Data-driven Multivariate Statistical Fault Diagnosis Method And Application

With the functions gradually improved and the structures increasingly complicated, the large-scale automation systems are the types of important technical equipment being used extensively in national defense and national economy realms, which safety and reliability become crucial.It is quite difficult to establish an accurate physical model being used to fault diagnosis for such systems in which there are strong nonlinear on their condition behaviors and serious state coupling among the different parts and being subject to the non-gauss noise and various indetermination factors.Since the non-contact measurement technology and the soft measurement technology and a variety of intelligent instruments in the wide spread of industrial processes are excellently evolving, the massive process data which reflecting the operational status of systems is collected and stored. However, there is a "data rich, information plaque lack of "embarrassing situation in industrial process. In the case of the fault diagnosis model with mechanism is difficult to established for systems, how to use the accumulated mass data from the online or outline operation process, digging deeply out the data characteristics and the inherent laws in the running on the systems process for the related equipment fault diagnosis, and finally to achieve the purpose of improving product quality and ensuring production safety and reducing economic losses, which has become an issue needed to urgent address in aerospace, chemical, manufacturing, transportation and logistics areas, but also is one of the hot focus in recent years of engineering and academic areas.Therefore, this article is completed by analyzing dynamic non-stationary characteristics from large automated system running and aiming at needs of the diagnostic process. To achieve the purpose of enhancing accuracy and reliability in the based data-driven fault diagnosis technical, this article use of existing measurement techniques and measurement tools for obtaining the mass process data. A deeply researching on data-driven multivariate statistical fault diagnosis carried out data-driven multivariate statistical fault diagnosis by using the appropriate combination of corresponding mathematical theory and engineering practice. The work is good for the further development of the present imperfect data-driven fault diagnosis the theoretical system and achieving data-driven fault diagnosis in the practice of good applications to provide a degree of technical support and the corresponding theoretical guidance. By using the various methods existing in the present data-driven fault diagnosis system and the proposed new ways in this article, the massive data of 6135D diesel engine under the leak fault is sampled and 6 data-driven fault diagnosis methods are respectively exploited in the data samples. Data experimental results show that the various studies carried out and some of the innovative achievements made in this article are significant in both theory and application. This major research work carried out by the following:①Using the wavelet decomposition in various scales to extract detailed features of a noise signal containing the constant bias, graded, mutation and high frequency sinusoidal failure, detailed analyzing the multi-scale features of the dynamic non-stationary data and the fault characteristics distribution law of the data is obtained. Wavelet filtering for the edge effect appears in the result produced by the filtered signal glitches and spikes, with Toeplitz matrices and the Gram-Schmidt orthogonalization algorithm designed the edge of the orthogonal filter, thereby reducing the edge effect on the adverse impact of signal filtering.②The relevant priciples of signal singularity detection are analyzed in this paper. In order to solve the problem that the wavelet-singular point detection is more sensitive to the noise and signal denoising will lose some improtant fault information, a new method of sigular point is proposed based the multiscale products of wavlet which can sharpen the important features of signal while weakening noise. Therefore, multiscale products can distinguish edge structures from noise more effectively.③Aiming at multi-scaling and time-varying of the dynamic non-stationary process data the analysis on multi-scale characteristics of the signal is studied and the edge correcting filter is designed. Using online multi-scale filtering (OLMS) ideology, combined with the recursive principal component analysis (APCA) and multi-scale principal component analysis (MSPCA), multi-scale line filtering based on multivariate statistical fault diagnosis is proposed.④Proposes an online moving window multi-scale principal component analysis(MW-MSPCA) datadrivenbased fault diagnosis method for tracking the non-stationary dynamics of the process which contains time-varyingand multi-scale data. In this data-driven diagnosis technique, wave threshold denoising is used to solve the conflict betweenthe statistical model deviation and data correlation decreasing;the statistical models is updated by using moving windowprincipal component analysis in various scales;the contributions of individual process variables to the process behaviorchanges is illustrated in a 3-dimensional contribution chart, which determines the root cause failure ;and a quantitative evaluation mechanism is also given to evaluating the accuracy in such algorithms.⑤Select 6135D diesel engine as an object of study, to analyze the failure mechanism and the signal features at cylinder head vibration, making the measurement data under the engine failure in the leak as an experimental data sample, implementing applied research by employing respectively 6 kinds of algorithms, such as the traditional principal component analysis(PCA),based on-line multi-scale filtering multi-scale Principal Component analysis(OLMS-R-MSPCA), and the experimental results show that the proposed method is effective and feasible.