Dimensionality Reduction Nonlinear Partial Least Squares Method for Quality-Oriented Fault Detection

Abstract

Unlike traditional fault detection methods, quality-oriented fault detection further classifies the types of faults into quality-related and non-quality-related faults. Therefore, quality-oriented fault detection has attracted significant attention in industrial applications due to its ability to provide more comprehensive fault information. Various approaches have been presented to cope with this challenge. Nevertheless, these approaches often struggle to effectively dissect the process variable space, leading to limitations in quality-oriented fault detection. Motivated by this background, this study introduces a kernel principal component analysis (KPCA)-based quality-oriented partial least squares method, offering a more suitable decomposition and a more straightforward monitoring logic. In contrast to the standard kernel partial least squares method, the proposed method employs the KPCA method to capture the nonlinear characteristics inherent in the original process variable space, subsequently reducing its dimensionality to yield the kernel principal component space, which not only encapsulates the nonlinear characteristics of the original process variable space but also mitigates noise and unknown interferences in the data, thereby achieving dimensionality reduction. On this basis, an orthogonal decomposition of the kernel principal component space is achieved using generalized singular value decomposition technology of the load matrices of the kernel principal component and quality variable spaces, significantly enhancing monitoring performance. Finally, the validity and superiority of the proposed method are demonstrated through two case studies.