Incipient Fault Prognosis Of Gearbox Based On Semi-supervised Self-organization Map Neural Network

The incipient fault signals are easy to be contaminate by noise as the signals are very weak, therefore, the fault information are difficult to be extracted by traditional fault diagnosis methods. And it is difficult for traditional methods to diagnose more than one category of faults. Consequently, it has practical significance to research incipient fault diagnosis methods.Self-organizing Map is a very useful tool for clustering and visualization. However, the directly using of the GNSOM and DPSOM often leads long training time, low accuracy and bad visualization quality. Linear discrimination analysis (LDA) solve this problem very well, it can reduce the dimension of the initial faults feature dataset and eliminate the redundant influence of features. So the training speed and the accuracy of LDA-GNSOM and LDA-DPSOM are improved comparing with the classical SOM. The Iris dataset simulation indicates the effectiveness of these two methods.To further expedite the training speed and improve the visualization quality, the semi-supervised LDA-GNSOM and semi-supervised LDA-DPSOM are also proposed based on their unsupervised learning methods. It can learn unlabel samples according the guide of label samples, so training time is reduced. The bear and gear experimental results indicate the effectiveness of the methods for incipient fault diagnosis.Experiments of gear faults are conducted on the transmission testing platform. The fault signals are analyzed by time domain and frequency domain statistical analysis, which demonstrate the complexity of incipient fault signals. The proposed semi-supervised LDA-GNSOM and semi-supervised LDA-DPSOM methods are both applied to detect and classify the incipient faults. Semi-supervised learning performs better performance comparing with unsupervised learning, and experimental results show the effectiveness for gearbox incipient fault diagnosis.