A Discriminative Feature-Based Fault Diagnosis Network for Planetary Gearboxes Under Variable Working Conditions

Abstract

In planetary gearbox fault diagnosis under variable working conditions, the method based on unsupervised domain adaptive is to correct the data shift between different working conditions. However, the current methods only focus on extracting invariant features, ignoring the extraction of discriminant features, resulting in a large number of samples near the classification boundary and serious class confusion in the knowledge transfer process. Furthermore, the current methods only use the measure of feature discrepancy to align the marginal feature distribution and ignore the fine-grained information of samples with the same label but in different working conditions. In fact, it is sometimes ineffective to rely solely on the measure of feature distance for extracting domain-invariant features. Given these shortcomings, we propose subdomain confusion adaptive networks based on discriminative learning (SCAD). First, a new Softmax variant, C-Softmax, is proposed, which adjusts the features to enhance their discriminatory power. Second, a new domain adaptive strategy is proposed to measure and reduce the confusion of the target domain based on aligning the marginal distribution and condition distribution of the data in different working conditions. Compared with other methods, SCAD has higher diagnostic accuracy in eight fault diagnosis tasks under different working conditions of the planetary gearbox.