Research On Fault Monitoring Method Of Complex Industrial Process Based On Transfer Learning

As an important measure to ensure process safety and product quality,industrial production process monitoring has attracted more and more attention with the large scale and complexity of industrial production process.The actual fault detection and diagnosis problem is the classification problem.Then,different monitoring methods are put forward to classify the actual industrial process faults.With the rapid development of computer technology and instrument technology,a large number of production data are stored in the industrial process,thus the data-driven process monitoring method has been widely studied and applied.However,data-driven methods usually require training data and test data to satisfy the condition of independent and identically distribution.In the actual industrial production process,due to different production conditions,raw materials and equipment parameters,the old process data and the new process data can not meet the condition of independent and identically distribution.At the same time,the existing statistical learning method and machine learning method can not accurately monitor the new process.Transfer learning can make use of the production experience and expert knowledge of old process to guide the new process to establish monitoring model,which effectively improves the efficiency of learning and reduces the waste of knowledge.In this thesis,the process monitoring method based on transfer learning is studied combined with the monitoring problem of the TE production process and the grinding classification production process.The main research contents are as follows:(1)Firstly,the process monitoring methods are summarized,and then the research status of data-driven process monitoring method is analyzed.This thesis expounds the advantages and disadvantages of the statistical class process monitoring method and the traditional machine learning method,and then leads to transfer learning,and expounds the research status and related methods theory of transfer learning,and puts forward the overall research strategy of this thesis.(2)According to the actual production process,due to the changes of plant conditions,raw materials and equipment parameters and other conditions leading to the production process data distribution changes,it is difficult to establish a monitoring model directly for the new process with less production experience and less historical data,and the monitoring accuracy of the model is not ideal.The transfer learning model based on graph regularization is introduced in this thesis.According to the experience and expert knowledge of the old process,the definition of the labeled data and the state label matrix of the sample is given.Using the improved algorithm of non-negative matrix factorization to extract common latent factors between process data,a more accurate basis matrix is obtained,and a transfer learning bridge is established.At the same time,the neighborhood weighted graph method is used to calculate the sample similarity,and the local and global consistency information of the process data is fully preserved,so as to alleviate the negative migration.By solving the objective function,the state label of the monitoring data is obtained,and then the running state of the process is judged.The effectiveness and practicability of the method are validated by using TE process data and actual grinding classification production process data.(3)In the actual industrial production process,the production data contains noise because of the malfunction of production equipment and the mistake of human operation.And it will seriously affect the monitoring effect.In view of this situation,this thesis proposes a process monitoring method based on transfer learning with anti-noise performance.The method uses low rank representation method to map the old process data and the new process data into the common subspace,in which the old process data and the new process data satisfy the same distribution and the new process data can be expressed linearly by the old process data.And the noise matrix can be obtained in the process of mapping,so as to achieve the effect of anti-noise.The effectiveness and practicability of the proposed method are verified by simulation experiments using TE process data and actual grinding classification process data.