Introduction
Ultra-Wide Band (UWB) technology has become a cornerstone for precise indoor positioning and navigation. However, the reliability of these systems hinges on accurate channel estimation—specifically, the ability to distinguish between Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) signal environments. While conventional Deep Learning (DL) architectures like Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks excel at this classification, their high computational and memory requirements often preclude deployment on low-power, battery-operated edge devices.
This research introduces a fully unsupervised Spiking Neural Network (SNN) framework designed to bridge this gap. By leveraging biological brain dynamics through a Liquid State Machine (LSM) architecture, the proposed system achieves performance comparable to supervised DL models while offering a pathway toward ultra-low-power neuromorphic hardware implementation.
System Architecture
The proposed pipeline utilizes a three-stage unsupervised process to transform raw radio frequency (RF) data into environment classifications.
1. Feature Engineering and Spike Encoding
The system processes two primary data streams: 10-dimensional RF feature vectors (signal statistics) and Channel Impulse Response (CIR) sequences (raw signal shapes). These continuous values are converted into discrete spike trains using rate encoding, ensuring compatibility with the temporal dynamics of spiking neurons.
2. Liquid State Machine (LSM) Encoders
The core of the feature extraction process lies in dual "Liquid Encoders" composed of Leaky Integrate-and-Fire (LIF) neurons. These recurrent networks utilize fixed, untrained synaptic weights to act as a reservoir. This reservoir transforms input spikes into high-dimensional spatio-temporal patterns, capturing complex signal characteristics without the need for expensive backpropagation training.
3. Spiking SOM Classifier
The outputs from the RF and CIR liquid encoders are concatenated and processed by a Spiking Self-Organizing Map (SOM). This classifier learns to categorize signals in a completely unsupervised manner using Spike-Timing-Dependent Plasticity (STDP), a biologically inspired learning rule where synaptic strength is adjusted based on the relative timing of pre- and post-synaptic spikes.
Experimental Evaluation
The framework was evaluated using the eWINE benchmark, a standard public dataset for UWB research focusing on signals from the DW1000 RF chip. To investigate the impact of temporal context, the researchers tested two variations of CIR data: 50-sample and 120-sample segments.
Results and Analysis
The experimental results demonstrate that the unsupervised SNN approach is highly effective for UWB channel estimation. The primary configuration, Strategy 1 (RF + 50-sample CIR), achieved a test accuracy of 80.2%.
Key Findings:
- Competitive Accuracy: The 80.2% accuracy is on par with several supervised deep learning strategies (ranging from ~78% to 82%). While it does not yet surpass state-of-the-art meta-learning methods (83.8%), it achieves comparable results with a significantly simpler, unsupervised architecture.
- The Role of Liquid Dynamics: Ablation studies revealed that the liquid encoders are critical to performance. Removing these components (as seen in Strategies 4, 6, 8, 9, and 10) resulted in accuracy drops of approximately 30%, confirming that recurrent spiking dynamics are essential for effective feature extraction.
- Feature Fusion: The integration of both RF and CIR data proved vital. Strategies utilizing only one data type (e.g., Strategy 3 or 5) showed significantly lower performance compared to the fused approach.
Conclusion
This work demonstrates that unsupervised SNNs can perform complex UWB channel estimation tasks with high accuracy. The reliance on fixed-weight liquid encoders and STDP-based learning makes this architecture exceptionally well-suited for neuromorphic edge hardware like Intel’s Loihi.
Future research will focus on closing the accuracy gap with supervised models by optimizing encoder structures and validating the framework on physical neuromorphic chips to quantify real-world power efficiency. Additionally, exploring automated ANN-to-SNN conversion techniques may provide further avenues for enhancing performance while maintaining low-power characteristics.
Summary updated