Study On Information Entropy Feature Extraction And Fusion Methods In Fault Diagnosis

Along with the increasing of system complexity, the running state parameters take on the dynamical characteristics of non-stationary and multi-source coupling,so the feature extracting of the signals and fusing diagnosis of multi-source information in fault diagnosis are the research hotspot and nodus all along.In this paper, aiming at the key technology of the fault diagnosis—feature extracting of signals,the study was carried out based on information entropy and data fusion theories in two aspects that are the quantitative description of the characteristics of the multi-level signal and quantitative fusing analyse of multi-source signal, and multiple information entropy indexes for different signal transform space and information entropy fusion models for different fusion level are proposed.Firstly, two main information measurements—information entropy measurement and complexity measurement are discussed on their measurement index, algorithms and application characteristics,and aiming at the limitations of the two information measurements,a new information component probability estimation algorithm is put forward to describe the energy distribution character of the signal, and a complexity information entropy module applied for generalized signal transform space is proposed to realize the quantitative describing of the complexity characteristics of the signal and be as a theoretical foundation of the multi-source and multi-level signal feature extraction.Secondly, based on the proposed complexity information entropy module, the methods on the multi-level feature extraction for singular variable in different signal transform space and the association feature extracting for multi-source parameters are researched separately. One the one hand, the complexity information entropy is combined with different signal analyzing algorithms in time domain, frequency domain and time-frequency domain separately to construct the multiple information entropy feature indexes, such as singular spectrum entropy, power spectrum entropy, multi-resolution analysis complexity entropy and multi-resolution singular spectrum entropy etc., which could realize the quantitative description of the energy distribution character of the signal in different signal transform space. One the other hand, the complexity information entropy is expanded to two-dimensional signal space to construct the information entropyindexes for the relating analysis of multiple variables, which called joint complexity information entropy and information transfer feature index to describe the joint energy distribution character and its variety characteristic. The theoretical methods are proved to be effective by the experimental analysis on typical simulating signal, response parameter of nonlinear system and failure signal of the rotor system.Thirdly,by combination of the information entropy feature extraction and data fusion idea, the information entropy models and realization methods for the fusion analyzing of the multi-source signal are proposed, which include the information entropy index used in data level fusion—double channel energy spectrum entropy and the multiple channel singular spectrum entropy.the multi-source information relating optimization and information transfer feature analyzing model used in feature level fusion,the information entropy rule and optimized wavelet-neural network mapping model used in decision-making level fusion. The proposed information entropy fusion models could realize the fusion analyzing of the multi-source information in different levels and could get the quantitative information or fusion result about the state character of the system.Finally, the application studies on the information entropy feature extraction and fusion method is processed on the fault diagnosis of the rotor system. The value response of the rotor system in typical fault state—transverse crack, rub-impact and transverse crack coupling rub-impact are gained by rotor dynamics analysis, and the nonlinear characteristics of the response in different parameter conditions are quantitative described by the information entropy algorithms to realize the simulating analysis of the information entropy characters of different rotor fault state. Furthermore, the fault diagnosis experiment is processed on experimental rotating machine, and the effective recognition of the running state of the system is realized by the information entropy feature analysis and fusion diagnosis to the detected multi-source signals. So the proposed information entropy model is proved to be effective for the coupling feature extraction of the multiple signals and fault diagnosis of the complex system.