FEdgeNAV: Edge-Based Federated Learning Framework for Wheel Odometry Fault Detection in Industrial AGV Fleet Navigation

Abstract

Automated Guided Vehicles (AGVs) rely on accurate wheel odometry for reliable navigation; however, wheel slip and varying terrain conditions introduce positioning errors that degrade navigation performance. This paper presents FEdgeNAV, a lightweight edge-based federated learning framework for collaborative wheel odometry fault detection in heterogeneous industrial environments. The proposed framework employs a lightweight 1D CNN-GRU model trained collaboratively across distributed edge clients using Federated Averaging (FedAvg), while keeping raw navigation data local by exchanging only model parameters. Wheel odometry faults are identified through residual analysis between the estimated and measured vehicle speeds. Experimental evaluation on the NavINST dataset demonstrated accurate vehicle-speed estimation and residual-based wheel-odometry fault detection across heterogeneous operating conditions. Furthermore, the proposed framework achieved computationally efficient inference on the NVIDIA Jetson Orin Nano, demonstrating its suitability for real-time deployment on resource-constrained edge platforms for industrial AGVs.