DICA enhanced SVM classification approach to fault diagnosis for chemical processes I. MONROYa, R. BENITEZb, G. ESCUDEROc, M. GRAELLSa. a Chemical Engineering Department, bAutomatic Control Department, cSoftware Department. EUETIB, Universitat Politècnica de Catalunya (UPC), Barcelona, SPAIN. Presentation Type Preference: Oral Presentation. Enter Theme/Sub Theme: On-Line Systems - Fault Diagnosis and Supervision Keywords: Dynamic independent component analysis (DICA), SVM, Fault diagnosis Type of contribution: Academy. Presenter professional category: PhD Student. As a result of the inherent dynamic and nonlinear characteristics of chemical process and their increasing complexity, on-line monitoring and fault diagnosis are gaining importance for plant safety, process economy, reliable maintenance and product quality. Standard multivariate statistical monitoring methods, such as principal component analysis (PCA) and partial least squares (PLS) do not explicitly take into account possible time correlations in the observations or deviations from Gaussianity of the latent variables[1]. However, in most cases state variables are driven by stochastic processes in the form of uncontrollable disturbances and random noise which may present both auto and cross-correlation. In order to address this problem, dynamic principal component analysis (DPCA) was proposed[2]. This approach constitutes a process monitoring method that uses an augmenting matrix with time-lagged variables and has been shown to be valid in different practical applications[3,4]. Recently, Independent Component Analysis (ICA) has been developed as a statistical technique that extracts statistically independent components from multivariate observed data. By using higher order statistical properties of the data, ICA provides a better identification of the underlying factors in the data than standard PCA techniques[5,6,7]. The extension of DPCA to ICA led to a new method called dynamic independent component analysis (DICA) which applies ICA to the augmenting matrix with time-lagged variables[3,8]. This method is able to extract the major dynamic features or source signals from the process and to find statistically independent components from auto- and cross-correlated inputs. This paper addresses fault diagnosis by combining DICA and Support Vector Machines (SVM). DICA is used for process monitoring and feature extraction in order to provide an enhanced representation of process information. On a second stage, SVM is used for fault detection and diagnosis as a classification algorithm. This hybrid DICA-SVM technique has been computationally implemented in MATLAB by using the free packages FastICA[9] and light-SVM[10]. The resulting fault diagnosis system (FDS) is then applied to the Tennessee Eastman[11] process as case study. Results are assessed by different performance indexes widely accepted in the machine-learning and process control literature. Preliminary results show that the proposed feature extension improves SVM classification performance, thus indicating that the presented hybrid approach constitutes a very promising alternative for fault diagnosis in chemical processes. References: 1. P Nomikos and J.F Macgregor. AIChE Journal, 40, 8 (1994). 2. W-F Ku, R-H Storer and C Georgakis. Chemometrics and intelligent laboratory systems. Vol 30(1995). 3. J Lee, C Yoo and I Lee. Chemical Engineering Science 59 (2004). 4. J Chen and K Liu. Chemical Engineering Science 57 (2002). 5. R.F Li and X.Z Wang. Computers and Chemical Engineering 26 (2002). 6. L Jiang and S Wang. Proceedings of the third Conference on Machine learning and cybernetics, Shanghai 26-29 (2004). 7. J Lee, S.J Qin and I Lee. AIChE Journal, 52, 10 (2006). 8. A Chen, Z Song and P Li. Lectures notes in computer science. Vol 3644 (2005). 9. Laboratory of Computer and Information Science. Helsinki University of Technology. 10.Thorsten Joachims. Department of Computer Science. Cornell University. 11.J.J Downs and E.F Vogel. Computers and Chemical Engineering, 17, 3 (1993). En este trabajo se presenta una nuevo planteamiento de monitorización estadística de procesos basado en la aplicación de Análisis de componentes independientes (ICA) y Análisis dinámico de componentes independientes (DICA) con las Máquinas de Vectores de Soporte (SVM) para monitorizar procesos con variables autocorrelacionadas y correlacionadas y mejorar el rendimiento de diagnosis de fallos. El objetivo de ICA es encontrar una representación lineal de datos no gaussianos con componentes estadísticamente independientes, mientras que DICA aplica ICA a una matriz aumentada con variables retrasadas en tiempo, lo cual podría extraer la mayor dinámica del proceso. Las matrices de componentes independientes son usadas como el conjunto de datos de entrada de las SVM para la diagnosis de fallos. La metodología es aplicada a diagnosticar los fallos del caso de estudio TE y el mejor rendimiento de diagnosis es obtenido usando sólo ICA+SVM con las condiciones que se usan en este artículo (ventana de tiempo de 2 muestras para DICA). Este trabajo será presentado en junio del presente año 2009.