Research On Monitoring Method Of Grinding Classification Process Based On Domain Adaptation

With the increasing complexity of industrial processes,it is becoming increasingly important to detect and diagnosis faults in industrial processes.Fault detection and diagnosis are important and challenging problems in many engineering applications and continue to be an active area of research in the control community.With the rapid development of computer technology,a large number of process data are collected and stored,making the data-driven method become the mainstream technology of today’s process monitoring.In many machine learning algorithms based on data-driven,a major assumption is that the training and the test data are drawn from the same distribution.However,in the grinding classification production process,due to the uncertainty and non-linearity of product overflow particle size,ore properties and equipment parameters,the production data of the old process(source domain)and the new process(target domain)often fail to meet the same distribution conditions.If we want to achieve new production process fault diagnosis,we can use the information of similar industrial process as much as possible to improve the accuracy of fault diagnosis and reduce the distribution difference between the fields,that is,the idea of domain adaptation is introduced into machine learning.Domain adaptation has proven to be promising in image classification,object recognition,sentiment analysis and document categorization across different customer datasets.However,the application of recognizing fault information in the industrial field is very few.Based on the analysis of the existing process monitoring methods,combined with the different data distribution but similar cross-domain fault detection and diagnosis problems in the grinding classification process,this thesis studies the process monitoring method based on domain adaptation.The main research contents are as follows:(1)We summarize the existing process monitoring methods,expound the problems faced by the data-driven process monitoring methods,and then introduce the significance of transfer learning research.We elaborated on the research status of domain adaptation problems in transfer learning,and the overall research strategy of this thesis is proposed.(2)In the modern beneficiation production process,there are usually several similar grinding and grading process production links,and there are differences in the production time of each link.The operation cycle can be flexibly adjusted according to the production load or maintenance instructions.In addition,due to changes in production conditions,production materials and equipment parameters,the distribution of data in the production process changes.For new processes with less production experience and historical process data,it is difficult to establish a monitoring model directly for the process,and the accuracy of the prediction model is not ideal.At this time,over-fitting and negative transfer are easy to occur.The extreme learning machine domain adaptation model is introduced in this thesis.The method projects the extreme learning machine classifier parameters of the target domain into the source domain parameter space,so that it is the same as the classifier parameter distribution of the source domain.In addition,the parameter approximation term and regular term constraint are introduced in the objective function,considering the possibility of a negative migration and overfitting in the migration.The advantage of this algorithm compared with the previous domain adaptation algorithm is that its classifier parameters and transfer matrix are simultaneously optimized,and the objective function solving process is relatively simple.We verify the practicability of the method mainly for the actual grinding grade production process data,and also use the typical TE process data to verify the effectiveness of the research method.(3)The inherent non-linear characteristics of the grinding and grading production process seriously affect the monitoring effect,and in the actual production process,because the new process is not put into production for a long time,the available tag data cannot be provided for establishing the supervised model.Aiming at this situation,this thesis proposes a process monitoring method based on joint distribution adaptation and manifold alignment.The method learns two projection matrices that map the monitoring data of source domain(labeled data)and target domain(unlabeled data)into two low-dimensional subspaces where the distributional shift and structural shift are reduced simultaneously.And the manifold alignment and kernel method are introduced into the domain adaptation,which can preserve the neighborhood relationship within each set and make the distance of the corresponding points in the projection coordinates as close as possible.It is also suitable for the complex process with nonlinear trait.Finally,the practical verification of the actual grinding classification process data is carried out,and the effectiveness of the method is verified by TE process.