Research Of Fault Detection And Diagnosis Based On Semi-Nonnegative Matrix Factorization For Industrial Process

In recent years,the automation level of industrial production equipment has been constantly improved,and the production scale of modern industrial processes has shown the characteristics of large-scale and complication.If the complex industrial system fails and cannot be detected and eliminated accurately in time,it will lead to serious loss of personnel and property.In addition,the widespread application of technologies such as distributed control systems in industrial processes has produced a large amount of industrial process data.Therefore,the methods based on data-driven have been widely used in the field of fault diagnosis for industrial process.The methods based on data-driven do not need to establish accurate model of systems,and only use real-time sampled data and stored large amounts of historical data for fault diagnosis.They can be applied to various practical systems through control algorithms based on large amounts of data,so they have high efficiency and universality.Though there is a long time and rapid development of fault detection and diagnosis based on data-driven,there are still many problems unsolved.Considering the nonlinear and non-Gaussian characteristics of complex industrial processes,the fault diagnosis technology based on data-driven needs to be further investigated.To eliminate the disadvantage of low fault detection accuracy of traditional nonlinear methods,a fault detection method based on Kernel Semi-Nonnegative Matrix Factorization(KSNMF)is proposed.This method fully takes into account the characteristics of nonlinear industrial process data,the kernel method is used to describe the real nonlinear relation of original data.The KSNMF method first uses the kernel principal component analysis method to perform whitening preprocessing,including remove the correlation of data variables and scale the sample variance.The purpose of these are to reduce redundant information interference and to normalize the whitened sample admeasurement,so that the change inside the sample more stable.Semi-Nonnegative Matrix Factorization(Semi-NMF)method is used to find the partial representation of the original data,mine the intrinsic structure of the data,and embody the concept of "the parts constitute the whole".Semi-NMF method not only reduces dimensionality,but also preserves local structure of data in low-dimensional subspace.Taking Tennessee Eastman process as the research background,a new method is simulated and verified by TE process data.To eliminate the disadvantage of low fault diagnosis rate of Fisher Discriminant Analysis method,a fault diagnosis method based on Kernel Semi-Nonnegative Matrix Factorization(KSNMF)and Uncorrelated Optimal Discriminant Vectors(UODV)method is proposed.The main idea of the KSNMF-UODV method is to first use the KSNMF model to extract the local characteristics of data retained in the potential variables,and then utilize them as the training set of the UODV method to establish the fault diagnosis model and reduce the fault diagnosis rate.Finally,the feasibility and effectiveness of this method are verified by simulation for TE process.