Research On Fault Diagnosis Method Of TE Process Based On Attention Mechanism

In recent years,with the rapid development of science and technology and the increasing complexity of chemical processes,automation has become higher and higher.However,any minor fault can cause serious consequences,so timely and accurate fault diagnosis of chemical processes to ensure the regular operation of equipment is a hot issue for research in various engineering disciplines.In this thesis,we take Tennessee Eastman(TE)process as the research object and realize fault diagnosis by combining deep neural network and attention mechanism modeling based on TE process data.Due to the complex characteristics of chemical process data,such as nonlinear,non-Gaussian,high dimensional,time-varying,and multimodal,the following problems exist in the research of TE process fault diagnosis methods.(1)In the TE process,many data variables and time series lead to a low accuracy rate of conducting fault diagnosis.In the traditional fault diagnosis method,the adjacent samples are assumed not to interfere,and the time series relationship between data is not considered,resulting in the limitation of fault diagnosis.(2)It is difficult to diagnose the problem because the characteristics of minor faults are not prominent.In the TE process,some faults are not obvious,and although they do not have a severe impact,they may cause other accidents if they are not found and taken in time.Therefore,these minor faults should not be ignored.(3)TE process multimodal fault is difficult to diagnose.TE process usually operates in multiple operating modes.Because of the differences in data distribution in different modes,the neural network fault diagnosis model trained with single-mode data cannot be effectively applied to other modes.This thesis designs fault diagnosis algorithms for unimodal and multimodal TE processes to address the above problems.The main innovations and researches of the thesis are as follows.(1)For the problem of many TE process data variables with time series and difficult to extract fault features,this thesis proposes a parallel TCN module and a one-dimensional convolutional neural network(1D-CNN)in the TCN module slides with time step in each sub-block and considers adjacent samples and closely related variables to extract local time-related features,and then constitutes global features by tandem sub-blocks of local features.(2)To address the problem that tiny faults are challenging to identify,this thesis combines the channel attention mechanism with the parallel TCN module and proposes the TCN-SENET network model.The model pays attention to the interdependence of each channel of global features and adaptively makes adjustments to find the channels with high importance so that the extracted fault features are richer in level.Smoothing labels are also introduced to suppress the overfitting problem.The experimental results show that the described algorithm improves the fault diagnosis accuracy of TE processes.(3)To address the complex problem of multimodal fault diagnosis in TE processes,this thesis adopts the deep residual variance self-encoder and self-attention mechanism and applies the idea of feature fusion for multimodal fault diagnosis.The experimental results show that the described algorithm has a high generalization capability for multimodal TE processes.