Semi-Supervised Novelty Detection for Precise Ultra-Wideband Error Signal Prediction

TL;DR

The paper tackles the challenge of unreliable UWB localization in dynamic, non-line-of-sight environments by introducing EPSNoDe, a semi-supervised novelty-detection framework that learns a map-specific UWB quality fingerprint using an overcomplete autoencoder trained on nominal data. Novelty is detected through reconstruction errors, with per-anchor and total error metrics guiding localization reliability assessments. Different input configurations (RNG, MA, PCA) are evaluated, revealing that RNG and MA best capture environmental perturbations via CIR reflections, while PCA underperforms, highlighting the importance of feature selection. The evaluation on a controlled office environment demonstrates the method's ability to identify novelty zones and quantify fit to ground truth via KDE and KL divergence, suggesting potential as sensor-fusion priors or navigation costs to enhance UWB-based localization in evolving environments.

Abstract

Ultra-Wideband (UWB) technology is an emerging low-cost solution for localization in a generic environment. However, UWB signal can be affected by signal reflections and non-line-of-sight (NLoS) conditions between anchors; hence, in a broader sense, the specific geometry of the environment and the disposition of obstructing elements in the map may drastically hinder the reliability of UWB for precise robot localization. This work aims to mitigate this problem by learning a map-specific characterization of the UWB quality signal with a fingerprint semi-supervised novelty detection methodology. An unsupervised autoencoder neural network is trained on nominal UWB map conditions, and then it is used to predict errors derived from the introduction of perturbing novelties in the environment. This work poses a step change in the understanding of UWB localization and its reliability in evolving environmental conditions. The resulting performance of the proposed method is proved by fine-grained experiments obtained with a visual tracking ground truth.

Figures (7)

• Figure 1: UWB signal attenuation types depending on the obstacles within and around the environment. A novel method for UWB Error Prediction with Semi-Supervised Novelty Detection (EPSNoDe) is proposed to identify the map areas where the dynamic nature of the environment may affect the UWB signal.
• Figure 2: Architecture of the proposed UWB EPSNoDe model - It consists of an overcomplete autoencoder ($N < N_{E1} < N_{E2}$ and $N_{D1} > N$). The legend illustrates all the layer types. The input dimension varies according to the input type.
• Figure 3: Sketch of the environment used to test the algorithm. Three scenarios of novelty can be identified: A, B, C.
• Figure 4: The total error is computed for each test conducted. Each row corresponds to the architecture type applied to the dataset, and each column corresponds to the dataset type employed in the test. EPSNoDeRNG uses only ranging distances, EPSNoDeMA involves the Moving Average of the CIR along with the ranging distances, and EPSNoDePCA involves the application of Principal Component Analysis to the CIR along with the ranging distances.
• Figure 5: 3D plot obtained using EPSNoDeMA architecture in case B with the obstacle placed within the grid map in the top-right corner. The graph shows the total error computed applying Eq. \ref{['eq:toterr']} for each map grid point. The highest errors are detected on the top-right corner of the map where the bars are higher.