Research On Bearing Fault Diagnosis Technique Based On Neural Network

Rolling bearing is one of the most important parts in mechanical equipment, especially the rolling equipment. Mechanical equipment is widely used in our society, such as aeronautics and astronautics, transportations, economical productions, metallurgy and chemical industry, Rolling bearing usually work in the critical environment and always prone to failure. Once the bearing failure occurs, it will seriously impact on economic profits, and even do harm to human’s safety. How to keep the rolling bearing’s health, or how to detect the failure immediately and change new bearing in time is of great importance to our daily life. So, it is very meaningful to study the technology of fault diagnostic of the rolling bearing.The vibration signal data of rolling bearing which is operating in four types of conditions, the normal condition, inner race fault condition, rolling element fault condition and outer race fault condition, is the search object of this article. Feature parameters are extracted by the means of wavelet transform and Hilbert-Huang transform, and then input to the neural network. After the weights and thresholds are optimized, the performance of neural network is also optimized. Then this neural work is used in the pattern recognition in diagnostic process.The rolling bearing fault diagnosis technology research and development history is shortly described in this paper. Then the structure of rolling bearing, the types of bearing fault data, and the fault characteristic frequencies are introduced. Time domain analysis and frequency domain analysis are used as the simple methods to analyze the data. Considering that the fault data of rolling bearing has the characteristic of non-stationary, the wavelet analysis, Hilbert-Huang transform based on adaptive parameters are used to extract the fault features of the four types of fault data. These all three methods have showed the good effects and the parameters extracted by these three methods are combined as a fault feature vector which can respective the characteristics of the rolling bearing fault types. Through the comparison of the application effects of neural network, Levenberg-Marquardt BP neural network is chosen in pattern recognition. Euclidean distance based feature selection method is used so that the most representative parameters could be chosen to shrink the dimensions of feature vectors. After the weights and thresholds are optimized by genetic algorithm for neural network, neural network’s training speed and precision are largely promoted. Analysis result of the test samples shows that this method has a high accuracy in classification of fault modes and that it has strong suitability for rolling bearing fault detection and diagnosis.