Research On NMF - SVM Fault Diagnosis Method Based On Distance Space

The complexity of modern industrial system is more and more high, in order to reduce impacts of system failure such as the decline of the product quality and the increase of the production cost, the enterprise urgent need some process monitoring method that applicable to practical industrial system to ensure safe and effective operation of the system. On the one hand, the field bus technology and computer technology is widely used in industrial process control system, which make a large amount of data to be preserved, so the method of using these processes data establish fault diagnosis model received wide attention. But the large amount of data may also lead to a problems that huge data is a lack of information, so how to extract key information reflect the running state of system from a large number of data, eliminate the redundant information, noise and error in the process of data, is very important to establish effective fault diagnosis model.This paper first makes a brief analysis on the fault diagnosis methods currently in industrial process, and put forward some limitations of traditional multivariate statistical method performance in the practical industrial processes, such as nonlinear process, running state is not single. For the characteristics of multimode in industrial process, by analyzing the spatial distance relationship between the various modes of normal data and fault data will be mapped to the space distance, so eliminate the influence of multimode, then using non-negative matrix factorization (NMF) as a method for dimensionality reduction and feature extraction. NMF is a multi variable statistic analysis method developed in recent years, it can make more effective use of local feature information of the process data, and easy to understand and explain, and then use support vector machine (SVM) suitable to solve the problem of small sample to establish fault diagnosis model. For NMF stability is not good, using principal component analysis method (PCA) to determine the initial value of the iterative algorithm. At last, I designed a fault diagnosis system based on NMF-SVM in distance space, and it can be more intuitive, fast to get the result of fault diagnosis.This paper mainly through constructing distance space to eliminate influence on the accuracy of fault diagnosis model because of the process operation is not single, combined with NMF to feature extraction and dimensionality reduction, and using SVM to establish the fault diagnosis model, finally, through the simulation experiment on TE process data, this method compared with the basic SVM algorithm and PCA-SVM algorithm in the fault diagnosis accuracy has been improved obviously.