Application Of Wavelet De-noising And Probabilistic Neural Network To Recognize The Signal Of The Engine Sound

To collect and analysis the sound when the engine is running is an effective way to understand the state of the engine. Therefore, recognizing the signal of the engine sound is an important part to the engine state's detection and classification. For the current status that it uses artificial auscultation usually to judge the engine sound, this paper provided a method which uses the wavelet de-noising and probabilistic neural network to recognize the signal of the engine sound.Firstly, this paper analyzed the concept and nature of the continuous wavelet transform and the discrete wavelet transform which lay the foundation of introducing wavelet de-noising. Through comparing and analyzing the three methods' advantages and disadvantages which combine modulus maxima de-noising, wavelet thresholding de-noising and wavelet space de-noising, it pointed that wavelet thresholding de-noising is a better choice. Because the signal of the engine sound has the mutational and discontinuous characteristics, and wavelet thresholding de-noising will produce pseudo-Gibbs phenomenon, this paper proposed a method that wavelet thresholding de-noising based on the translation invariance method, and determined the size of the translation, and resolved the problem of selecting the threshold function. By analyzing the blocks signal which added different noises, it confirmed that the method of translation invariant wavelet de-noising has a good de-noising effect. This paper analyzed the common theory in recognizing the signal of the engine sound by probabilistic neural network, and described the concept of the artificial neuron, the classification and learning way of the neural network. On the basis of analyzing the Bayes theory and Parzen window theory in depth, this paper summarized the characteristics of the neural network, and described the learning algorithm systematically.This paper selected the sound signal produced by a four-stroke engine with a horizontal bar. And it described the technical parameters of the engine, the types of the sound signal(abnormal sound mainly), and determined the operating conditions, the measuring point and the test environment of the signal of the engine sound, analyzed the hardware and software of measuring the engine sound, and made a norm which can capture the signal of the engine sound correctly.Based on these research results, this paper made a comparative analysis that combines the normal engine, the box with abnormal sound engine, the right cover with abnormal sound engine, the left cover with abnormal sound engine, and whether they are de-noised. This paper confirmed that it was necessary to use wavelet de-noising and it was correct to use the 1/3 octave value as the feature vector of the engine sound signal. By substituting the eigenvectors into probabilistic neural network to train and forecast, the results showed that wavelet de-noising and probabilistic neural network can recognize the signal of the engine sound well.