现有的联邦学习（federated learning，FL）广泛应用于工业物联网，特别是在工业安全风险识别与异常检测方面，可实现分布式模型训练。然而，FL在面向工业安全风险的训练任务中仍面临高通信和计算成本、节点不可靠等挑战。为了解决上述问题，该文提出一种动态分簇联邦学习算法（dynamic clustering framework for federated learning，DCF-FL）进行动态联盟形成与资源分配。首先，基于主观逻辑提出实时的综合信誉评估机制，筛选并动态选取可信任的摄像头节点作为簇头，以提升训练鲁棒性。其次，构建了一个混合整数非线性规划问题，联合考虑设备的计算能力与无线信道条件，选择参与FL的节点并优化资源分配以最小化通信与计算开销，同时保证模型精度；通过深度强化学习驱动的解耦方法求解该问题，将选择决策与资源分配子问题分离求解。仿真实验结果表明，在15个节点的典型场景下，DCF-FL的通信与计算综合成本比FedAvg算法和FedProx算法分别降低了76.9%和61.3%；在非独立同分布数据环境下，DCF-FL的识别准确率基本在20%~80%之间波动，远高于对比算法（低于20%）。

Objective: Industrial Internet of Things (IIoT) has emerged as a critical infrastructure for safety monitoring in high-risk industrial environments such as petrochemical plants and power systems. Real-time identification of safety risks, including abnormal behaviors and hazardous events, is essential for ensuring operational reliability and preventing accidents. Federated learning (FL), a distributed machine learning paradigm, enables collaborative model training across multiple edge devices while preserving data privacy. However, its application in IIoT-based safety risk identification still faces considerable challenges, including unreliable edge nodes, heterogeneous computing and communication capabilities, and high communication overhead due to frequent parameter exchanges. These issues lead to degraded model accuracy, reduced system robustness, and increased latency. To address these limitations, this paper proposes a dynamic clustering framework for FL (DCF-FL) to improve the reliability and efficiency of safety risk identification in IIoT environments. Methods: The proposed DCF-FL framework integrates three key components: a reputation evaluation mechanism, dynamic cluster head selection, and joint optimization of device selection and resource allocation. First, a comprehensive reputation evaluation mechanism based on subjective logic is designed to assess the trustworthiness of edge devices. This mechanism incorporates multiple factors—such as data contribution, local model accuracy, and channel conditions—to dynamically select reliable devices as cluster heads, thereby enhancing training robustness. Second, the FL process is organized into dynamically formed clusters, where the cluster head coordinates local training and model aggregation. A dynamic replacement strategy for cluster heads is introduced to mitigate single-point failures and adapt to varying network conditions. Third, the joint optimization problem of device participation and resource allocation is formulated as a mixed-integer nonlinear programming problem, which considers communication and computation costs while maintaining model accuracy. To efficiently solve this complex problem, a deep reinforcement learning (DRL)-based decoupling approach is adopted, where device selection and resource allocation are handled in separate stages. The DRL agent iteratively learns optimal decision policies through interaction with the environment, improving resource utilization and reducing the overall system cost. Results: Extensive simulations are conducted using public datasets, including BoWFire and UCF-Crime, along with a custom dataset for industrial safety scenarios. The experimental results demonstrate that DCF-FL substantially outperforms the FedAvg and FedProx baseline methods. In a typical scenario with 15 devices, DCF-FL's total communication and computation cost is reduced by 76.9% and 61.3%, respectively, compared with the baseline algorithms. Moreover, the DRL-based optimization achieves near-optimal performance, with only a 4.6% gap compared with the theoretical optimum obtained by traditional branch-and-cut methods, while maintaining much lower computational complexity. In terms of model performance, DCF-FL exhibits faster convergence and lower test loss, especially in non-independent and identically distributed data settings. The accuracy of risk identification remains consistently higher than that of the baseline methods across most training rounds, demonstrating the effectiveness of dynamic clustering and reputation-based node selection. Additionally, the proposed reputation mechanism effectively distinguishes unreliable devices, preventing their negative impact on model training and improving system stability. Conclusions: This paper presents DCF-FL, a novel FL framework tailored for industrial safety risk identification in IIoT environments. By integrating subjective logic-based reputation evaluation, adaptive cluster head selection, and DRL-driven resource optimization, DCF-FL effectively addresses the challenges of node unreliability and high resource consumption in traditional FL systems. The results indicate that DCF-FL achieves a favorable balance between model accuracy and system efficiency, making it suitable for resource-constrained and heterogeneous industrial scenarios. Future work will focus on further improving the stability of model accuracy under highly dynamic conditions and extending the framework to more complex real-world applications.