Research On Fault Detection Method Of Chemical Equipment Based On Probabilistic Graphical Model

With the digital transformation of the manufacturing industry and the continuous progress of technology production,chemical equipment is showing a trend towards larger,more complex,faster,heavier,and highly automated processes which leads to increasingly complex chemical processes.If a malfunction occurs,it will bring direct or indirect significant economic losses to the enterprise and may potentially endanger people’s lives,resulting in very serious consequences.Fault diagnosis helps technicians to detect,isolate,and identify faults,and carry out troubleshooting.Bayesian networks are a probabilistic graphical model effective in handling various uncertainty problems.Data-driven fault detection provides effective scientific theoretical guarantees for accurate fault detection.In response to the characteristics of non-stationary time-series data in chemical processes and the application requirements of fault detection,this paper proposes a complete fault detection process for time-series data in chemical processes.It includes data modeling,fault detection based on dynamic Bayesian networks(DBN)combining multi-variable processes,root cause diagnosis,and fault propagation path recognition scheme.The proposed method generates evidence from monitored process data and uses this information to update process knowledge in the DBN.For detection,this paper proposes a new fault detection method based on Bayesian network marginal probability.After detecting a fault,the fault propagation path is identified based on the causal relationship between process variables.The main research is as follows:1.To deal with the non-stationary characteristics of time-series data in chemical processes,a hidden Markov model based on Euclidean distance of data points is proposed for data modeling and analysis.Fragmentation processing of data is performed to provide a foundation for fault detection in chemical processes.2.Bayesian structure learning based on Bayesian scoring and Markov chain Monte Carlo sampling is studied.A particle filter-based Markov chain Monte Carlo sampling algorithm is proposed and its performance is verified on a public dataset.Compared to other algorithms of the same type,the proposed algorithm improves the reconstruction accuracy of Bayesian structure learning and the convergence speed of the sampling method.3.A fault detection model for chemical processes based on non-homogeneous dynamic Bayesian networks is constructed for the Tennessee Eastman(TE) process in the chemical industry.Whether a fault occurs is determined by the average marginal probability of the Bayesian network and the fault node is located.Furthermore,the fault propagation path is inferred through the structure of the Bayesian network.