Overview
This work addresses accuracy degradation in Bluetooth Low Energy (BLE) direction-finding specifically within multipath environments. The research focuses on developing a lightweight, machine learning (ML)-based pipeline for detecting non-line-of-sight (NLOS) channel conditions using BLE Constant Tone Extension (CTE) In-phase-Quadrature (IQ) features. The proposed Nyström Kernel Approximation (NKA) coupled with a Support Vector Classifier (SVC) head is presented as a method for channel classification.
Research Context
BLE direction-finding holds promise for indoor industrial localization applications. However, its accuracy is susceptible to biases introduced by reflections and scattering encountered in multipath environments. The detection of line-of-sight (LOS) and non-line-of-sight (NLOS) conditions is a recognized challenge in wide-band radio systems. For BLE direction-finding, there has been a noted lack of narrow-band channel-feature representations, scalable kernel-based feature transformations, and dedicated datasets to facilitate data-driven channel classification.
Approach
The research established a controlled BLE measurement setup designed to generate labeled LOS/NLOS data. This data was collected within two distinct propagation environments. A quality-driven ML pipeline was then developed for BLE CTE IQ features. Key steps in this pipeline included:
- Standardization: Robust quantile-based standardization was applied to the features. This step aimed to mitigate the influence of outliers and heavy-tailed distributions.
- Feature Analysis: Principal Component Analysis (PCA) and Adaptive Kernel Density Estimation (AKDE) were used to analyze the standardized features. This analysis aimed to verify scenario-dependent statistics and determine LOS/NLOS separability.
- Feature Transformation and Classification: Nyström Kernel Approximation (NKA) was employed to construct low-rank nonlinear feature maps. This was followed by the application of a lightweight Support Vector Classifier (SVC) head for the task of LOS/NLOS detection.
- Comparative Evaluation: The performance of the Nyström–SVC classifier was compared against two other models: Random Forest (RF) and Multilayer Perceptron (MLP).
- Threshold Selection: Pipeline-calibrated posterior probabilities were utilized for cost-aware threshold selection. This aimed to enable efficient real-time LOS/NLOS detection in resource-constrained localization systems.
Findings
- The NKA approach demonstrated an improvement in accuracy of approximately 7-14% relative to a raw baseline.
- While the Multilayer Perceptron (MLP) achieved higher absolute accuracy, the Nyström–SVC approach presented a more favorable trade-off concerning training complexity, inference cost, and memory footprint.
- PCA and AKDE analysis of standardized features verified scenario-dependent statistics and revealed LOS/NLOS separability.
Why This Matters
The development of lightweight NLOS detection methods is relevant for improving the accuracy of BLE direction-finding systems. Addressing the challenges posed by multipath environments can enhance the reliability of indoor industrial localization applications, particularly in systems with limited computational resources.