Uncertainty aware federated averaging approach for privacy secured collaborative remaining useful life prediction of rolling element bearing

Centralized prediction of remaining useful life (RUL) has demonstrated promising result during predictive maintenance of rolling element bearing. However, centralized learning paradigms for RUL prediction of bearings face significant challenges in industrial scenarios. Firstly, sufficient life-cycle degradation data is difficult to obtain from a single-edge client. Secondly, concerns related to copyright issue contribute to the continued isolation of user data. Thirdly, state-of-the-art methods often overlook integrating uncertainty as feedback to enhance predictive learning for reliable RUL estimation. To address these challenges, this article proposes an uncertainty-aware federated averaging (UAFA) approach within a federated learning framework. Firstly, in the framework, stochastic gradient descent is performed with each local bearing client by monte-carlo dropout (MCD) based long short-term memory network. During local training, a dynamic modulation factor is used to adapt the uncertainty-aware learning rate and the uncertainty components are sent to the central server upon training completion. Finally, client models are aggregated using UAFA and evaluated on multiple bearing datasets. Experiments on several run-to-failure tests show that the UAFA-based framework achieves higher accuracy and lower uncertainty than state-of-the-art (SOTA) aggregation methods. Moreover, UAFA consistently outperforms existing approaches across diverse feature types and client counts, demonstrating strong robustness and generalizability.