Temporal-Aware Self-Attention Enhanced Bidirectional Gated Recurrent Unit for Dynamic Fault Detection in Industrial Processes

Timely and accurate fault detection is crucial for ensuring safety and reliability in industrial processes. However, complex temporal dynamics and nonlinear interactions make this task challenging. This paper proposes a novel deep learning architecture, termed Temporal-Aware Self-Attention Enhanced Bidirectional Gated Recurrent Unit (TASA-BiGRU), for dynamic fault detection. The model integrates a temporally-aware self-attention mechanism with sinusoidal positional encoding and a bidirectional GRU (BiGRU) network to jointly capture global dependencies and contextual sequential patterns. In the proposed framework, the input sequence is first enhanced by attention and then modeled by the Bi-GRU, followed by a decoder that reconstructs the original input. The reconstruction error is used as an indicator of abnormality, and the model is trained in an unsupervised manner using only normal operation data. Experiments on the Tennessee Eastman process demonstrate that TASA-BiGRU achieves superior detection performance compared with baseline models, confirming the effectiveness of combining temporally-aware attention and bidirectional recurrent learning for detecting dynamic industrial faults.