FAULT DIAGNOSIS BASED ON MULTI-SCALE CLASSIFICATION USING KERNEL FISHER DISCRIMINANT ANALYSIS AND GAUSSIAN MIXTURE MODEL AND K-NEAREST NEIGHBOR METHOD

Abstract

Effective fault monitoring, detection and diagnosis of chemical processes is important to ensure the consistency and high product quality, as well as the safety of the processes. Fault diagnosis problems can be considered as classification problems as these techniques have been proposed and greatly improved over the past few years. However, a chemical process is often characterized by large scale and non-linear behavior. When linear discriminant analysis is used for fault diagnosis in the system, a lot of incorrect diagnosis will occur. As solution, this paper presents a novel approach for feature extraction and classification framework in chemical process systems based on wavelet transformation and discriminant analysis. The proposed multi-scale kernel Fisher discriminant analysis (MSKFDA) method used the combination of kernel Fisher discriminant analysis (KFDA) and discrete wavelet transform (DWT) to improve the classification performance as compared to conventional approaches. A DWT is applied to extract the process dynamics at different scales by decomposed the data into multiple scales, analyzed by the KFDA and only dynamical characteristics with important information was reconstructed by inverse discrete wavelet transform (IDWT). Then, Gaussian mixture model (GMM) and K-nearest neighbor (KNN) method were individually applied for the fault classification using the output from the MSKFDA approach. These two classifiers are evaluated and compared based on their performance on the Tennessee Eastman process database. The proposed framework for GMM and KNN classifiers had achieved average classification accuracies of 84.72% and 82.00%, respectively, with the results show significant improvement over existing methods in fault detection and classification.