Research On Fault Monitoring And Diagnosis Based On Probabilistic Graphical Model In Chemical Process

With the development of modern industries,the chemical process is becoming more and more complex and large-scale.In order to make sure the large-scale production process operates smoothly,the plant-wide monitoring has become an essential technical approach.However,how to efficiently monitor the state of production process and diagnose the root causes of abnormalities is a very important and complex problem.The accurate and effective production process variable relational network construction is the basis of the plant-wide monitoring and fault diagnosis analysis.Firstly,according to the characteristics of the large-scale chemical production process,which has numerous variables and strong coupling data,this thesis proposes a MIC-CMI-GC-based hybrid score search constraint network structure learning method,which integrates maximum information coefficient,conditional independence analysis and Granger causality test and combines the corresponding score function,loop detection structure optimization techniques.The proposed method effectively solves the problem of easy to obtain local optimal solution and large search space in the traditional structure learning method based on score search.Secondly,based on the proposed network structure learning method and the expectation-maximization-based parameter learning algorithm,a static Bayesian network model is constructed for the static production process.Then,on the basis of static Bayesian network,the dynamic Bayesian network model is determined through applying the concept of time-slice and analyzing the time-varying data to construct the transfer network for the dynamic process,which effectively solves the problem that variables have strong autocorrelation in chemical production process and improves the accuracy of fault monitoring and diagnosis results.Thirdly,in view of the problem of lack of fault samples in the actual chemical process,this thesis adopts the idea of data-missing hypothesis and uses Bayesian network inference machine to realize data reconstruction.Through calculating the T~2 and SPE statistics of each variable and constructing variable-fault index and whole-process fault index,the monitoring of chemical production process can be realized.According to the abnormal state,the fault contribution index of each variable is calculated,and the evidence variable is obtained,and the fault root is traced by using Bayesian network model.The validity and practicability of the proposed fault monitoring and diagnosis method are verified by TE process.Finally,due to the high professional requirements of Bayesian network modeling and the difficulty of using it to solve uncertain problems,"Chemical process fault monitoring and diagnosis analysis platform based on Bayesian network"is developed.The monitoring diagnosis Bayesian network model and application are deployed on the server and client respectively,which realizes the functions of online monitoring and analysis,fault root diagnosis and factory management.The software improves the usability of Bayesian network model,and provides technical support and decision-making assistance for operators.