Incipient fault detection in industrial processes remains challenging, particularly for notorious faults 3, 9, and 15 in a chemical process benchmark, namely the Tennessee Eastman process (TEP). This paper proposes a novel unsupervised framework, namely self-modulated KAN-enhanced direct cross-attention (SMK-DCA). It constructs heterogeneous features by integrating raw data, process-aware features, and sliding-window singular values. Kolmogorov–Arnold networks (KAN) enhance nonlinear expressiveness before a cyclic DCA mechanism enables comprehensive interactions among heterogeneous features. A feature-wise linear modulation (FiLM) adaptively calibrates representations, while a sparse autoencoder with multi-target reconstruction amplifies subtle fault signatures. By leveraging KAN’s superior approximation capability and cyclic multi-view fusion, the proposed method effectively captures incipient fault-induced variations often overlooked by conventional approaches. Extensive experiments on TEP demonstrate that SMK-DCA effectively detects incipient faults 3, 9, and 15, while obtaining the best average detection rate across all faults among the compared MSPM and deep learning methods. Furthermore, validation on real-world data from an IGBT power system confirms the generalization capability of the proposed method across different industrial domains.
Yu et al. (Thu,) studied this question.