A Novel BDPCA-SMLSTM Algorithm for Fault Diagnosis of Industrial Process

Fault diagnosis constitutes a crucial yet intricate endeavor within the industrial process field. Stemming from the robust feature representation capabilities inherent in deep learning models, the exploration of intelligent fault diagnosis methodologies grounded in deep learning has attracted significant attention and become a prevalent research focus. However, there are still many problems to be solved in fault diagnosis modeling, such as high-dimensional space, feature selection and overfitting. In this paper, considering industrial processes are nonlinear and dynamic, a slowness with Mogrifier-LSTM based on bidirectional dynamic inner principal component analysis (BDPCA-SMLSTM) is proposed to carry out the fault diagnosis task. First, after BDPCA extracts the most predictable dynamic latent variables (DLVs) from forward and reverse directions, the reconstructed DLVs are fed into SMLSTM to classify fault types. All the hyperparameters are determined by the particle swarm optimization (PSO) algorithm. Finally, the effectiveness of the proposed method is illustrated with the cases of the Tennessee Eastman process.