Augmenting Beyond Generation: A Statistical Feature Methodology to Imbalanced AUV Fault Diagnosis

Autonomous underwater vehicles (AUVs) operate with enclosed platforms and strict safety constraints, making labeled fault data scarce and imbalanced, which limits intelligent diagnostic models. Mainstream generative augmentation-effective for dense, high-frequency vibration signals-often fails on AUV telemetry, where low-sampling-rate, sparse state data dominate and yield unrealistic syntheses. This study proposes a practical framework, the Statistical Feature Enhanced Data Augmentation Network (SFE-DAN), which integrates statistics-guided augmentation with multi-scale representation learning to address class imbalance under sparse sensing. SFE-DAN couples a statistics path-descriptive features enhanced by a Feature Attention Net (FAN) for data-aware reweighting and feature-space perturbation-with a raw-state path using a Multi-Scale Attention Block (MSAB) to capture multi-receptive-field temporal cues; a Gated Linear Unit (GLU) fusion aligns both paths for robust decision-making. Experiments on a practical AUV platform show consistent gains over six representative methods. Ablation studies further confirm each component's contribution. The framework offers a data-efficient paradigm for AUV fault diagnosis under unavoidable imbalance and low-rate sensing.