A Multi-Scale Cross Transformer Network-Based Fault Diagnosis Method for Industrial Process

The process of industrial production is becoming increasingly complex in modern times, posing higher demands for safety. It's worth noting that data from modern industrial processes often exhibit multivariate, temporal and multiscale characteristics. To address these problems, this paper presents a novel fault diagnosis method based on the Multi-Scale Cross Transformer model. We integrate encoder, convolution and patching mechanisms, making it capable of extracting temporal features from different time scales, effectively capturing short-term dynamics and long-term trends. Additionally, a two-stage attention mechanism is introduced to automatically learn correlations across both time and space dimensions. Experiments conducted on the Tennessee-Eastman process demonstrate the superiority of our proposed method. It shows over 10% accuracy enhancement in diagnostic performance compared to state-of-the-art methods. Furthermore, the method can gradually map data to a discriminative feature space with better intra-class compactness and inter-class separability, reflecting its reliability and validity.