A novel deep‐learning fault detection model of MSLR‐transformer for chemical process

Although Transformer is good at capturing long‐range dependencies in data, it has high computational complexity and has difficulty extracting multi‐scale features efficiently. Therefore, its fault detection performance is limited when dealing with chemical process data that is cross scale and with rich temporal feature. To address the above problems, a fault detection method based on multi‐scale sparse low‐rank‐Transformer (MSLR‐Transformer) is proposed in this paper. First, a parallel convolutional structure with multi‐scale receptive fields is constructed. The structure utilizes differentiated convolution kernels to simultaneously extract multilevel temporal features, which enhances the ability to perceive multiscale dynamic changes. Second, a sparse screening mechanism based on response strength is introduced for sifting out redundant information and highlighting highly responsive features, thus reducing localized noise interference in the process data. Then, the features are de‐ranked by a low‐rank linear attention based on kernel function mapping, which reduces the computational complexity. At the same time, the capture of global dependencies is achieved. Finally, the proposed method is applied to penicillin fermentation (PF) process and Tennessee Eastman (TE) process to verify the effectiveness. The experimental results show that the method is superior to the traditional model in fault detection.