Intelligent Diagnosis Method of Gear Faults based on MGAN and CNN

Abstract

Aiming at the problems of insufficient samples of gear fault signals collected in practical engineering，insufficient training，low fault recognition rate and easy pattern collapse of generative adversarial networks when using common deep learning network for pattern recognition under noise interference. An intelligent diagnosis method of gear faults based on MGAN （Mixture Generative Adversarial Nets） and CNN （Convolutional Neural Networks） is proposed. The real gear signals are transformed into time-frequency signals by time-frequency transformation technology after morphological filtering，and new samples are synthesized by MGAN to obtain a balanced data set，which overcomes the problem of insufficient samples. At the same time，the influence law of main parameters of MGAN on the quality of synthesized samples is analyzed. The balanced data set is used to train CNN for fault diagnosis，which effectively improves the fault diagnosis rate. Through the comparative test and bench test，the effectiveness and advantages of this method in accurately identifying faults and overcoming the collapse of neural network mode under the condition of insufficient samples are verified，which provides a new research idea for typical fault extraction and intelligent identification of gearboxes.

Keywords

• Generative adversarial network;Model collapse;Insufficient samples;Fault diagnosis