A Research On Overload Fault Diagnosis For Vacuum Pump Based On Acoustic Emission Signal

With the development of space science and technology and the development of manned spaceflight engineering,especially the progress the space propulsion technology has achieved during recent years,greater demands of safety and reliability are being placed on the experiment equipment by the spacecraft simulation of spacecraft.And the normal operation of the vacuum pump is the foundation of a successful experiment.So,it's of great importance to detect the operating status of the vacuum pump.This paper has mainly studied an on-line monitoring and diagnosis system for vacuum pump,which is based on the principle of acoustic emission signal.According to the principle of the acquisition of acoustic emission signal,a signal acquisition system of the vacuum pump is built.The system can be divided into two parts: hardware and software.Hardware is mainly composed of acoustic emission sensors,preamplifiers and digital acquisition card.Software is mainly composed of two parts: a program for the acquisition of signal which is written by LabVIEW,and a program for data processing which is written by Matlab.After the system was verified in the laboratory,the normal signals and the overload signals were collected at the test site.The acoustic emission signals we achieved are mixed with a lot of noise,firstly,we use a denoising method which is based on the singular value decomposition to denoise the signal,and a new method for the selection of the separation order is proposed in this paper,which combines the contribution rate of the singular values and the slope change of the singular values.And then the effectiveness of the denoising method has been verified by the simulation of the experiment signals.Secondly,in this paper,three different methods are used to extract the feature vector of the signals:(1)Six eigenvalues,such as distortion,kurtosis index,Peak,and so on,are extracted from the signal samples to form the feature vector of the time domain.(2)The signal samples are decomposed for seven layers by using the wavelet packet,and the relative energy spectrum of the first eight bands,which are obtained by the signal decomposition,is extracted as the feature vector of the frequency domain of the signals.(3)The signal samples are decomposed by empirical mode decomposition,and the relative energy spectrum of the first nine intrinsic mode function,which are obtained by the signal decomposition,is extracted as the eigenvector of the frequency domain of the signals.The result shows that,the recognition of feature vector extracted by the wavelet packet decomposition is the highest,which is up to 96.7%,and the time it spends is short also.In this paper,two Pattern recognition methods are used to recognize the signals: the support vector machine and the extreme learning machine.And the two methods are compared from two aspects: the accuracy of the recognition and the running time of the programs.The results show that both the support vector machine and the extreme learning machine can distinguish the normal signals from the overload signals,and the accuracy of the recognition of the support vector machine is higher than that of the extreme learning machine,while the running time of the program of the extreme learning machine is shorter.Considered that,the running time of both can meet our requirements,the acoustic emission is a better choice in this study.