Lightweight Non-Line-of-Sight Channel Detection for ML-assisted Bluetooth Direction Finding

Technical Summary: Lightweight Non-Line-of-Sight Channel Detection for ML-assisted Bluetooth Direction Finding

Problem Statement
Bluetooth Low Energy (BLE) direction finding (DF) is increasingly utilized for indoor industrial localization and retail services. However, its accuracy is significantly compromised in non-line-of-sight (NLOS) environments where multipath reflections and scattering bias angle-of-arrival (AoA) estimates. While Machine Learning (ML) approaches for NLOS detection are well-established for wide-band systems (e.g., UWB, cellular) using Channel Impulse Response (CIR) features, they are not directly applicable to BLE. BLE operates in the constrained 2.4 GHz ISM band with narrow bandwidth, precluding the use of wide-band CIR features. Furthermore, existing public BLE datasets are not optimized for LOS/NLOS classification, and there is a lack of rigorous ML frameworks tailored to native BLE observables, specifically Constant Tone Extension (CTE) In-phase-Quadrature (IQ) statistics. Additionally, practical BLE deployments often occur on resource-constrained embedded platforms, necessitating lightweight algorithms that balance accuracy with low computational complexity, memory footprint, and inference latency.

Methodology
The authors propose a quality-driven, lightweight ML pipeline designed specifically for BLE CTE IQ features. The methodology encompasses four main stages:

1. Data Acquisition and Generation: A controlled measurement campaign was conducted using u-blox XPLR-AOA-3 hardware in two distinct indoor environments: a square room ("R") with metal structures and an office corridor ("O"). A graphene-coated blocker was used to enforce NLOS conditions by eliminating direct signal paths while allowing reflections. This process generated labeled datasets containing 64-dimensional IQ-derived feature vectors for both LOS and NLOS scenarios.
2. Feature Engineering and Preprocessing:
   • Robust Standardization: To mitigate the influence of outliers and heavy-tailed effects common in NLOS data, a robust quantile-based standardization (RobustScaler) was applied, utilizing the Interquartile Range (IQR).
   • Exploratory Analysis: Principal Component Analysis (PCA) was used to verify statistical separability between LOS and NLOS classes. Adaptive Kernel Density Estimation (AKDE) was employed to model class-conditional probability density functions (PDFs), revealing that NLOS distributions exhibit heavier tails and greater dispersion compared to LOS.
3. Kernel Approximation and Feature Mapping: To capture nonlinear structures without the high computational cost of full kernel methods, the study utilizes Nyström Kernel Approximation (NKA). This constructs low-rank nonlinear feature maps. Several kernel types were evaluated, including Random Fourier Features (RFF), Radial Basis Function (RBF), Polynomial, and Sigmoid.
4. Lightweight Classification: The pipeline employs a lightweight Support Vector Classifier (SVC) head operating on the kernel-transformed features. This approach is compared against baselines including Random Forest (RF) and Multilayer Perceptron (MLP). The framework emphasizes "identification-and-discard," where unreliable NLOS channels are detected and excluded from subsequent triangulation.

Key Contributions

• Dataset Generation: Creation of labeled LOS/NLOS datasets from controlled BLE AoA measurements in two distinct propagation environments, accompanied by statistical validation using PCA and AKDE to confirm class separability.
• Kernel-Feature Selection: Demonstration that Nyström Kernel Approximation (NKA) significantly improves LOS/NLOS detection performance by capturing intrinsic nonlinearities in the feature space, outperforming linear baselines.
• Systematic Framework: Proposal of a quality-guided framework for selecting kernel types, hyperparameters, and feature budgets, specifically tailored for resource-constrained BLE devices.
• Low-Complexity Detection: Development and comparison of lightweight classification pipelines that produce calibrated posterior probabilities. This enables cost-aware threshold selection for real-time deployment, balancing misclassification costs.

Results
Experimental results across four dataset configurations (Scenarios "R" and "O", and synthetic "Mixed" and "Proportional" combinations) indicate:

• Accuracy Improvements: NKA-based approaches improved accuracy by approximately 7–14% relative to raw linear SVC baselines. For instance, in the "O" scenario, accuracy reached 93% with NKA-SVC compared to 86% for linear SVC at a 7K training size.
• Model Comparison: While MLP achieved the highest absolute accuracy (approx. 90.8%), it incurred higher training complexity and sensitivity to hyperparameters. Random Forest showed competitive accuracy but lacked memory efficiency.
• Efficiency Trade-offs: The Nyström–SVC approach (specifically using RFF) offered the most favorable trade-off. It reduced inference time by approximately 44% compared to MLP and significantly reduced training time compared to other nonlinear models, while maintaining a small memory footprint (approx. 17 KB for RFF+SVC vs. 120 KB for RF).
• Robustness: Models demonstrated insensitivity to noise, with accuracy degradation limited to 1–3% when additive white Gaussian noise (AWGN) at 20 dB SNR was introduced.
• Calibration: The pipeline successfully generated calibrated posterior probabilities, allowing for adaptive threshold selection based on the desired trade-off between sensitivity and specificity.

Significance and Claims
The paper claims that the proposed framework provides a practical and effective path toward robust NLOS classification-and-discard for industrial BLE AoA localization. The authors emphasize that while deep learning models (MLP) offer high accuracy, the proposed kernel-based linear SVC approach is superior for resource-constrained environments due to its favorable balance of training complexity, inference cost, and memory usage. The work validates that BLE-specific feature transformation is essential for data-driven channel classification, as generic wide-band approaches are insufficient. Finally, the study suggests that integrating these quality indicators into a positioning engine can enable confidence-aware weighting of AoA measurements, thereby improving overall localization reliability without requiring specialized hardware acceleration.