Study On The Extraction Of Engine's Fault Sound Signals Based On The Neural Network Prediction Model

Using sound information to diagnose fault is the method that had been continuing to use in the past. But it hadn't got fast development because it has not combine with the modern signal processing technology very well. But if the processing results of acoustic signal are used to stimulate characteristic of human audio system, then the performance may be even better and it is much easier to be accepted by operators. So fault diagnosis based on acoustic signals is very important in real world.In the fieldwork noise detection, because of complication of the real acoustic environment, we often face with complex spatial acoustic field that covers large area, or is unsteady or has multi-sound source. Especially when the machine to be inspected is out of order, the fault signal, the steady normal working signal and background noise mix together. The signal detected in this condition is actually a mixture of multi sources and their reflections. When certain machine is out of order, the sound that represents characteristic of the fault may be canceled or covered totally by signals form other sources and accuracy and reliability of fault diagnosis are degraded drastically. This greatly limits the usage of audio fault diagnosis in real world. It is obvious that how to acquire real acoustic signal of the machines to be inspected is key to improve feasibility of audio sound diagnosis. Because of lacking in effective method to separate signals, nowadays there is no much progress in the study and application of noise fault diagnosis. So there needs an algorithm urgently which can separate the fault sound-source signal to improve the realization of audio fault diagnosis.Most of engine's fault sound signals are due to the part abrasion, crack, flexible of touching surface or parts of engine, so the engine's fault sound signals are the acoustic signals by strike. Because these striking course were finished shortly, engine fault signals are stochastic vibrating attenuating impulse signals, which often reflect significant engine's fault information. Moreover the normal acoustic signals of engine are multi-source, complexity, and have much noise, we can just extract the fault signals under the high signal-to-noise ratio using the traditional signal processing methods, without efficiently extracting the fault signals under the low signal-to-noise ratio.Because neural network has abilities of self-learning, self-organizing, associative memory and parallel processing, it is widely used. Today, there are many applied cases about neural network in automatic control, information processing, robot, medical diagnosis, pattern identification and fault diagnosis. So we can adopt neural network method to study project in this paper.As we know, the engine working sound signals is determined by common movement of the complex and regular parts, therefore the engine working sound signals is characterized by random, internal discipline determinism, which is complied with the chaos signals. So they are closely contacted.In this paper, we study the chaotic theory, the main character of the chaos signal and the theory of phase-space reconstruction in chaos system. Use GP algorithm to calculate correlation dimension, then calculate embeding dimension. Use autocorrelation algorithm to calculate delay time. By calculating the most Lyapunov exponent, we proved that engine working sound signals is chaotic. We gave the relevant simulation results, which make a good foundation for building engine neural network prediction model and fault signal extraction based on this model in next chapterIn this paper, the basis knowledge of neutral network and main applications are introduced. And we study Radial Basis Function neural network, which has many good features and is widely used, including its basis framework, algorithm, features and improving feature methods. We combine with the theory of phase-space reconstruction in chaos system and Radial Basis Function neutral network to establish RBF neural network prediction model, which based on time series from Duffing equation or engine working sound time series. Based on this model, we extract the fault sound signals. So the fault sound signals are enhanced. We do experiments according to several different data. And we get the satisfied results. We give the relevant simulation results.It contains some noise in the extracted fault sound signal. In order to identify engine's fault signal, the noise reduction methods of signal using wavelet transform is used to make signal stronger. Wavelet transform inherited and developed localization idea of Gabor transform. At the same time, it overcomes the shortage of Gabor transform such as the fact that the window can not change with frequency and the lack of orthogonal bases. Wavelet is not only an ideal tool to analyze local frequency spectrum, it's also have good local characteristic in time domain. It provides a powerful analytical method for acoustic signal analysis.In this paper, we realize the basic theory of wavelet mainly, introduces the basic idea of series wavelet, disperse wavelet, sdudy the noise reduction disposal of signal using disperse wavelet, than uses this method to the disposal of the engine's fault sound signals. This course effectively realizes noise reduction of the engine's fault sound signals. This paper researches the noise reduction methods of signals using wavelet, chooses method based on threshold value and suitable threshold value to reduce noise. Through simulation examples, performance of noise reduction is very well and the precision of fault diagnosis is enhanced.