Research On Fault Detection And Diagnosis Method Based On Kernel Convex Non-negative Matrix Factorization

As the internal connection of the industrial production process control system becomes closer and closer,the occurrence of local faults often affects the overall production.The whole body is involved in a single attack.When a fault occurs in a certain place,if timely fault detection,fault type diagnosis and fault elimination cannot be obtained,it may cause the production of the entire system to stagnate and bring a lot of economic losses.In the entire course of industrial development,various fault diagnosis methods have emerged,which are applied in different industrial production processes and different sectors.It is precisely because the application site and the basic principle are inconsistent that the existing methods have their own advantages and disadvantages.A large number of sensors and intelligent instruments can generate a large amount of data reflecting the running status of production process,and using these real-time data can complete the fault diagnosis through the built model.This kind of diagnosis algorithm based on a large amount of data training can be applied to actual systems and has great practical value.Aiming at the high-dimensional and non-linear characteristics of sampling data in an industrial production process,a fault detection method based on kernel convex non-negative matrix factorization algorithm is proposed.The innovation of this method is to reduce the dimension of the nonlinear data in the kernel space,and perform the variance scaling to keep the samples of different faults consistent in fluctuation.Moreover,further extract the local characteristics of data based on the global information,ensuring that the edge information of data can be fully utilized in the process of fault detection and recognition,in order to find the detailed changes of different data categories to promote the accuracy of detection.The method first uses the kernel function to reconstruct the original input data in high-dimensional space,and the principal component analysis method to perform whitening preprocessing on mapped data in order to eliminate the correlation between data variables and reduce or enlarge the sample variances to make the internal changes of sample more stable.Then,the convex hull non-negative matrix decomposition method is used to find the essential structure of the whitened data.At the same time,the graph regularization constraint is utilized to keep the inherent geometric structure of data set from changing throughout the decomposition process.Finally,based on the kernel convex non-negative matrix factorization algorithm,N2 ?SPE statistics are established,and the N2 and SPE control limits are obtained by the kernel density estimation method,and the samples to be tested are detected for failure.Simulation experiments using Tennessee Eastman process fault detection data show that the kernel convex non-negative matrix factorization algorithm has a better detection effect for industrial process data fault.A fault diagnosis method based on the support vector machine algorithm of kernel convex non-negative matrix factorization is further proposed.This method uses the kernel convex non-negative matrix factorization algorithm to map the original input matrix into the high-dimensional kernel space for data reconstruction,and whitens the reconstructed samples,so that the nonlinear samples can be linearly reduced dimension.The sample matrix is decomposed by the convex hull non-negative matrix factorization algorithm to obtain the basis vector matrix 'and the coefficient matrix.By using N2 and SPE statistics to build a data fault monitor,the fault data is diagnosed for the type of fault.The base vector matrix obtained by the decomposition of fault samples is used as the training data set of classifier,and the multi-classifier is obtained by training.Then,by simulation using three different types of fault data in the TE process,the effectiveness of the proposed method is verified.At the same time,compared with the classification results of multi-classifier by general support vector machine,the results show that the new method helps to promote the classification accuracy of the classifier.