Research On Process Monitoring Method Of Complex Industrial Based On Model Transfer

In recent years,the industrial production process is becoming increasingly complex,process monitoring technology has attracted much attention,and a variety of process monitoring methods emerge in endlessly.With the rapid development of computer technology,a large number of industrial process data are collected and stored,and the data-driven method has unique advantages.Data-driven methods usually require training data and test data to be independent and identically distributed.However,in the actual industrial production process,due to the changes in production conditions,raw materials and equipment parameters,training data and test data often can not meet the conditions of independent and identical distribution,and the existing statistical learning methods and machine learning methods are no longer applicable.Transfer learning make use of the old process production experience and expert knowledge to guide the new process to establish a monitoring model,effectively improving the efficiency of learning while reducing the waste of knowledge.As a kind of transfer learning,model transfer has been widely studied for its advantages of fast modeling and less requirement for target domain samples.Combining with the problems of fault detection and diagnosis in TE(Tennessee Eastman)production process and grinding and classification production process,the process monitoring method based on model transfer is studied in this thesis.And the main research contents are as follows:(1)Firstly,the related methods of process monitoring are summarized,and the transfer learning is introduced.The existing model transfer and related methods are studied in detail,and the overall research strategy of this thesis is proposed.(2)In view of the actual industrial production process,because of the change of production conditions,raw materials and equipment parameters,the distribution of production process data changes and becomes a new process,and the new process is likely to have some new faults,but the number of new faults can not be predicted.In order to solve the problem,this thesis proposes a new LSSVM(Least Squares Support Vector Machine)incremental algorithm,and improves the existing model transfer framework.Combining them,a process monitoring method based on incremental model migration is obtained,which can monitor all new faults in the new process and set up solutions to the problems caused by this kind of class extension.In the implementation of the method,the new process fault can be regarded as a linear combination of all known faults,and the incremental learning method can be used to diagnose the faults sequentially;the marginal probability distribution of the two processes can be minimized to improve the transfer efficiency;the state labled matrix of monitoring data is obtained by solving the objective function,and then the state of the process is settled.Finally,the effectiveness and practicability of the method are verified by TE process data and actual grinding and classification production process data.(3)In the actual industrial production process,due to the lack of experience and knowledge of the new process,the labeled samples are limited,the normal state and various fault state data may be seriously unbalanced,the monitoring effect is easy to be affected.In view of the situation,this thesis proposes a model transfer process monitoring method to solve the problem of sample imbalance.Firstly,this method improves the existing adaptive weighting framework,expands the two kinds of adaptive weighting framework to multiple classes,redesigns the weighting mechanism,and then combines it with the model transfer framework,realizes the model transfer by aligning the output weights of the new process model and the old process model.At the view of data,SMOTE(Synthetic Minority Over-sampling Technology)algorithm is used to generate partial pseudo-data for a small number of samples;At the view of algorithm,the small number of samples,including the pseudo-data generated by SMOTE algorithm,have different contributions to classification.It is necessary to set the best weight for each sample data,give the samples with large guiding significance a larger weight,and reduce the weight of the samples without guiding significance or with small guiding significance.Finally,the effectiveness and practicability of the method are verified by simulation experiments using TE process data and actual grinding and classification process data.