Research On Concept Drift Detection And Application For Industrial Process

The data-driven modeling method has been widely used in the soft measurement of difficult-to-measure parameters such as product quality and pollution emission in industrial processes.The inherent concept drift phenomenon in these complex processes leads to a decrease in model accuracy.Concept drift is mainly caused by factors such as equipment wear and production environment changes in the industry,which is difficult to predict and quantify.In order to improve the performance of the updated model,it is necessary to effectively identify process concept changes and accurately determine new samples that can characterize concept drift.However,existing concept drift detection methods still have many unresolved research problems in drift characterization ability,detection timeliness and weakly supervised learning.Moreover,most methods are developed around classification tasks,which are lack of application analysis of regression measurement in actual industrial processes.Therefore,the research on the concept drift detection algorithm for industrial processes is of great significance to improve the accuracy of industrial soft-sensing models.To solve the above problems,concept drift detection methods such as distribution hypothesis testing supported by difficult-to-measure parameter errors based,comprehensive estimation index based,and sample output and feature space based are designed in this paper.The proposed methods are applied to the municipal solid waste incineration(MSWI)process.Moreover,the concept drift detection APP for pollutants discharged from the MSWI process is designed and developed.The main research of this paper are as follows:(1)Design and research of concept drift detection method based on distribution hypothesis testing supported by difficult-to-measure parameter errors.The method based on measurement errors of difficult-to-measure parameter have weak concept representation capabilities,and existing researches lack the analysis of sample distribution in regression tasks.In view of the above problems,a concept drift detection method based on distribution hypothesis testing supported by difficult-to-measure parameter errors is proposed.First,support vector regression(SVR)is used in the outlier sample detection window to obtain the outlier samples contained in the real-time process data.Then,the Euclidean distance between the outlier sample and the historical sample set is calculated in the distribution detection window.Next,a test drift index combined with a variety of distribution test methods that can characterize the distribution changes contained in outlier samples is defined,so as to realize effective identification of drift samples.Finally,synthetic and real industrial process data sets are used to verify the effectiveness of the proposed method,which shows better performance than existing methods.(2)Design and research of concept drift detection using a novel comprehensive estimation index.Concept drift detection in industrial processes requires high timeliness,and there is one-sidedness in the method based on sample distribution hypothesis testing.Aim at these problems,a new concept drift detection method using a novel comprehensive estimation index is proposed.At first,the online measurement output of the new sample is obtained by the model based on historical samples.Then,the measurement error and sample distribution evaluation indices are calculated in two different windows,respectively.Further,a new comprehensive estimation index is defined by weighting the above two indices.Finally,the concept drift samples identified by the novel index are combined with the selected historical samples to update the measurement model.Simulation results based on synthetic,benchmark and real industrial process data sets show the effectiveness of the proposed method.(3)Research and research of semi-supervised concept drift detection method by combining sample output space and feature space.The existing methods are limited by the difficult-to-measure parameters' true values,which are difficult to be effectively applied to industrial processes.Thus,a semi-supervised concept drift detection method by combining sample output and feature space is proposed.First,unsupervised mechanism based on principal component analysis(PCA)is used in the sample feature space to identify concept drift samples.Then,semi-supervised mechanism based on temporal-difference(TD)learning is used in the sample output space to label the pseudo-true value for the identified concept drift samples.Further,the Page-Hinkley detection method is used to confirm the concept drift samples.Finally,the new samples obtained by the above steps are combined with historical samples to rebuild the measurement model.The simulation results based on synthetic and real industrial process data sets show the proposed method has better performance than existing methods.Moreover,the cost of sample annotation is effectively reduced and the drift adaptability of the measurement model is enhanced.(4)Development of the concept drift detection APP for pollutants discharged from municipal solid waste incineration process.First,the APP has been analyzed from the perspective of users and functions,and its main functions have been determined,that is,allowing users to easily and efficiently detect concept drift and update the measurement model.Secondly,the development plan and technical route were designed and perfected.Among them,the concept drift detection algorithm and the basic measurement model are realized by Matlab on the PC side.The former detects the changes of sample output and feature space based on comprehensive estimation index,and the latter calculate the absolute and relative measurement errors of the samples based on the Gaussian process regression(GPR)model.The main function of My SQL database is to store account information,process data and algorithm parameters.Through its communication with Intelli J IDEA,functions such as user login and registration,process data reading,algorithm parameter setting,and detection result display can be realized on the Android user side.The concept drift of MSWI process data can be detected by users through the APP,and the measurement model can be updated with one click.This APP uses Java programming,and it is developed by the tools of Android Studio,Matlab,My SQL and Intelli J IDEA.