Research On Fault Diagnosis Of TE Process Based On Data Driven

Data-driven controls and applied as a hot research at home and abroad in recent years, is widely used in various fields, based on data-driven important stage fault detection, diagnosis, recognition in academic research. Data-driven approach to the use of controlled system online and offline data analysis for process data analysis and processing, and achieved good results. In the process industries, traditional modeling methods through various small-scale simultaneous linear relationship between input-output subsystem, to establish a more precise mathematical models, such as principal component analysis (PCA), control charts and Fisher discrimination method (FDA) etc. and so on. However, the actual industrial process basic data showing non-normal distribution, principal component analysis (PCA) can handle easily lead normal limitations and failure analysis allowed false positives, false negatives. And in the face of multi-scale, multi-level complex system, the majority of small fault data showing non-Gaussian and nonlinear, and such methods can only handle linear data, these limitations become stage process industry monitoring glitch problem.With large-scale industrial processes, complex development, analysis, nonlinear correlation between the variables of the system, such as a key mathematical model. In view of this, this paper using independent component analysis (ICA), nucleus independent component analysis (KICA) and independent subspace - Partial Least Squares (ICSM-PLS) three methods of historical data for research, by analyzing the acquisition system features changes in the variables of the system running status signal analysis, treatment of industrial process to achieve accurate Africa Gaussian, nonlinear fault data, and minor failures to achieve accurate monitoring purposes. The main work is as follows:1) For principal component analysis (PCA) can only deal with the limitations of Gaussian signal data, we use ICA method instead of the traditional PCA, through the ICA correlation between the original data to eliminate, to achieve the non-Gaussian nature of the fault signal analysis and recognition.2) With the increase of industrial processes observed sample signal matrix nonlinear complexity, for the ICA should not deal with the limitations of the more complex nonlinear observer data, using KICA approach by maximizing the extraction of negative entropy converts non-Gaussian information to minimize negative entropy, enhanced nonlinear processing capability of the system to overcome the ICA method for analyzing nonlinear systems poor performance shortcomings.3) Fault diagnosis in industrial processes, the traditional ICA algorithm on the entire variable spatial modeling is easy to overlook minor exceptions, resulting in a lower detection rate of traditional fault diagnosis methods. To solve this problem, this paper ICSM-PLS method, more flexible integration strategy for signal analysis, subspace algorithms to detect small faults in the system. ICSM-PLS method based on independent component contribution to strengthen the local variables of the system, signal processing and analysis, improved fault diagnosis of industrial processes.In this paper, the simulation of industrial production in the actual situation on the TE process, industrial process to extract data in a separate component of the test results were analyzed by. Experiments show that the method is effective to detect the operating state of industrial processes, to achieve the purpose of the process of industrial process monitoring.