Kernel entropy quality correlation analysis for nonlinear industrial process fault detection

Abstract

Quality-oriented fault detection plays a critical role in industrial processes, significantly boosting modern industrial efficiency since its inception. While kernel canonical correlation analysis is commonly used for nonlinear quality-oriented fault detection, it has certain limitations. To address these issues, this paper proposes a kernel entropy quality correlation analysis. The proposed approach initiates with nonlinear mapping to project the process variable space into a higher-dimensional space, effectively capturing nonlinear features within the data. By extracting the primary features contributing to the Renyi entropy of the dataset, a kernel entropy latent variable space is constructed, facilitating both nonlinear mapping and dimensionality reduction. Subsequently, canonical correlation analysis is employed to elucidate the relationship between the kernel entropy latent variable and the quality indicators. To rationally decompose the kernel entropy latent variable space according to the quality indicators, two decomposition strategies are proposed: the singular value decomposition-based method and the generalized singular value decomposition-based method. Moreover, this paper provides a theoretical justification for the validity of these two decomposition strategies. Finally, the effectiveness of the proposed method is validated through two numerical examples and two industrial case studies.