关键词: 深度学习, 故障诊断, 多头自注意力机制, 双通道结构, 双阶段诊断

Abstract:

Deep learning models exhibit strong feature extraction capability for chemical-process fault diagnosis. However, under nonlinear and high-dimensional operating conditions, a single network often fails to capture both long- and short-term temporal dependencies simultaneously, leading to information loss and feature redundancy. To address this issue, this paper proposes a two-stage, dual-channel, multi-head self-attention fault classification model. In the first stage, a parallel GRU–LSTM dual-channel architecture is employed to learn complementary temporal representations: the LSTM branch captures long-term temporal dependencies, whereas the GRU branch focuses on short-term dynamics. Multi-head self-attention is then used to fuse and reweight the extracted features, emphasizing informative patterns while suppressing redundancy. Subsequently, based on fault confusion values, faults are divided into E-type and H-type categories. H-type faults are further processed in the second stage, where dynamic kernel PCA is applied to compute the Hotelling's (T²) and SPE statistics and construct an augmented feature matrix, thereby improving sensitivity to weak shifts, gradual degradation, and propagation-related anomalies. Experiments on the Tennessee Eastman process demonstrate that the proposed model outperforms representative baseline methods in detection and diagnosis accuracy, robustness, and generalization capability.

Key words: deep learning, fault diagnosis, multi-head self-attention mechanism, dual-channel structure, two-stage fault diagnosis