关键词: 过程系统, 安全, 神经网络, 故障诊断, SlowFast网络, Transformer模型, 风险预警

Abstract:

Considering the high dimensionality, nonlinearity, and strong temporal correlations of chemical process data, this paper proposes a multi-scale spatiotemporal fusion fault diagnosis method based on the SlowFast network and Transformer. The method is designed with a dual-channel feature extraction architecture: the Slow pathway employs temporally down-sampled convolutions to capture macro-level process features, while the Fast pathway retains high temporal resolution convolutions to extract local dynamic characteristics. By integrating the self-attention mechanism of Transformer, long-range dependencies are effectively modeled. A sliding-window strategy is adopted to construct spatiotemporal samples, enabling accurate fault classification and real-time monitoring. Experimental results on the TE process dataset demonstrate that, with a four-layer Transformer block stack, the proposed model achieves outstanding diagnostic performance. Compared with existing models such as DCNN (88.20%), LSTM (95.37%), CNN-LSTM (96.64%), and DCRNN (91.70%), the proposed approach significantly improves fault diagnosis accuracy (98.44%). By introducing the SlowFast architecture from the video analysis domain into chemical process time-series modeling and combining it with the global perception capability of Transformer, the classification effectiveness is substantially enhanced. Furthermore, a novel early-warning evaluation metric is proposed, and the relationship between the diagnostic performance and interpretability of SlowFast is discussed based on experimental results, providing a new technical approach and theoretical reference for fault diagnosis in complex industrial processes.

Key words: process systems, safety, neural network, fault diagnosis, slowfast network, transformer model, risk early warning