DFL-GC: A Decentralized Federated Framework for Planetary Gearbox Fault Diagnosis Under Domain Shift and Class Imbalance

In planetary gearbox fault diagnosis, Federated Learning (FL) enables clients to train models while preserving data privacy collaboratively. However, achieving satisfactory performance typically assumes that data are independent and identically distributed across clients, and that class distributions within each client are balanced. In practice, these assumptions rarely hold: data collected at different clients under distinct working conditions induce domain shift, and within each client, predominantly normal operation leaves few fault samples, leading to class imbalance. Domain shift can destabilize model aggregation, and class imbalance biases learning toward the majority class. Therefore, addressing the coexisting domain shift and class imbalance is essential for the practical application of FL. To support serverless collaboration under parameter-only communication, a Decentralized Federated Learning (DFL) framework with a Gaussianized Feature Alignment Module (GFAM) and a Class-Balanced Kullback–Leibler loss (CBKL), named DFL-GC, is proposed and aggregated via a ring topology without a central server. During local training, each client applies GFAM to map local features to a standard Gaussian distribution and refines representations with a self-attention mechanism, reducing inter-client distribution discrepancy and improving diagnostic performance; CBKL applies client-specific class-frequency-aware weighting to Kullback–Leibler divergence terms to strengthen intra-class prediction consistency and interclass prediction separation for minority classes, improving minority-class diagnosis. Experiments on two planetary gearbox datasets show improved domain alignment and robustness to class imbalance, yielding superior diagnostic performance and outperforming seven comparative methods.