Research On Data-driven Methods For Industrial Process Monitoring And Fault Diagnosis

Industrial production occupies a leading position in national economy.As an important tool to ensure the reliable operation and safe production of industrial processes,process monitoring and fault diagnosis technology have always been a popular research topic in the field of control.Today,modern industrial processes are constantly developing in the direction of scale,integration,complexity and intelligence,and process monitoring and fault diagnosis are constantly facing new difficulties and challenges.With the development of supervisory control and data acquisition technology,monitoring data in modern industrial processes has become extremely abundant and easily accessible,and data-driven process monitoring and fault diagnosis methods have been widely focused and researched.Based on deep learning and graph theory techniques,this thesis proposes a process monitoring and fault diagnosis method that can be used for high-dimensional nonlinear industrial processes.The main research contents of this thesis are as follows:(1)The concepts and methods related to process monitoring and fault diagnosis are discussed in detail,the theory related to Variational autoencoder and Bayesian networks is analyzed,and the structure and characteristics of the Tennessee Eastman benchmark(TE)simulation process are discussed.(2)A process monitoring method based on a Variational autoencoder is proposed for the characteristics of high dimensionality,non-Gaussian and non-linearity of process data,which can use process monitoring data to determine whether the process is operating normally.A probability model is established using the Variational autoencoder to map the real-time collected process data to the latten space;The process monitoring statistic is designed by the feature that latent variables conform to Gaussian distribution,and the threshold value of the statistic can be calculated by the kernel density estimation method;By mapping the process data to the latent space,the parameters of the probability distribution of the mapped data are used to calculate the monitoring statistic and compare it with the threshold value to determine the process operation status,and the process monitoring is realized.Finally,the proposed method is validated using TE process data.(3)For the fault location problem,a fault diagnosis method based on Bayesian networks is proposed.The model building of Bayesian network is a prerequisite for using Bayesian networks to complete fault diagnosis.Combined with the scoring search idea of Bayesian network structure learning,this thesis proposes a Bayesian network structure learning algorithm based on the discrete whale optimization algorithm.Firstly,the initial population is obtained using mutual information and maximum spanning tree,then use the scoring function of the Bayesian network as the fitness function,and use the discrete whale optimization algorithm to obtain the Bayesian network with the highest score,so as to obtain the optimal Bayesian network structure.Finally,the effectiveness of the structure learning algorithm is verified by the Asia dataset,and the algorithm is applied to the TE process to construct the Bayesian network structure of the TE process.The kernel principal component analysis method is used to determine the evidence variables,and the inference capability of the Bayesian network is used to achieve fault location.