Gear Fault Feature Extraction And Pattern Recognition Technology

Gear is the important part that is extensively used in the mechanical equipment, its damage and failure often cause trouble of transmission system or complete equipment, thus lead the great incident. The gearbox, the core of power transmission, its condition monitoring and fault diagnosis has received more and more attention. And it was indicated that sixty percent of gearbox fault came from gear failure. Based on this, this thesis studies detailed the technique of feature extraction and pattern recognition of gear fault. The main contents are listed as follows:First, the mechanism of gearboxes vibration noise is studied in the thesis. To strictly describe its vibration, a mathematics model is established and the key factors of affecting gears' vibration, which are gear teeth stiffness and driving errors, are detailedly analyzed. Based on the research, this thesis analyzes spectrum characteristics of gear long and short period fault.Secondly, it is systematically studied character extraction methods of gears' signal in this dissertation. It is used various methods such as time-domain analysis, frequency-domain analysis and demodulation analysis to comprehensively analyze gear fault signal, and extract features of representative long and short period fault.Thirdly, this paper makes some discusses on BP algorithm principle and its limitations, then modifies it. It is used modified BP algorithm to train ANN, and analyze topology of ANN and ways how to select its train parameters. Through characteristic analysis of gear long and short period fault, it is constructed a BP neural network to diagnose the fault from the extracting fault characters in this thesis. It is practicable that BP neural network is applied to diagnosis of gear long and short period fault through the network analysis results.Finally, aiming at the problem that it was difficult to obtain gear fault sample to train ANN, causing shortage of ANN train and inaccurate diagnosis result, it is included support vector machine(SVM) that is based on statistical learning theory(SLT) to apply in gear long and short period fault diagnosis. And compared with the training result of BP neural network in the small-sample situation, SVM has stronger classification ability and generalization performance than BP neural network, and it can conform to the diagnosisrequirements of practical engineering application.