A Semi-Supervised Fault Diagnosis Framework for a Gearbox Based on Generative Adversarial Nets

Abstract

It is very significant to realize effective fault diagnosis of a gearbox in modern industrial systems. Undeniably, the traditional intelligent fault diagnosis methods such as back propagation (BP) neural network, recurrent neural network (RNN), extreme learning machine (ELM), Long Short-Term Memory (LSTM) and convolutional neural network (CNN) might have a satisfactory performance in accuracy. However, the premise of this high accuracy is to add labels to all samples manually, which will undoubtedly increase the cost of failure detection. In this article, a semi-supervised fault diagnosis framework for a gearbox is proposed by utilizing GAN. First of all, fast Fourier transform (FFT) is adopted transform 1-D vibration signals into 2-D frequency spectrograms to fit the input format of GAN. Then, the frequency spectrograms are input into the GAN model to achieve fault diagnosis with few marked samples. Finally, an experiment study is carried out to confirm the excellent result of our approach in accuracy and stability. The results indicate that its performance in stability and accuracy is quite excellent.