Hydrophone Spectrum-Sequence-Driven FedNova-MobileViT Model for Ship Propeller Cavitation Recognition

Source abstract

Propeller cavitation produces non-stationary acoustic signatures that vary with vessel operating condition, sensor position, and background noise, while operational constraints often prevent raw hydrophone recordings from being pooled across platforms. This study proposes a FedNova-MobileViT framework that learns four cavitation states from distributed hydrophone spectrum sequences without routine raw-audio transfer. Continuous recordings are converted into normalized log-power spectrum sequences with 128 frequency bins and 64 consecutive frames. A shallow depthwise-convolutional stem preserves local ridges and broadband bursts, MobileViT blocks exchange information across separated time- frequency neighborhoods, and FedNova normalizes client updates to reduce bias from unequal local optimization progress. Evaluation uses five heterogeneous acoustic clients representing towing-tank, harbor-approach, open-water, maneuvering, and maintenance conditions. The proposed model achieves 96.7% accuracy, 95.9% macro-F1, a false-alarm rate of 0.041, a missed-alarm rate of 0.047, and 38 ms sequence-level inference latency under the reported embedded GPU profile. Its macro-F1 exceeds FedAvg-MobileViT and FedProx-MobileViT by 2.5 and 1.7 percentage points, respectively. FedNova reaches 95% of its final macro-F1 after 42 communication rounds, compared with 61 rounds for FedAvg under the same sampling policy. The evaluation also examines class-wise recall, calibration, noise robustness, component ablation, communication payload, and energy consumption to characterize recognition quality and deployment cost. These results indicate that temporal spectrum modeling and normalized aggregation are complementary for cavitation recognition under heterogeneous client data, although broader vessel-level validation and repeated-seed uncertainty analysis remain necessary.