EDANN-Based Cross-Domain Diagnosis Study of Gearbox Faults

Aiming at the problem that the traditional intelligent diagnosis model for gearbox faults cannot apply the previous knowledge gained under different operating conditions, a transfer learning model with enhanced domain-adversarial neural networks (EDANN) is proposed. Firstly, EDANN uses atrous spatial pyramid pooling (ASPP) to expand the feature extraction range and capture richer signal features. Secondly, the maximum mean square discrepancy (MMSD) layer assigns different weights to the source and target domain features mapped into the same space, in order to achieve the purpose of reducing the feature discrepancy and domain adaptation. Thereafter, the domain adversarial loss value generated during the training of the model is employed to impede the ability of the domain discriminator to distinguish between the feature representations of the source and target domains, which in turn compels the feature extractor to learn invariant features that are domain-specific. Finally, the domain-invariant features are classified by the fault classifier to complete the diagnosis of the gearbox health under the variable operating conditions. Simulation results show that the cross-domain diagnosis accuracy of the improved network is better than that of the original network, deep domain confusion network (DDC) and efficient network (EfficientNet), which basically meets the industrial requirements.