LightMamba: Lightweight Mamba Architecture for UWB NLOS Classification

A lightweight Mamba-based neural network for UWB (Ultra-Wideband) NLOS (Non-Line-of-Sight) classification, achieving real-time inference with high accuracy.

Abstract

This repository contains the implementation of LightMamba, a novel lightweight architecture based on Mamba state-space models for UWB signal classification. Our approach efficiently processes Channel Impulse Response (CIR) data to distinguish between Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) conditions in UWB communication systems.

Key Contributions

• Novel Architecture: First application of Mamba architecture to UWB signal classification
• Lightweight Design: Optimized for real-time inference on resource-constrained devices
• Multi-modal Input: Jointly processes CIR features and auxiliary sensor data
• Comprehensive Evaluation: Tested on multiple public datasets with different feature dimensions
• Inference Optimization: Achieves <2ms inference time per sample on CPU

Datasets Supported

Dataset	CIR Features	Aux Features	Scenarios	Source
eWINE (1030)	1016	14	Indoor/Outdoor	eWINE Project
Industrial (1056)	1016	40	Industrial	Industrial UWB
Multi-env (164)	150	14	Office/Industrial/University	Multi-environment

Architecture Overview

Input: CIR [B, T] + Auxiliary Features [B, A]
  ↓
Feature Extraction & Encoding
  ↓
Mamba Sequence Modeling
  ↓
Attention Pooling
  ↓
Classification Head
  ↓
Output: LOS/NLOS Probability

Reproducibility

Environment Setup

pip install torch numpy pandas scikit-learn matplotlib seaborn
pip install mambapy  # Mamba implementation

Quick Start

1. Download Dataset: Place dataset files in respective directories (e.g., 1030/dataset/)

2. Training:

   cd 164/  # or 1030/ or 1056/
   python train.py

3. Inference Benchmarking:

   python infer_single_sample.py --num_runs 1000 --save_json results.json
   

Project Structure

LightMamba/
├── 1030/                    # eWINE dataset experiments
│   ├── model/              # Mamba model implementation
│   ├── utils/              # Data processing utilities
│   ├── train.py           # Training script
│   └── infer_single_sample.py
├── 1056/                    # Industrial dataset experiments
│   └── ...existing code...
├── 164/                     # Multi-environment dataset experiments
│   └── ...existing code...
└── README.md

Datasets

The datasets are available on following websites: https://github.com/JaronFontaine/Industrial-UWB-localization-CIR-dataset; https://github.com/JaronFontaine/UWB-dataset-from-an-office-industrial-and-university-environment; https://github.com/ewine-project/UWB-LOS-NLOS-Data-Set.

License

This project is licensed under the MIT License. See LICENSE file for details.