Data Mining Applied To Fault Detection Shinho Jeong

Data Mining Applied To Fault Detection Shinho Jeong Jaewon Shim Hyunsoo Lee Digital Media Lab, ICU

Abstract q Aims of work Ø Neural Network Implementation of the Non-linear PCA model using Principal Curve algorithm to increase both rapidity & accuracy of fault detection. q Data mining? Ø Extracting useful information from raw data using statistical methods and/or AI techniques. q Characteristics Ø Maximum use of data available. Ø Rigorous theoretical knowledge not required. Ø Efficient for a system with deviation between actual process and first principal based model. q Application Ø Process monitoring Ü Fault detection/diagnosis/isolation Ø Process estimation Ü Soft sensor 2 DML, ICU

Fault Detection(1) Fault introduction 3 DML, ICU

Fault Detection(2) q Major concerns Ø Rapidity Ü Ability to detect fault situation at an earlier stage of fault introduction. Ø Accuracy Ü Ability to distinguish fault situation from possible process variations. q Trade-off problem Ø Solve through Ü Frequent acquisition of process data. Ü Derivation of efficient process model through data analysis using Data mining methodologies. 4 DML, ICU

Fault Detection(3) q Inherent problems ① Multi-colinearity problem Ü Due to high correlation among variables. þ Likely to cause redundancy problem. þ Derivation of new uncorrelated feature variables required. ② Dimensionality problem Ü Due to more variables than observations. þ Likely to cause over-fitting problem in model-building phase. þ Dimensional reduction required. ③ Non-linearity problem Ü Due to non-linear relation among variables. þ Pre-determination of degree of non-linearity required. þ Application of non-linear model required. ④ Process dynamics problem Ü Due to change of operating conditions with time. þ Likely to cause change of correlation structure among variables. 5 DML, ICU

Fault Detection(4) q Statistical data analysis Ø Uni-variate SPC Ü Conventional Shewart, CUSUM, EWMA, etc. Ü Limitations þ Perform monitoring for each process variable. Ä Inefficient for multi-variate system. þ More concerned with how variables co-vary. Ä Need for multi-variate data analysis Ø Multi-variate SPC Ü PCA þ Most popular multi-variate data analysis method. þ Basis for regression modesl(PLS, PCR, etc). 6 DML, ICU

Linear PCA(1) q Features Ø Creation of… Ü Fewer => solve ‘Dimensionality problem‘ & Ü Orthogonal => solve ‘Multi-colinearity problem‘ new feature variables(Principal components) through linear combination of original variables. Ø Perform Noise reduction additionally. Ø Basis for PCR, PLS. q Limitation Ø Linear model => inefficient for nonlinear process. 7 DML, ICU

Linear PCA(2) q Theory Encoding mapping Decoding mapping 8 DML, ICU

Linear PCA(3) q ERM inductive principle q Limitation q Alternatives Ø Extension of linear functions to non-linear ones using… Ü Neural networks. Ü Statistical method. 9 DML, ICU

Kramer’s Non-linear PCA q Limitations Ø Difficult to train the networks with 3 hidden layers. Ø Difficult to determine the optimal # of hidden nodes. Ø Difficult to interpret the meaning of the bottle-neck layer. 10 DML, ICU

Non-linear PCA(1) q Principal curve(Hastie et al. 1989) Ø Statistical, Non-linear generalization of the first linear Principal component. q Self-consistency principle ① Projection step(Encoding) ② Conditional averaging(Decoding) 11 DML, ICU

Non-linear PCA(2) q Limitations Ø Finiteness of data. Ø Unknown density distribution. Ø No priori information about data. q Additional consideration ② Conditional averaging => Locally weighted regression, Kernel regression Ø Increasing flexibility(Span decreasing) Ü Span : fraction of data considered to be in the neighborhood. ~ smoothness of fit ~ generalization capacity 12 DML, ICU

Proposed Approach(1) q Creation of Non-linear principal scores 13 DML, ICU

Proposed Approach(2) q Implementation of Auto-associative N. N. 14 DML, ICU

Conclusion q Future works Ø Implementation of the proposed Neural Network structure. Ø Application for a case study 15 DML, ICU