• Similar method for regression
• Non-parametric method: works for any distribution (but high storage and computational requirements)
![Page 28: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/28.jpg)
Nalco Seminar, October 2, 2002 28
k-NN rule - Example
Bayes decisionboundary
7-NN decisionboundary
![Page 29: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/29.jpg)
Nalco Seminar, October 2, 2002 29
Multiplayer perceptrons
![Page 30: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/30.jpg)
Nalco Seminar, October 2, 2002 30
PMC (cont.)
![Page 31: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/31.jpg)
Nalco Seminar, October 2, 2002 31
Training of MLP’s! Principle: maximize a measure of fit
! R(w) is a non linear function of w → minimized using an iterative gradient-based non linear optimization algorithm.
desired output
weight vector input vector
network output regularization term
![Page 32: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/32.jpg)
Nalco Seminar, October 2, 2002 32
Example
![Page 33: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/33.jpg)
Nalco Seminar, October 2, 2002 33
Is there a "best" learning system ?No free lunch theorem:
No classifier is better than others for all problems
![Page 34: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/34.jpg)
Nalco Seminar, October 2, 2002 34
Pros and cons of learning algorithms
- restrictive assumptions- simple to implement- fast learning and operation
LDA
- slow learning- arbitrary decision boundariesMLP
- high storage and time requirements in operation
- no learning- arbitrary decision boundaries
k-NN rule
ConsProsClassifier
![Page 35: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/35.jpg)
Nalco Seminar, October 2, 2002 35
Tuning learning algorithms! Each classification/regression method has one or
more tuning parameters:
! Each tuning parameter controls the complexity of the model: greater complexity results in smaller bias, but greater variance→ bias/variance dilemma
number of hidden units nH, λMLPd, kk-NN rule
number of input features dLDA
![Page 36: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/36.jpg)
Nalco Seminar, October 2, 2002 36
The bias-variance dilemma
How to determine the optimal
model complexity ?
→ model selection
![Page 37: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/37.jpg)
Nalco Seminar, October 2, 2002 37
Model selection ! Model selection: given M models, find the one with
the smallest expected prediction error
! Problem: we know only the training error
which is a strongly biased (optimistic) estimate of EPE.
![Page 38: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/38.jpg)
Nalco Seminar, October 2, 2002 38
The hold-out estimation method
The simplest approach, when enough data is
available.
Train Validation Test
≈ 50 %fit the models
≈ 25 %model selection
≈ 25 %estimate
generalization errorof selected model
![Page 39: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/39.jpg)
Nalco Seminar, October 2, 2002 39
Cross-validation
! For k=1,…,K! fit the model using the data with part k removed! test the resulting model on part k
! Combine the K estimates of prediction error
Train TestTrain Train Train
1 2 3 4 5
Random partition of the data set (typically 5 · K · 10):
model fit without κ(i)tuning parameter
subset of example i
![Page 40: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/40.jpg)
Nalco Seminar, October 2, 2002 40
What to do is the estimated prediction error is too high ?
1) Add new features
validation
training number of features
error
α*
Additional features may or may not reduce the error (too many features may be harmful !)
![Page 41: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/41.jpg)
Nalco Seminar, October 2, 2002 41
What do do is the estimated prediction error is too high ? (cont.)
2. Add new examples: this can only reduce the error, but may be costly. How many ? → the number of examples allowing to achieve a given error can be predicted by extrapolating the learning curves.
![Page 42: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/42.jpg)
Nalco Seminar, October 2, 2002 42
3. Reject ambiguous patterns (classification)
What do do is the estimated prediction error is too high ? (cont.)
ambiguityregion
![Page 43: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/43.jpg)
Nalco Seminar, October 2, 2002 43
Rejection/error tradeoff
0 100%
rejection rate
error rate
perr
maximumadmissibleerror rate
correspondingrejection rate
![Page 44: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/44.jpg)
Nalco Seminar, October 2, 2002 44
Implementation
Analysis• choosing sensors• defining the feature space
and the output space• collecting the data
Design• fitting learning model(s)• selecting or combining the
best models• assessing the performances
Implementation• robustness, novelty detection• missing data• model refinement and adaptation
need more information
(examples, features, sensors)
need to retrain model, or
build new model
![Page 45: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/45.jpg)
Nalco Seminar, October 2, 2002 45
Outlier/novelty detection! Learning framework: (X,Y) have joint probability distribution
Pr(X,Y).! The training set composed of validated, quality-controlled data → realization of a random sample from Pr(X,Y).
! In operational conditions, the distribution of (X,Y) may changedue to:! different operating conditions! sensor faults! occurrence of new, previously unseen system states
! The output from the learning system may become unreliable, unless some outlier and novelty detection mechanism is implemented
![Page 46: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/46.jpg)
Nalco Seminar, October 2, 2002 46
Outlier: example
outlier:- sensor fault ?- new class ?
![Page 47: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/47.jpg)
Nalco Seminar, October 2, 2002 47
A robust pattern recognition system
feature vectorx ∈ Rd
predictionclassifier
outlierdetection
rejectedpatterns
![Page 48: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/48.jpg)
Nalco Seminar, October 2, 2002 48
Distance rejectionSimplest approach:• summarize the learning setusing prototypes• reject input patterns for whichthe distance to the closest prototype exceeds a given threshold.
• Determination of the prototypes = clustering problem• Algorithms: SOM, c-means
?
![Page 49: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/49.jpg)
Nalco Seminar, October 2, 2002 49
Example
ambiguityrejection
distancerejection
![Page 50: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/50.jpg)
Nalco Seminar, October 2, 2002 50
Missing data
x1
x3
x5
x2
x4
PR s
yste
mx1
?
?
x2
x4
PR s
yste
m
?sensorfailure
![Page 51: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/51.jpg)
Nalco Seminar, October 2, 2002 51
Missing data reconstruction! Let I = indices of missing features
J = indices of available features! Approach: estimate E(XI|XJ=xJ)
featuresmust be correlated !(redundancy helps)
![Page 52: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/52.jpg)
Nalco Seminar, October 2, 2002 52
Adaptation of learning systems
new systemstates
outliers
updateddecision boundary
![Page 53: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/53.jpg)
Nalco Seminar, October 2, 2002 53
Adaptation of learning systems (cont.)
! Continuous adaptation of a learning system requires:! outlier/novelty detection mechanisms! unsupervised algorithms for discovering new classes! a posteriori knowledge of class labels for predictive accuracy
improvement! Can be done on-line but difficult
! stability/plasticity dilemma! many tunable parameters
! More safely done off-line, using human supervision! some expertise in data analysis is necessary! increases maintenance costs
![Page 54: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/54.jpg)
Nalco Seminar, October 2, 2002 54
Case study: prediction of optimal coagulant dosage in WTP
SEINE
TankSandFilter
GACFilterOzonePulsator
DrinkingWater
C
UV
pH
T
DO
T°
T
TOC
pH
Z
Cl
T°Coagulantdosage
C : ConductivityCl : Residual ChlorineDO : Dissolved OxygenT° : TemperatureT : TurbidityTOC : Total Organic CarbonUV : Ultraviolet absorptionZ : Residual Ozone
GAC Filter :Granularactivated
carbon filterSludge
Viry-Chatillon treatment plant
Clarifier
![Page 55: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/55.jpg)
Nalco Seminar, October 2, 2002 55
Requirements! Determine optimum coagulant dosage using on-line
measurements of raw water quality parameters: turbidity, pH, conductivity, temperature,….
! Operation without human supervision:! robustness against erroneous data! estimation of missing input data! detection of and adaptation to changes in water
characteristics! Portability of the system to different sites:
! methodology for fitting the whole system automatically to new data
![Page 56: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/56.jpg)
Nalco Seminar, October 2, 2002 56
The system
predictedcoagulant
dosing rate
rawdata
Self-Organizing MapMultilayer Perceptron
validatedreconstructed
data
rejecteddata
training andmodel selection
algorithm
![Page 57: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/57.jpg)
Nalco Seminar, October 2, 2002 57
Learning algorithm
![Page 58: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/58.jpg)
Nalco Seminar, October 2, 2002 58
Results
0 20 40 60 80 100 120 140 160 180 2001
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
Sample
Coa
gula
nt D
ose
(ppm
)
![Page 59: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/59.jpg)
Nalco Seminar, October 2, 2002 59
Simulation of sensor fault
06/25/98 06/27/98 06/29/98 07/01/98 07/03/98 07/05/98 07/07/988
10
12
14
16
18
20
22
Time
Dis
solv
ed O
xyge
n (m
g/l)
Normal sensor
Faulty sensor
Reconstructed sensor
![Page 60: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/60.jpg)
Nalco Seminar, October 2, 2002 60
Results
06/26/98 06/28/98 06/30/98 07/02/98 07/04/98 07/06/98 07/08/982
2.5
3
3.5
4
4.5
5
Time
Coa
gula
nt D
ose
(ppm
)
Actual coagulant dose
Predicted coagulant dose
Confidence interval
06/26/98 06/28/98 06/30/98 07/02/98 07/04/98 07/06/98 07/08/982
2.5
3
3.5
4
4.5
5
Time
Coa
gula
nt D
ose
(ppm
)
Actual coagulant
Predicted coagulant dose
Without preprocessing With preprocessing
![Page 61: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/61.jpg)
Nalco Seminar, October 2, 2002 61
Conclusions! Pattern recognition (supervised learning) techniques
allow to build statistical models of the relationship between input and output variables, using observation data.
! Applications:! software sensor design! system diagnosis! data mining: text/image categorization, credit scoring,
financial decision making, …! The three phases in the development of a PR system:
! analysis (choice of sensors, definition of features, data)! design (model fitting and selection)! implementation (robustness, adaptation)
![Page 62: Pattern Recognition for System Monitoring - An Overviewtdenoeux/dokuwiki/_media/en/talks/nalco03.pdf · Nalco Seminar, October 2, 2002 2 Outline! Pattern recognition for process control](https://reader031.vdocument.in/reader031/viewer/2022031009/5b926fea09d3f23a718bc7ec/html5/thumbnails/62.jpg)
Nalco Seminar, October 2, 2002 62
Conclusions (continued)! The design of a pattern recognition system requires a
close cooperation between! domain experts (choice of input and output spaces, selection
of a representative learning set), and! statisticians (selection of learning techniques, interpretation
of results).! end-users (knowledge of operational constraints and
objectives)! Two pitfalls:
! expect too much from statistical techniques when too few data is available
! expect too much from huge data sets when domain knowledge is weak or the learning task has not been thoroughly specified