Applications Researches On Fault Prediction Of Complex Industrial Processes Based On Extended Finite State Machine

With the development of industrial technology, the demands for safety assurance are improved constantly. So, how to prejudge the safety problem on the basis of safety evaluation and effective protection become very important in industrial processes. As the time of data mining is coming, the rising of artificial intelligence brings new vigour and hope to the traditional predictive field. As industrial production process has grown ever more complex and large-scale, multi-fault maybe more practical in industry processes and can damage much more than a single fault. Therefore, the perfection of the prediction and the improvement of predictive accuracy has become an urgent problem which need studying on this topic. Aiming to solve the problem on data processing with nonlinear properties and multi-fault prediction, the following research has been done on the improvement of the fault prediction method:First of all, there is additive noise in industrial data, and the data change caused by fault of some variables is so small. Considering all that, the mean-filtering using moving window is adopted to undo the noise which generated by instrument equipment and unknown input disturbance. Mean-filtering using moving window also can ensure the real-time and accuracy of data acquisition in industry system. The time-delay mutual information is applied to do the cross-correlation analysis and time-delay analysis. Secondly, a feedback differential evolution optimized extreme learning machine (FDE-ELM) with time delay-based extended finite state machine (TD-EFSM) method was proposed in this article to solve the single fault prediction problem. FDE-ELM is a network model for non-linear industrial process that is proposed to record the time series information for industrial data, so that the prediction accuracy can be improved. The time-delay dependent network is built according to the result of TDMI method, and the network is introduced to the FDE-ELM model. Then, compare the prediction result with the control limit. If there are any system variables beyond the control limit, the EFSM was activated for the fault reasoning and recognition. After recognition, the fault can be predicted precisely. The simulation experiment results on TE process demonstrate the validity of this method. Thirdly, based on the research of single fault prediction, the research for multi-fault prediction has been done. A Time-series extended finite state machine (TS-EFSM) based relevance vector machine (RVM) multi-fault prediction approach was proposed in this article. RVM is a prediction method with good comprehensive performances, which just needs less original data and gives a higher prediction precision. RVM has better stability and wider applicability than ELM, and it take less predictive time than FDE-ELM. Because of the complexity of multi-fault prediction, the composition operators were introduced to enhance the accuracy of reasoning. The results of TE simulation show that the proposed multi-fault approach can achieve the goal of multi-fault prediction successfully.According to the research results, all the fault prediction approaches based on EFSM performance well. The EFSM not only can recognize the fault source precisely but also can do visualization of reasoning procedure, and visualization is a valuable reference for operators.