DEEP RESIDUAL NETWORK WITH HYBRID DILATED CONVOLUTION FOR GEARBOX FAULT DIAGNOSIS

Abstract

Commonly used methods for gearbox fault diagnosis involve feature extraction from measured signals to capture its state variation, followed by a fault identification process. These methods are regarded as feature-based process and the extracted features, such as RMS value and kurtosis, are used as input for fault diagnosis. However, fault-related transient impulses, which are embedded in the signals, are lost in feature extraction, leading to reduced diagnosis accuracy. To overcome this shortcoming, the deep residual network with hybrid dilated convolution (ResNet-HDC) is constructed for gearbox fault diagnosis in this paper, which possesses two advantages: 1) deep residual network for deep feature extraction, and 2) hybrid dilated convolution for blurred signal handling. Experimental study performed on a gearbox test rig has shown that the ResNet-HDC is effective for gearbox fault diagnosis.