Research On Approaches Of Fault Diagnosis Based On Bayesian Network And Multivariate Statistical Analysis

In modern manufacturing industries, with the rapid development of mass production and complexity, in order to ensure the processes long-term running safety and reliability, process monitoring and fault diagnosis is becoming one of the most important and challenging research area in the filed of process control. The complexity manufacturing system makes the history data increasing and the relation between variables getting more complicated, and to some extend, grasping the accurate casual relations between variables would affect the diagnostic accuracy. Multivariate statistical technique is an important branch in the research of fault diagnosis but is weak in diagnosing the root causes of the system change. The fault diagnosis for the variables with casual relations is studied in this research. The research work can be summarized as follows:Firstly, combining the engineering domain knowledge with the generic PC algorithm, a typical algorithm to learn the structure, an integrated modeling approach is proposed to discover the causal relationships. The causal relationships are represented by a causal Bayesian network, in which engineering domain knowledge is embedded in various critical stages of a generic causal discovery algorithm. Secondly, according to the constructed casual Bayesian network and the diagnostic methods based on Multivariate statistical technique, a new method is proposed by effectively integrating Bayesian network with the T2 decomposition method. The method can be realized through three steps: establishing multiple linear regression models based on Bayesian network, detecting the diagnosis using Hotelling T2 control chart, and converging all decompositions to a unique decomposition, meanwhile, the control limits for the terms of the decomposition are given.Finally, the application of the method in actual cases shows that the proposed method is effective and feasible and its performance is better than univariate control charts, PCA method and MYT decomposition method in the aspect of two different situations. In addition, the results of many simulation experiments indicated that the proposed method is better than MYT decomposition method in respect of diagnostic accuracy and it is more convenient.