FedDwa: A lightweight federated learning with dynamic weighted average aggregation method for machines RUL prediction

Remaining useful life (RUL) prediction of machines is crucial for modern industrial scenarios. Deep learning methods have been extensively employed in machine RUL prediction due to their flexibility and excellent modeling ability. However, existing methods inevitably face the following problems. (1) These deep learning-based prediction methods are over-parameterized with excessive model complexity, which makes them difficult to deploy in edge devices. (2) The availability of limited data from different clients cannot be reasonably evaluated due to privacy protection and potential conflicts of interest. To overcome these challenges, this paper proposes a lightweight federated learning with dynamic weighted average aggregation (FedDwa) method. The method utilizes a central server and multiple edge clients to train the model without sharing data. First, an innovatively local prediction model called adaptive sparse self-attention graph convolution gated recurrent unit is developed to improve the prediction performance under finite samples of a single client while reducing the model complexity. On this basis, the designed dynamic weighted average aggregation method is capable of dynamically assigning the weights on different edge client models to mitigate the unfavorable effects of low-precision local models. Finally, experimental results on two real-world machine degradation datasets not only validate the effectiveness and superiority of the proposed FedDwa, but also prove its strong adaptability to limited data and privacy preservation scenarios in widespread industrial applications.