Improved Fault Diagnosis Using Dynamic Kernel Scatter-Difference-Based Discriminant Analysis

Abstract

Performing fault diagnosis of nonlinear processes involving data with serial correlations, nonlinearities, and overlapping signatures is a challenging task. This article proposes a dynamic kernel scatter-difference-based discriminant analysis (DKSDA) for resolving such complex data so as to improve fault resolution for efficient diagnosis. The DKSDA considers a suitably time lagged extension of the original data and allows its transformation into high-dimensional feature space via nonlinear kernel functions, and then solves the scatter difference form of the Fisher criterion. This fault diagnosis method successfully addresses the problem of dynamic correlations that are typically associated with chemical process measurements and efficiently captures the nonlinearities in data. A systematic procedure is proposed to configure the interactive parameters, namely, the number of lags, the scatter difference, and the kernel width, which govern the performance of DKSDA. The procedure involves a two-dimensional grid search at two levels to minimize a performance criterion defined in terms of the misclassification of DKSDA scores evaluated by cross validation using the nearest mean classifiers. The performance of the proposed method is evaluated by applying it for the diagnosis of overlapping and nonoverlapping faults of Tennessee Eastman challenge process and the overlapping faults of a general multivariable nonlinear dynamic process. The comparison of results with the recently reported methods demonstrates the superior performance of DKSDA for nonlinear process fault diagnosis involving complex overlapping data.