Federated Learning for improved prediction of failures in Autonomous Guided Vehicles

Abstract

Autonomous Guided Vehicles (AGVs) are nowadays an indispensable component of production lines in smart manufacturing. Managing the fleet of AGVs covers not only the delegation of operational tasks but also the monitoring of AGVs activity and health condition by applying tailored Machine Learning-based methods to detect anomalies in various signals gathered by edge IoT devices mounted on board. Detecting anomalies requires appropriate prediction of selected signals based on multiple types of sensor readings. Momentary energy consumption is one of the signals that can indicate abnormal states in AGVs. In this paper, we show that the prediction of this signal can be improved with the Federated Learning (FL) approach that involves exchanging experience gained by particular AGVs. This paper significantly extends the conference paper (Shubyn et al., 2022) with the new multi-round approach to building global prediction models and recent experiments on real data streams produced by AGVs designed by the AIUT company. The results of our experiments prove that in the AGV operational environments with distributed knowledge Federated Learning performs better than traditional centralized approaches and that frequent synchronization of experience may lead to better prediction quality.