🛰️ UWB Indoor Positioning using CNN-based NLoS Classification & ML-based Position Estimation

This repository contains the full implementation of a UWB-based indoor positioning system developed as part of my Master’s thesis and extracted research paper.
The project focuses on:

• NLoS/LoS classification using a novel CNN architecture
• Machine learning & ensemble learning for precise position estimation
• Signal preprocessing using Channel Impulse Response (CIR)

All simulations, dataset preprocessing, CNN training, and ML-based positioning models are included in this repository.

---

📌 Project Overview

In UWB indoor positioning systems, Non-Line-of-Sight (NLoS) propagation introduces positive ranging bias, which significantly reduces localization accuracy.
To address this issue, this project proposes:

1. A deep CNN-based classifier to distinguish LoS vs NLoS using raw CIR data
2. Regression-based positioning using classical machine learning and ensemble learning
3. Performance evaluation using MAE, CDF, and real-vs-predicted plots

The proposed CNN achieved over 92% classification accuracy and significantly improved positioning performance under NLoS conditions.

Note: The UWB dataset used in this project is publicly available and can be accessed from the following official source: Dataset
Due to file size limitations, the raw dataset is not included in this repository.

---

🧠 System Architecture

1️⃣ Data Preprocessing Pipeline

Implemented in:

• Create Dataset.py

Main steps:

• Extracting absolute values of complex CIR
• Removing unnecessary columns
• Stacking and merging multi-anchor CIR measurements
• Building final training/testing CSV datasets

**Figure:** Sample CIR data used as the input to the CNN network

---

2️⃣ Novel CNN-Based LoS/NLoS Classification

Implemented in:

• CNN_New_Stracture.py

In this work, a novel 1D CNN architecture is proposed specifically designed for UWB Channel Impulse Response (CIR) signals to achieve high-accuracy LoS/NLoS classification.

Key model structure:

• 1D Convolutional Neural Networks (Conv1D) optimized for temporal CIR features
• Max-Pooling layers for dimensionality reduction
• Fully Connected layers:
  • FC1 = 128
  • FC2 = 64 or 128
• Dropout regularization:
  • 0.5 (main optimized model)
  • 0.8 (overfitting analysis)

Output:

• Binary classification: LoS / NLoS

Figure: Proposed novel CNN architecture for LoS/NLoS classification based on UWB CIR signals

---

3️⃣ ML & Ensemble Position Estimation

Implemented in:

• positioning.py

Models used:

• K-Nearest Neighbors (KNN)
• Decision Tree
• Random Forest
• XGBoost
• Meta-Learners:
  • Linear Regression
  • Neural Network (MLP)

Techniques:

• K-Fold Cross Validation
• Weighted Ensemble using inverse MAE
• Stacked Meta-Learning
• Execution time benchmarking
• Empirical CDF of positioning error

Figure: The output of the CNN classifier is fed into the regression network along with tag-to-anchor distance and RSS features to estimate the target position along the X, Y, and Z axes.

---

📂 Repository Structure

├── requirements.txt

├── Create Dataset.py # CIR preprocessing & dataset generation

├── CNN_New_Stracture.py # CNN LoS/NLoS classifier

├── positioning.py # ML & Ensemble-based positioning

├── Report.pdf # Full Master’s thesis

├── Paper.pdf # Extracted research paper

└── README.md

---

📌 Citation

If you use any part of this code, methodology, or results in your own research, please cite the following paper:

🔗 https://ieeexplore.ieee.org/abstract/document/10533361