24th EANN 2023, 14 - 17 June 2023, León, Spain

Forecasting Functional Time Series using Federated Learning

Raúl Llasag Rosero, Catarina Silva, Bernardete Ribeiro


  The need for accurate time series forecasting has questioned the potential of Federated Learning (FL) in solving regression problems with privacy-preserving and collaborative prognosis requirements. While recent Machine Learning (ML) studies have shown accurate predictions in time series forecasting using functional principal component analysis, the potential of integrating this approach with FL has not been previously evaluated. This paper depicts the potential of combining functional time series regression with FL through the implementation of a Functional Multilayer Perceptron (FMLP). Experimental results on one of the most innovative industrial maintenance strategies, Predictive Maintenance (PM), demonstrate that the integration of FMLP with the well-known Federated Averaging (FedAvg) algorithm achieves accurate time series forecasting while preserving data privacy. These results, obtained using NASA C-MAPSS datasets, outperformed traditional ML and Deep Learning (DL) approaches in estimating the Remaining Useful Life (RUL) of aircraft components.  

*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.