Federated Learning for Collaborative Prognosis

Abstract

Modern industrial assets generate prodigious condition monitoring data. Various prognosis techniques can use this data to predict the asset’s remaining useful life. But the data in most asset fleets is distributed across multiple assets, bound by the privacy policies of the operators, and often legally protected. Such peculiar characteristics make data-driven prognosis an interesting problem. In this paper, we propose Federated Learning as a solution to the above mentioned challenges. Federated Learning enables the manufacturer to utilise condition monitoring data without moving it away from the corresponding assets. Concretely, we demonstrate Federated Averaging algorithm to train feed-forward, and recurrent neural networks for predicting failures in a simulated turbofan fleet. We also analyse the dependence of prediction quality on the various learning parameters.