Wheeled Mobile Robot Dead Reckoning Based on Trans-GCN Model

To address the challenge of low positioning accuracy caused by sensor uncertainties in mobile robot dead reckoning systems, this study proposes Trans-GCN, a novel position prediction model that integrates Graph Convolutional Networks (GCN) with a Transformer architecture. The model leverages data-driven AI principles and sensor-specific characteristics to uncover hidden dependencies between wheel speed and inertial data, thereby enhancing navigation accuracy. Initially, the sensor data is segmented using a sliding window approach and represented as multiple graph structures. GCN is employed to capture spatial dependencies by learning the complex topological structures inherent in the data. Subsequently, positional encoding of graph feature signals is embedded into the Transformer, enabling more efficient extraction of global node features. An adaptive learning rate is introduced to enhance flexibility and efficiency in information propagation. The integrated model performs multi-sensor data modeling and feature fusion to predict the two-dimensional displacement increments of the mobile robot at each sampling interval, ultimately reconstructing the navigation trajectory. The model is trained under GNSS availability and used to predict robot positions during GNSS signal degradation or outages. Six sets of experiments were conducted on the publicly available NCLT dataset and a self-collected dataset. Results demonstrate that the proposed model achieves a trajectory fitting accuracy of 89.2%–97.7% in scenarios with partial or complete GNSS failures. The proposed model also improves training and inference speeds by 19.6% and 26.0%, respectively, compared to state-of-the-art methods, validating its superior performance in dead reckoning.