A branch selective kernel network based on slow features with interpretability for fault detection and diagnosis in chemical processes

Early fault detection and diagnosis (FDD) in chemical processes can significantly enhance operational reliability and reduce energy consumption. In recent years, data-driven methods such as multivariate statistical analysis (MSA) and deep learning (DL) have become the preferred approaches for FDD. However, the nonlinear and time-varying dynamic characteristics of modern chemical systems remain challenging to address using traditional methods. Additionally, the features learned by deep neural networks from process data are often difficult to interpret in the context of diagnostic tasks. To overcome these challenges, this paper proposes an interpretable Branch Selective Kernel Network (BSKNet) based on slow features. The network consists of kernel slow feature analysis (KSFA), branch fusion, and a Shapley Additive Explanation (SHAP) interpreter. Specifically, the kernel slow feature algorithm is used to reconstruct the features in the spatial domain, ensuring that the transformed slow features exhibit spatial correlation. Features from different regions with varying temporal dynamics are then extracted using branches with different convolution kernels, and an attention-weighted kernel selection mechanism is employed to fuse branch information. Finally, interpretability is provided both locally and globally by quantifying feature contributions through SHAP values, which help identify the root cause features with the greatest contributions. This method introduces a hybrid image-based chemical FDD framework (KSFA-BSKNet-SHAP) and is validated through experiments on the Tennessee Eastman Process benchmark dataset. Compared to popular methods, the proposed approach demonstrates superior fault detection and diagnosis performance.