| Safety and reliability in industrial production processes are crucial.Considering the complexity of modern industrial systems,traditional physics-based and knowledge-driven methods are too time-consuming and difficult to build factory fault detection models.Data-driven fault detection and diagnosis techniques do not require complex a priori knowledge,and only need to monitor the abnormal conditions of the whole system by identifying and locating fault data,which is less costly in the fault detection process and has gained process monitoring field It has gained wide attention in the field of process monitoring.Thanks to the correlation among variables in the process of plant monitoring,researchers try to find the relationship between process variables and the final output quality variables,and divide the faults into quality-related and quality-independent parts through matrix decomposition and detect them separately,which can avoid unnecessary warnings and thus reduce economic losses.Quality-related fault detection and diagnosis based on quality is currently a hot topic in the research of process monitoring algorithms.The specific work in this thesis is as follows.(1)Most of the current process monitoring algorithms are based on statistical analysis of kernels and deal with the nonlinearity of the raw data through kernel methods,such as kernel principal component analysis(KPCA),kernel partial least squares(KPLS)and kernel typical correlation analysis(KCCA)algorithms.The mapping of kernel functions makes it easier to distinguish process variables from mass variable correlations,so the space after kernel mapping is divided into two parts,mass-related and mass-independent,by relevant techniques such as direct decomposition and singular value decomposition,and monitored separately.Considering the iteration period and the long time spent in the decomposition process and the fault detection timeliness,this thesis proposes the kernel principal component regression algorithm without iteration,which not only saves the detection time but also has higher fault detection accuracy.(2)With the increasing number of data dimensions and types of faults,the ability of kernel methods to deal with nonlinearities gradually decreases because the parameters of the kernel and the number of principal elements change according to the types of faults,lacking universality.Under the positive influence of the wave of artificial intelligence in recent years,the research based on deep learning has been flourishing.Supported by the training of massive data,deep network models have excellent performance in nonlinear fitting,such as deep neural networks(DNNs).They also have the ability to capture implicit correlations,such as convolutional neural networks(CNN),recurrent neural networks(RNN),and long-short time neural networks(LSTM).In this thesis,the accuracy of fault detection is further improved by introducing neural networks applied to process monitoring.The characteristics of the algorithms and faults are also analyzed to obtain the reasons for their high accuracy.(3)The current depth framework makes it difficult to obtain the relationships between different variables and time series of industrial systems simultaneously,which reduces the accuracy of fault detection and diagnosis.Considering the great time series processing potential of LSTM,a new variational self-encoder(TVAE)based on dual-view LSTM embedding is proposed in this thesis.Firstly,the data measured by the sensor in time period T are obtained through a sliding window,and then the time series and variable dependence information are encoded separately using the dual-view embedding to obtain the time-dependent as well as variable-dependent features of the measured data,and then the abnormal scores of the variational self-encoder are used to determine whether they are faulty data.Since the relative positions of the original variables are changed in the neural network mapping process,the feature decomposition of the encoding matrix is not possible,but the encoding process makes the fault features more obvious,so this thesis identifies its fault features by the defined neuronal group convolution and performs classification diagnosis.A new paradigm for industrial process fault detection and diagnosis is provided.Experiments on the Tennessee Eastman process show that the TVAE method outperforms existing methods.With the rapid development of industrial processes,safety and economy in production processes are highly valued,and the need for quality-related hierarchical detection is increasingly evident.Data-driven process monitoring algorithms are only studied for historical plant data,and there is ample research value in the rational use of this data for real-time hierarchical detection of faults in modern industrial systems. |
