A temporal-aware dual-attention network for fault diagnosis in industrial processes

In recent years, deep learning approaches based on attention mechanisms have been extensively applied in the fault detection and diagnosis of complex industrial processes. However, existing methods typically extract features from process data at the time-point level, which makes it challenging to capture the long-term temporal features of series-level fault information. Therefore, this paper proposes a novel industrial fault diagnosis method based on the temporal-aware dual-attention network (TDANet). The temporal semantic features are extracted from inter-series and intra-series variations through the temporal second-order attention mechanism. The multi-series feature fusion achieves coarse-to-fine feature fusion, enhancing the temporal representation of fault information. Experimental results on the benchmark and real-world datasets demonstrate that TDANet outperforms comparative models in terms of accuracy and F1 score, thereby validating the effectiveness of the proposed method.