Echoes of Accuracy: Enhancing Ultrasonic Indoor Positioning for Energy-Neutral Devices With Neural Network Approaches

With increasing interest in indoor positioning systems (IPS) across various domains such as industry, retail and healthcare, the search for optimal solutions to meet the needs of different applications has gained significant momentum. This work highlights the potential of hybrid RF-acoustic systems combined with advanced machine learning (ML) models for robust, scalable, and energyefficient indoor localization. The focus is on enhancing positioning algorithms for energy-neutral devices to improve accuracy, precision, reliability and ease of installation. Traditional model-based (MB) methods, relying on Lineof-Sight (LoS) components, often struggle in challenging Non-Line-of-Sight (NLoS) and reverberant environments. To address this, we propose data-driven neural network (NN) approaches capable of harnessing multipath components (MPCs) as additional information. The echoes in the room are exploited to improve accuracy. Various NN architectures, including multilayer perceptrons (MLPs), (circular) convolutional neural networks (C)CNNs, and graph neural networks (GNNs) are evaluated, in first instance using synthetic data. Results demonstrate that especially GNNs outperform MB methods, achieving superior accuracy in both LoS and NLoS scenarios. During the second phase, extensive real-life experiments are carried out. The GNN is evaluated using cross-validation, training on measurement data, and transfer learning (TL) within a reverberant NLoS environment. The cross-validation and TL demonstrate the practical feasibility. We report over 80% of improvement in 3D positioning error compared to the MB technique.