One-dimensional fully decoupled networks for fault diagnosis of planetary gearboxes

Abstract Planetary gearboxes are of great importance in extensive industrial fields, and various deep learning methods have been applied for fault diagnosis of planetary gearboxes. But those methods may incorrectly classify due to coupling of the norm and the angle of features, which are corresponding to intra-class variation and semantic difference respectively. In this paper, an improved decoupled network, called one-dimensional fully decoupled network (1D-FDCNet), was proposed to diagnose faults of planetary gearboxes by decoupling inner-class variance and semantic difference with decoupled operators. First, one-dimensional vibration signals of the planetary gearbox were converted into frequency spectrums by fast Fourier transform. And then, the one-dimensional decoupled network was constructed to extract features automatically from frequency spectrums. Finally, decoupled operators continued to be used in fully connected layers for strengthening semantic discrimination ability of the classifier. Due to the application of the decoupled operators, the classification performance of the proposed method was superior by comparing with the backpropagation neural network, the one-dimensional convolutional neural network and the normal decoupled network according to the trail results. The results of the experiment and the engineering application confirmed that the proposed method was more effective than other methods.