22 / 2025-03-24 10:29:58

Complex nonlinear industrial process fault detection and isolation using efficient dual-layer dynamic feature extraction

canonical variate dissimilarity analysis, random Fourier mapping, fault detection, nonlinear dynamic processes, fault isolation

Canonical Variate Dissimilarity Analysis (CVDA) has demonstrated promising potential in dynamic process fault diagnosis applications. However, its reliance on a linear single-layer approach for dynamic feature extraction often results in suboptimal fault detection performance, particularly in complex nonlinear industrial processes. To address this limitation, this study introduces an efficient Dual-Layer CVDA framework (DLCVDA) designed for monitoring complex dynamic processes. The contributions of this work are threefold. First, it establishes a dual-layer dynamic information extraction mechanism: the first layer captures linear dynamic features through CVDA processing, while the second layer extracts nonlinear dynamic features via a nonlinear CVDA applied in the residual space. Second, to enhance computational efficiency in nonlinear feature extraction, the framework employs Random Fourier Feature Mapping (RFF) instead of traditional kernel function mapping to facilitate nonlinear CVDA modeling. Third, the study incorporates input contribution analysis based on the Shapley value theory to enable precise fault variable isolation. Validation experiments conducted on a Continuous Stirred Tank Reactor (CSTR) process demonstrate that the proposed DLCVDA method outperforms both CVDA and KCVDA methods in fault detection performance while maintaining high computational efficiency. Additionally, the implemented fault isolation strategy effectively identifies the correct fault variables