Research On Fault Diagnosis Method Of Multimodal Process Based On Deep Learning

With the increasing complexity of modern chemical process,there are more and more unstable factors of operating conditions and operating environment.If abnormalities occur in the chemical process,they may turn into serious accidents and cause significant safety,environmental and economic impacts if not corrected in time.Fault detection and diagnosis is the key technology to ensure the safe production of chemical processes.How to detect and diagnose faults accurately and timely has become one of the research focuses in the chemical industry.Deep learning has recently received extensive attention due to its powerful feature extraction ability.As a multi-layer stack of simple nonlinear functions,the deep learning model can fully extract the characteristics of chemical data layer by layer.Therefore,the research on the fault detection and diagnosis method of chemical process based on deep learning has attracted extensive attention of scholars.Aiming at the nonlinear,time-dependent and multi-modal detection problems of chemical process,the thesis designs a fault detection method of variational self-encoder model based on gated cyclic unit.In the model training stage,GRU-VAE model is trained by using the data in normal mode,and the abnormal score is calculated according to the training set to get the monitoring threshold,which is set as the upper control limit.In the anomaly detection phase,the model detects the real-time data set,and considers the reconstruction loss beyond the range as fault data.In order to verify the effectiveness of GRU-VAE detection network,the thesis uses Tennessee Eastman(TE)process data to verify on the basis of consulting relevant literature.The experimental results show that the detection accuracy of this method can reach 96.725%,which has good detection effect and good application prospects,and provides a new idea for fault detection.Industrial processes may produce multiple modes in the production process,to solve the problem of lack of fault data or insufficient label data for some modes and too many model parameters.For this reason,the thesis proposes an ensemble-based multimodal process fault diagnosis method based on the combination of attention mechanism and deep subdomain adversarial adaptive networks(AM-DSAAN).By building sub-domain adaptive adversarial migration learning network,introducing attention mechanism,using Bi GRU and GRU in parallel to improve the model’s ability to extract fault features,realizing nonlinear transformation,aligning the relevant sub-domains of source and target domain distribution to minimize the difference of sub-domain distribution,and using the gradient inversion layer and domain classification layer of domain adversarial training to reach an equilibrium state to get the best migration features,this method can transfer knowledge from the source mode to the target mode of fault diagnosis.To verify the diagnostic performance of the proposed method for unmarked fault data,the proposed AM-DSAAN method was also verified under four modes of the Tennessee Eastman process,and the experimental results showed a good performance with a diagnostic accuracy of 95.17%.