Fault Diagnosis Technology Of Complex Process Industries Based On Distributed Monitoring Framework

With the high degree of automation and integration of industrial systems,the process industry has become increasingly large and complex.Traditional centralized monitoring methods suffer from poor fault tolerance and reliability,while the distributed process monitoring based method can effectively reduce the complexity of modeling with high sensitivity and robustness,so the distributed monitoring framework is a feasible solution to explore.To this end,the thesis proposes two types of targeted distributed monitoring schemes to improve the performance of existing industrial process monitoring systems in the context of the existing distributed monitoring framework and from the actual industrial requirements.To address the problems of difficult modeling of complex coupled system mechanisms and the existence of complex characteristics such as non-Gaussian among variables,a performance index-based distributed ICA-PCA model is proposed by combining PCA and ICA methods.The method is based on the distributed ICA-PCA model for industrial process fault detection.Firstly,the optimal measurement variables and process sub-blocks are selected based on the monitoring performance indicators,and the ICA-PCA monitoring model is constructed for each sub-block separately,and the statistical information of each block is integrated and fused by Bayesian inference to achieve large system fault detection.The method extracts the non-Gaussian and Gaussian information in the data,and improves the performance of the non-Gaussian process fault monitoring model with high model accuracy and adaptive chunking according to the monitoring performance.The distributed process fault diagnosis based on the continuous hidden Markov model(CHMM)is proposed for the case where the topology or mechanism knowledge of the industrial process is known.The method can reflect the dynamic dependencies contained in the sequence more accurately by dividing the system with known topological or mechanical knowledge of the industrial process and building a CHMM model library for each sub-block.Each model library can be monitored independently,and the topology of the industrial process is maintained between models.This method is suitable for monitoring time series processes with trend,seasonal or periodic characteristics,and has the features of simple modeling,small data computation,running speed block and high recognition rate.The feasibility and effectiveness of the two methods proposed in the paper are verified by the Tennessee-Eastman(TE)simulation platform.