Strategic integration of adaptive sampling and ensemble techniques in federated learning for aircraft engine remaining useful life prediction

Abstract

In industrial manufacturing, predicting the remaining useful life of machines is crucial for improving operational efficiency and reducing maintenance costs. However, data privacy concerns and commercial competition make traditional centralized data collection methods insufficient to meet these needs. Federated learning offers a decentralized training approach that protects data privacy, but existing research faces challenges such as inadequate performance of single models, data quality disparities, and improper client selection strategies. To address these issues, this study proposes an adaptive sampling-based ensemble federated learning framework. By integrating the predictions of multiple models, the framework reduces model errors and enhances prediction accuracy and generalization capability. Additionally, we design an adaptive sampling method that dynamically adjusts the client selection strategy based on data quality, focusing on clients with low-quality data to ensure that their contributions are effectively utilized. Experimental results show that the proposed framework significantly outperforms existing benchmark methods on the turbofan engine dataset, with a 12% reduction in RMSE and a 35% decrease in Score. Ablation experiments and sensitivity analysis confirm that the framework maintains reliable predictive performance and efficiency in dealing with issues such as data imbalance, missing data, and scale changes. Supplementary materials for this article are available online.