Research On Fault Feature Extraction And Diagnosis Of Rotating Machines Base On Independent Component Analysis And Generative Topographic Mapping

With the development of the fault diagnosis technology researches and applications, the fault feature extraction is found to be the most important and difficult problem, which is the bottle-neck problem of the fault diagnosis technology and relates directly to both the veracity of diagnosis and the reliability of early diagnosis. To solve the fault feature extraction problem completely, theories, methods and techniques for information processing, especially modern signal processing, have been depended on, and new methods, new theories and new techniques for fault feature extraction and diagnosis are being researched.In the modern signal processing technology, Blind source separation(BSS) technology , Independent Component Analysis (ICA) and a neutral network of Generative Topographic Mapping(GTM) are two kinds of non-linear mapping processing. Based on these conditions, this paper describes the exploratory work of the fault feature extraction and diagnosis in rotating mechanical systems with Independent Component Analysis and Generative Topographic Mapping. The main contents of this paper include:(1) The research content of fault diagnosis for rotating machines, which topic are fault information detecting ,state inspecting ,fault feature analysis, fault mechanism and fault recognition neutral network are summarized. Two kinds of fault feature extraction and diagnosis methods, including Independent Component Analysis (ICA) and a neutral network ,are briefly introduced in this thesis;(2) The ICA mathematical mode and its Nongaussian estimation theory are summarized. Preprocessing method is analyzed so that the initial signals suffice to the Nongaussian condition of applying ICA .Three ICA algorithm using differently Nongaussian measurement criterions which are kurtosis, negative entropy and mutual information for convergent condition , are introduced;(3) Being taken the typical rotating machines roll bearing as research objects, the initial signals is acquired from roll bearing vibration sensor. With low-pass filter, centralization and whitening preprocessing, the practical FastICA algorithm is applied to roll bearing fault feature extraction and separate a few of dependent components successfully. By the result of frequency analysis for each dependent component, roll bearing fault feature is extracted successfully. This experiment show that ICA algorithm can extract fault feature of rotating machines effectively;(4) The GTM model is studied to focus on expectation-maximization (EM) algorithm, magnification factors and visualization principles. For the GTM model, the probability density of data in a space of several dimensions can be determined using the EM algorithm in terms of a smaller number of latent variables in lower-dimensional space together with a predefined discrete form for the prior distribution. An approach to rotating mechanical fault recognition based on the GTM model is presented. This approach is applied to roll bearing fault feature extraction. It follows from the investigation that fault can be distinguished from the normal conditions in roll bearing.Finally, this thesis and solved problems are summarized, and some future research areas of fault feature extraction and diagnosis technique for rotating machines are highlighted.