Research On Industrial Process Monitoring And Fault Diagnosis Based On Machine Learning Algorithm

With the development of industrial technology,modern industrial processes present the characteristics of large-scale,high complexity,and high production requirements.Under this trend,any abnormality caused by instrument or equipment may affect the quality of the product and cause damage to the equipment,even lead to a safety accident.With the combination of modern industrial technology and information technology,real-time data collection,cloud database,online monitoring and other functions have been realized.But these functions are only suitable for monitoring simple industrial processes.In order to solve the process monitoring problems of large-scale industrial processes,it is necessary to find the abnormality in industrial process in time and prevent the occurrence of chain reaction.It is extremely important to monitor the industrial process and to make the research of fault diagnosis algorithms.In the research of process monitoring and fault diagnosis methods,the multivariate statistical process monitoring(MSPM)method is often applied to complex industrial processes due to its data-driven characteristics.The MSPM method usually is used in modeling for the different characteristics of the research object.Linearity and nonlinearity are two common characteristics in industrial processes.In this article the process monitoring and fault diagnosis of linear industrial processes and nonlinear industrial processes are studied.Principal component analysis(PCA)is a common method used in linear industrial process.In fact in PCA algorithm the interpretation is poor to the principal component.In order to solve the problem,the SPCA algorithm has been studied widely.The SPCA algorithm has low information utilization rate in residual space,and its statistics have problems in fault determination.To improve this method,a process monitoring and fault diagnosis method based on improved SPCA is proposed in this paper.In the improved SPCA method Lasso and Ridge penalty terms were used to optimize the principal component vector and to sparse the principal component vector.In this way the improved SPCA algorithm enhances the interpretation of principal component and improves the information utilization rate of residual space.To reduce the miss diagnosis rate and improve the model accuracy,the square prediction error in the improved SPCA algorithm is split into principal correlate variables and common process variables.And this is realized by the multiple correlation coefficient of process variables.By doing this the process monitoring can be realized better.In nonlinear industrial processes,the interval kernel principal component analysis(IKPCA)algorithm successfully solves the nonlinear and uncertain problems.But the accuracy of this method is low.Through research and analysis,it is found that there is a problem in the selection of Gaussian kernel parameters in this method.To solve this problem,an interval ensemble kernel principal component analysis(IEKPCA)algorithm is proposed in this paper.IEKPCA algorithm introduces the idea of ensemble learning to build multiple sub-models to solve the problem of Gaussian kernel parameter selection.However,multiple sub-models will produce multiple results,which will complicate the calculation of the final result.In order to simplify the calculation,the Bayesian decision is used to convert the result into the fault probability,and the final results are obtained by weighting.In order to verify the two algorithms proposed in this paper,the two algorithms are applied in the Tennessee Eastman process.And the corresponding experiments are designed to verify the performance of the algorithm.The evaluation of the algorithm is mainly based on the performance indexes such as accuracy rate,false alarm rate and missing alarm rate in process monitor.Through the experimental results and comparison with other algorithms,the effectiveness of the proposed process monitoring and fault diagnosis algorithm is verified,and it is proved that the two methods optimize the process monitoring and fault diagnosis problems of linear and nonlinear industrial process.