A deep-learning model based on MFE-Transformer for chemical process fault detection

Chemical process data usually exhibits strongly nonlinear characteristics, local information loss, and sample imbalance problems. These problems make it difficult for traditional models to accurately capture the features and dependencies of chemical process data, thus affecting the fault detection effect. In order to solve the above problems, a fault detection method based on multiscale feature extraction-Transformer (MFE-Transformer) is proposed in this paper. First, the linear and nonlinear features in the data are captured respectively using a multiscale feature extractor. Second, local information enhancement is achieved by capturing the local information between the data by a multilayer convolutional neural network. After that, the high coupling in the data is captured using the multi-head self-attention mechanism in Transformer to identify the complex interactions between variables. Finally, the loss contribution of easy-to-classify samples is reduced by the focal loss function, while the sample imbalance problem is solved by enhancing the number of minority class samples using sliding window data enhancement. By experimental validation on penicillin fermentation (PF) process and Tennessee Eastman (TE) process datasets, the results show that the method outperforms traditional models in terms of fault detection performance.