Federated learning is a deep learning optimization method that can solve user privacy leakage, and it has positive significance in applying industrial equipment fault diagnosis. However, edge nodes in industrial scenarios are resource-constrained, and it is challenging to meet the computational and communication resource consumption during federated training. The heterogeneity and autonomy of edge nodes will also reduce the efficiency of synchronization optimization. This paper proposes an efficient asynchronous federated learning method to solve this problem. This method allows edge nodes to select part of the model from the cloud for asynchronous updates based on local data distribution, thereby reducing the amount of calculation and communication and improving the efficiency of federated learning. Compared with the original federated learning, this method can reduce the resource requirements at the edge, reduce communication, and improve the training speed in heterogeneous edge environments. This paper uses a heterogeneous edge computing environment composed of multiple computing platforms to verify the effectiveness of the proposed method.

联邦学习是一种可以解决用户隐私泄露的深度学习优化方法，在工业设备故障诊断应用中具有积极意义。然而，工业场景中的边缘节点是资源受限的，在联合训练期间满足计算和通信资源消耗具有挑战性。边缘节点的异构性和自治性也会降低同步优化的效率。本文提出了一种高效的异步联邦学习方法来解决这个问题。这种方法允许边缘节点根据本地数据分布从云端选择部分模型进行异步更新，从而减少计算和通信量，提高联邦学习的效率。与原来的联邦学习相比，这种方法可以减少边缘的资源需求，减少通信，提高异构边缘环境中的训练速度。本文使用由多个计算平台组成的异构边缘计算环境来验证所提出方法的有效性。