ICA-Multiclass SVM as a Monitoring System of Complex Processes

Abstract

Nowadays due to advancement of technology, the processes have become increasingly complex. So perform eectively monitor and correct diagnosis of these, has become a challenging task. this because the current processes have in common presence of combinations and correlation between analog and digital variables, presence of noise, etc. This issued is overcome by multivariate statistical methods which can eciently represent the dierent correlations between variables. One of the most popular methods for process monitoring is the Hotelling's T 2 distance but a large amount of data makes the process monitoring a dicult task. Principal Component Analysis (PCA) is used for the data reduction which means the extraction of few numbers representing the most variance of the analyzed data. However the use of these two methods is limited to normal multivariate data set and in industry rarely normal data set is presented, hence dierent methods are used for this purpose. Independent component analysis (ICA) has been used in no-normal mul- tivariate processes. In this paper a monitoring and fault diagnosis system for complex multivariate processes is presented. The new proposal is based on Multiclass Support Vector Machines (MCSVM) and the case study was addressed to the automotive industry. A comparison against a similar technique is carried out. Simulation of the new proposal shows promisory results.