Advancements in thermal runaway process monitoring: Exploring a novel residual dissimilarity-based Kernel independent component analysis method

Abstract

Thermal runaway faults typically develop gradually within complex systems at a relatively low rate, often remaining imperceptible during their initial phases. If not detected by adequate monitoring systems, these faults may go unnoticed until their consequences escalate to a critical level, potentially resulting in significant system degradation or failure. To address the limitations of traditional monitoring methods, this paper introduces a novel non-linear dynamic and non-Gaussian fault detection approach, termed residual dissimilarity-based kernel independent component analysis (RDKICA). RDKICA employs canonical variate dissimilarity analysis to construct both a state space and a residual space, effectively reducing dimensionality while preserving essential features for fault identification. In these spaces, state dissimilarity captures small drifts in linear components, whereas residual dissimilarity captures small drifts in nonlinear components. Kernel independent components are then extracted from the residual dissimilarity to effectively characterize small drifts in nonlinear components and account for non-Gaussian noise. The efficacy of the proposed algorithm is demonstrated through a comprehensive case study of a thermal runaway benchmark, complemented by an ablation study. The results showcase the superior detection performance of RDKICA in comparison to existing algorithms.