FuzzyCS-Proto: fuzzy client selection with class-wise prototype learning for federated ship underwater structural surface condition classification

Underwater structural inspection is essential for ship maintenance and marine infrastructure management, but practical inspection data are often distributed across different vessels, ports, inspection platforms and sensing devices. Direct centralized training may be limited by data ownership, communication cost, and operational privacy, while standard federated learning is vulnerable to non-independent and identically distributed (non-IID) data distributions, underwater image degradation and severe class imbalance. To address these challenges, this paper proposes FuzzyCS-Proto, a quality-aware fuzzy client selection framework with class-wise prototype learning for federated underwater structural surface condition classification. The proposed method evaluates client reliability using local validation performance, prediction uncertainty, image quality, and class balance, and selects more informative clients for server aggregation. In addition, class-wise prototypes are extracted from local feature embeddings and aggregated with class-specific reliability to provide semantic anchors for cross-client representation alignment. A minority-class coverage constraint is further introduced to reduce the risk of excluding rare but safety-critical categories. Experiments are conducted on an image-level classification dataset derived from LIACi by cropping annotated surface-condition regions into class-specific image patches and generating normal samples from non-defective regions. The resulting dataset contains five classes: corrosion, defect, marine growth, normal surface, and paint peel. Four federated settings, including IID, Label-skew Non-IID, Quality-skew, and Label+Quality-skew, are constructed to evaluate different types of client heterogeneity. Experimental results show that FuzzyCS-Proto improves Macro-F1 over several typical baselines such as FedAvg and FedProx, achieving Macro-F1 scores of 0.6501, 0.6292, 0.6641, and 0.5074 under the four settings, respectively. Ablation, sensitivity, and Grad-CAM analyses further verify the effectiveness and interpretability of reliability-aware client selection and class-wise prototype guidance.