Remaining Useful Life Estimation in Aircraft Components with Federated Learning

Abstract


 
 
 In this work, we compare distributed collaborative learning techniques for Prognostic and Health Management (PHM) systems, focusing on predictive aircraft maintenance. Aircraft industry components are usually evaluated using remaining useful life (RUL) estimations to describe the amount of time left before system health falls. Such estimates have been commonly achieved with traditional degradation estimation methods. These estimation methods have been widely applied in centralized processing architectures, limiting the scalability of PHM systems.
 Concerns about data privacy and transfer of large amount of data have also been limiting the construction of decentralized processing architectures. Nevertheless, with the emergence of collaborative training methods of machine learning models, e.g. Federated Learning (FL), the previous referred concerns have being tackled by privacy-preserving communications while keeping data at the network edges. However, the effectiveness of federated learning algorithms using time-series data for prognostic and health management of aircraft systems has been minimally explored.
 In this work, we use feed-forward neural networks on centralized and decentralized scenarios to compare the prediction error minimization of FL algorithms, such as, Federated Average (FedAvg) and Federated Proximal Term (FedProx). Our experiments take into account gradient descendent minimization and averaging weights of neural networks. Using FedAvg, we obtained similar prediction errors to the centralized scenario but presenting uncertain predictions along the aggregation iterations. On the other hand, using FedProx, the prediction error curve progressively decreases along the aggregation iterations if μ takes values ~ 0.01.