Explainable federated learning-based fault diagnosis framework for safety assurance in multi-agent autonomous systems

Purpose This study introduces an explainable federated learning (XFL) framework for fault diagnosis in multi-agent autonomous systems, such as UAVs and self-driving vehicles. It addresses limitations of centralized approaches, including scalability issues, communication delays, data privacy concerns and lack of transparency in safety-critical environments. By enabling collaborative model training without central data pooling, the framework ensures real-time fault detection, interpretability, and regulatory compliance, contributing to safer operations in intelligent mobility, drone logistics, and urban robotics. Design/methodology/approach The framework employs Federated Averaging (FedAvg) for distributed learning across autonomous agents, with local models featuring a lightweight neural network including attention mechanisms and softmax output. SHAP (Shapley Additive Explanations) is integrated for local interpretability, with aggregated global profiles for feature relevance. Simulations mimic multi-agent environments with non-IID data, injecting faults such as sensor drift, dropout and adversarial perturbations. Evaluation uses metrics such as accuracy, convergence speed and explanation fidelity, benchmarked against FedAvg, centralized and local-only methods on edge hardware such as Jetson Nano. Findings The XFL framework achieved 92.6% classification accuracy after 20 rounds, outperforming FedAvg (88.7%) and centralized baselines (86.9%), with faster convergence (12 vs. 18 rounds). It demonstrated robustness, retaining 89.4% accuracy under faults such as Gaussian drift and FGSM perturbations. SHAP explanations showed consistent feature importance (e.g. gyroscope delta and latency spike) with Jaccard indices averaging 0.705 across clients. Real-time inference latency was under 60 ms on edge devices, validating practicality for deployment. Ablation studies confirmed contributions from SHAP reweighting and attention layers. Originality/value This work pioneers the integration of explainable AI (SHAP) into federated learning for fault diagnosis in multi-agent autonomous systems, bridging gaps in privacy-preserving, interpretable and robust safety solutions. Unlike prior studies focused on accuracy alone, it embeds transparency during training, enabling traceable decisions for regulatory compliance and public trust. The framework's entropy-based reweighting enhances convergence and resilience, offering scalable value for safety-critical applications such as UAV swarms and vehicle platoons, advancing intelligent manufacturing and special equipment reliability.