Pattern Recognition Of Abnormal Sound Engines Based On BP Neural Network

For the diagnosis in abnormal sound engines, the traditional methods of the pattern recognition usually based on experience and knowledge of experts and operators. There are many defects in the traditional methods, which contains difficulties in knowledge acquisition, inference inefficient and low adaptive ability. Meanwhile, it is difficult to acquire mathematics model of the diagnostic system, because the complex nonlinear relationship existed between characteristics and types of abnormal sounds. The artificial neural network provides a new solution to this problem due to its unique advantages, such as parallel distributed processing, self-adaptation, self-learning, associational memory and so on.Based on a type of JS motorcycle and using energy features of acoustic signals as the inputs of network, which was extracted by continuous wavelet, a BP neural network is established in this paper using standard BP algorithm in order to achieve the purpose of the classification of several types of abnormal sounds. However, the BP neural network using standard BP algorithm has some defects, for example, it's sensitive to the order of the sample data, slow convergence, and easy to get into the local minimum. Improvements are introduced in this paper, they are disrupting the order of sample input, adding additional momentum, adaptive learning rate and improved BP algorithm based on Levenberg-Marquardt. The corresponding algorithms of these improvements are the adaptive learning rate BP algorithm, the BP algorithm combined additional momentum with adaptive learning rate and improved algorithms based on Levenberg-Marquardt. This paper makes a detailed analysis about those algorithms which is mentioned above, and the effect of different parameter in the improved algorithms.Finally, comparing the speed of convergence and the accuracy of diagnosis, this paper gets the optimization algorithm and network. The optimization network is three layers network structure: input layer has eleven neurons, the hidden layer has twenty neurons and output layer has two neurons. The optimization algorithm is the improved algorithm based on Levenberg-Marquardt. In the end, the front panel of the diagnosis system is designed by MATLAB graphical user interface (GUI).