Ensemble Learning Based Industrial Process Monitoring

With the development of modern industry,the requirements for industrial process safety,product quality and economic benefits keep improving.Therefore,process monitoring has been one of the important research issues in the field of automation.As the industrial process becoming complicated and scaled,a large number of sensors are widely deployed on equipment,enabling massive production process data to be collected.At the same time,with the development of cloud computing,big data,artificial intelligence and other information technologies,data-driven methods has become research hotspots in the field of process monitoring.In recent years,a large number of data-driven process monitoring methods have been proposed.But these single methods are based on certain data assumptions and cannot be well monitored under different conditions.Therefore,ensemble learning method is proposed to deal with the above problem.A variety of effective process monitoring methods and overall integration model framework are proposed based on the research of base model performance difference,number of base models and model accuracy in the ensemble learning.The details of the research are organized as follows:(1)The impact of different base models on the integration result are not considered in most of the ensemble learning models.The different model results are treated with equal weight in the process of the fusion.Therefore,in view of the above problem,an analytic hierarchy process based model evaluation method is proposed,which considers multiple criteria to prioritize multiple models.The evaluation result is used as the model weight and combining with the Bayesian posterior probability for fault detection and classification.The effectiveness of the proposed method is verified on Tennessee Eastman process.(2)In the process of ensemble learning,more base models cannot obtain better result.On the contrary,a single model with poor performance may cause a negative impact on the final result.Therefore,an ensemble selection based improved random forest method is proposed for fault classification.Firstly,considering the diversity between base models and the performance of the single model,a static ensemble selection method is proposed.Some decision trees in the original random forests are selected to construct a new random forests for online fault classification.Secondly,considering the difference between online samples,a dynamic ensemble selection method is proposed.Different new random forests are constructed for different online samples,and a weighted probability fusion method is proposed to replace the original majority voting method.The effectiveness of the proposed method is verified by a simulation platform.(3)In order to further improve the accuracy of ensemble learning method,the deep learning model and the ensemble learning model are combined to generate a deep ensemble f-orest method for fault classification.This method expands the single-layer integration model into a multi-layer integration model,which can effectively extract feature information and improve classification accuracy.Three forests are constructed in each layer of the model,and then the feature extraction result of each layer is stacked with the selected features from original features.The new vector is the input of the next layer.The model training procedure is terminated when the model performance is no longer improved.The effectiveness of the proposed method is proved by a comparative study on Tennessee Eastman platform.(4)A systematic ensemble learning framework for industrial process monitoring is proposed in this paper.Firstly,the model selection is carried out based on the data characteristic analysis and task identification.Then,the selected models are evaluated by multi-criteria decision making method.Finally,the results of multiple models are combined by the decision fusion method to obtain the final result.This framework provides a systematic operational procedure of data-driven process monitoring for operators and engineers.According to the simulation result,it can be seen that the monitoring performance can be effectively improved by propose method.Finally,all the research work is summarized and the future work of the ensemble learning in process monitoring is prospected.