New Fault Diagnosis Methods Of Kernel Principal Components Analysis And Evidence Theory On Multi-Domain Feature

With the development of modern hydraulic system for the direction of high-speed, high power and high reliability, the function of electromechanical equipment hydraulic system gets more complicated than before, the structure of it gets bigger and at the same time lots of uncertain factor and information appears. To reduce the fault occurring and improve the productivity, it is very important to build a perfect system of equipment hydraulic system's maintenance management. Therefore, the early phase of the hydraulic system's fault diagnosis and detecting becomes the core problem in the predictive maintenance of the equipment management system. In recent years, the theory and technology of intelligent fault diagnosis has made a big progress which now can help people identify the pattern precisely and predict the fault in given conditions. But there are still some difficulties to be solved in this field, such as the limitation of the sensor, the time-variation of fault diagnostic object, the Incompleteness based on the one-source signal, the compound of equipment fault. They are all the factors preventing the development of the theory and technology of intelligent fault diagnosis.Kernel principal component analysis (KPCA) and D-S evidence theory are analyzed in this paper. The denoising envelop demodulation method is used to process the signal based on wavelet packet, and KPCA fault diagnosis based on noise signal and fault diagnosis method based on exponentially weighted dynamic KPCA are presented here. Methods are studies how to settle basic probability assignment, multi-source information fusion fault diagnosis method which combined Support Vector Machine (SVM) and evidence theory is proposed in this paper. This method makes full use of complementary and redundant information of all signal sources can improve the precise and reduce the rate of misdeclaration and missing report of fault diagnosis remarkably. Diversified compound fault patterns are set in experiment which can be seen as a special fault pattern and then the pattern was diagnosed directly by multi-source information fault diagnosis method which combined Support Vector Machine (SVM) and evidence theory.The main works in this dissertation are showed as follows: (1)To ensure the completeness of feature fault information, 16 feature parameters from the time, frequency and time-frequency domain are selected to constitute multi-domain feature vector in fault diagnosis of the hydraulic pump. On the basis of analyzing the basic theory of wavelet packets and envelope demodulation algorithm based on Hilbert transform, the pretreatment method of envelop demodulation based on wavelet packets band-pass filtering denoising is put forward. Then the method is used to process the gathered vibration and sound signal. The energy of each frequency band contains much fault information. Method of extracting band energy based on wavelet packet decomposition is given and 8 feature parameters are extracted in time-frequency domain. 5 dimensionless parameters in time domain and 3 feature parameters in frequency domain under 4 common faults are presented, it also analyses the degree of the parameters'sensitivity to these 4 faults.(2)The fault diagnosis method of KPCA based on sound signal of pump is expatiated. The basic theory of KPCA and its basic procedure when applied in the fault diagnosis is introduced. The detailed pre-processing procedure of the sound signal is presented. Feature vector in multi-information domains is extracted from time domain, frequency domain and time-frequency domain. At last fault diagnosis is carried out by KPCA. The diagnosis result is compared with the single feature vector of parameters in time domain, frequency and the time-frequency respectively, and then compared with diagnosis result based on vibration signal.(3)Fault diagnosis method based on exponentially weighted dynamic KPCA is proposed. Due to the dynamic and time varying state of the normal hydraulic pump, it adopts the refresh method of date by time sliding window, sets up new KPCA mode with the new date. It introduces exponential weighting coefficient, constructs a diagnosis model featured dynamic self-adaptability by using both the new and the old model. It then elaborates in detail its procedure of modeling and diagnosis, makes the diagnosis based on vibration signal, discusses the exponential weighting coefficient's influence on the diagnosis, and then makes a contrast with the traditional KPCA diagnosis result.(4)Multi-source information fault diagnosis method based on combining Support Vector Machine (SVM) and D-S evidence theory is proposed. The basic probability assignment is settled by using SVM, the basic step of fault diagnosis of D-S evidence theory is given. 5 channel Signals are monitored, which include 3 channel vibration signals in the x-axis, y-axis, z-axis, sound signal and outlet pressure signal. Recognition framework of fault diagnosis and the acquisition parameters are set, feature vector is extracted in multi-domain to complete the fault diagnosis of hydraulic pump, after signals are pre-processed. The result is compared in detail with the basic probability assignment acquired by the method of BP-neural network, and fault-tolerant capability of the new method when losing one certain channel is studied.(5)Compound faults of the hydraulic pump are diagnosed directly. The experiment puts the pump in the state of compound faults, and compound fault is seen as a special pattern of fault. Then diagnosis can be executed by using both SVM and D-S evidence theory fault diagnosis method.(6)The scheme is proposed of experimental system based on virtual instrument, system of 5 channels'axial piston pump operation monitoring is established. The system chooses 3 orthogonal components, that is vibration signal, sound signal and outlet pressure signal as its monitor signal. It analyses the vibration state's frequency range and the fault analysis frequency band of every pump and selects suitable sensors. It studies where to place the stations of the vibration sensor and sound sensor. Many faults and collecting parameters are given, the foundations of fault diagnosis methods are proposed.