Adaptive Wasserstein-maximum mean discrepancy generative adversarial network-based fault diagnosis of aviation fuel pumps under sample imbalance

Fault diagnosis of aviation fuel pumps is essential for flight safety; yet, practical deployment is often hindered by severe class imbalance, where rare fault samples are scarce and difficult to collect. To address this issue, this paper proposes an adaptive Wasserstein and maximum mean discrepancy enhanced generative adversarial network for imbalanced fault diagnosis using acoustic signals. Specifically, Wasserstein distance and maximum mean discrepancy are integrated into a unified objective to improve distribution alignment between real and generated data, thereby enhancing the fidelity and diversity of synthesized fault samples and mitigating common generative adversarial network issues such as mode collapse and unstable training. In addition, an adaptive optimization scheme incorporating learning-rate adjustment, reverse gradient flow regulation, and attention mechanisms is designed to stabilize adversarial learning and improve minority-class recognition. Experiments on an aviation fuel pump acoustic dataset demonstrate that the proposed network generates samples with the highest similarity to real fault signals and consistently improves diagnostic performance across a series of classifiers under imbalanced training conditions.