| With the development of modern process industrial processes towards large-scale,continuous,and integrated processes,the production processes are becoming increasingly complex,with various factors interdependent,posing great challenges to real-time monitoring and control in process industries.In modern process industries,a large number of process variables are collected,which are usually complex and intertwined,making it difficult to monitor a process across the entire plant.Traditional centralized process monitoring methods cannot meet the requirements of real-time,reliability,and safety.Therefore,data-driven distributed process monitoring has gradually become an important solution.This thesis is based on the existing distributed monitoring architecture,starting from practical industrial needs,and combining mechanism knowledge and machine learning algorithms to conduct theoretical research and simulation verification on process variable decomposition,nonlinear dynamic data modeling,and fault diagnosis problems under the process industry’s distributed monitoring architecture,as follows:(1)Starting from the historically collected industrial data,a distributed process monitoring method based on MI-Louvain decomposition and SVDD diagnosis is proposed to effectively monitor the entire plant process.Firstly,the entire process of the plant is mapped to an undirected graph model corresponding to mechanism knowledge and process structure.Mutual information(MI)is introduced to describe the correlation between different nodes,and a Louvain algorithm with MI correlation is proposed to finely decompose the process into reasonable sub-blocks.This method combines mechanism knowledge and data analysis to improve the rationality of process decomposition.Then,based on the support vector data description algorithm(SVDD),a fault detection model is constructed in each sub-block,and a corresponding variable contribution rate calculation method is designed to locate fault variables.Finally,based on Bayesian fusion decision,the detection results of all sub-blocks are comprehensively evaluated.The designed process monitoring method can extract non-Gaussian,nonlinear information from process data,improve the fault detection performance of nonlinear processes,and realize fault variable localization.(2)In view of the problem that the above monitoring scheme fails to fully consider the dynamic characteristics of data,a distributed dynamic process monitoring method is further proposed.The specific improvement strategy is reflected in the following two aspects: firstly,in the process variable decomposition problem unit,after mapping the entire plant process to the undirected graph model,copula entropy(CE)is used to estimate the weights between nodes in the graph model instead of MI.This method divides process variables into reasonable sub-blocks because CE has the advantage of estimating the correlation between variables more concisely than MI.Secondly,in the detection unit of each sub-block,a fault detection model based on dynamic recursive support vector data description(DR-SVDD)is constructed.This model extracts nonlinear dynamic information from process variables.Compared with the process monitoring method based on static SVDD detection model mentioned earlier,it effectively improves the detection rate of the entire plant process faults.The thesis verifies the effectiveness and feasibility of the two distributed monitoring methods,static and dynamic,proposed in this thesis based on the Tennessee Eastman(TE)process. |
