Skip to content

Latest commit

 

History

History
202 lines (136 loc) · 6.92 KB

DAReportGroup01.md

File metadata and controls

202 lines (136 loc) · 6.92 KB

UWB LOS/NLOS Classification Project

Group 01 | Data Analytics Course | University of Glasgow | March 2025

Team Members and Contributions

[List of team members with their roles and contributions]

1. Problem Definition

1.1 Background

Indoor positioning systems require accurate distance measurements between anchors and tags. In Ultra-Wideband (UWB) systems, Non-Line-of-Sight (NLOS) conditions significantly degrade ranging accuracy, leading to positioning errors. This project focuses on solving two critical challenges:

  1. Classification Challenge: Accurately identifying whether a UWB measurement is from a Line-of-Sight (LOS) or Non-Line-of-Sight (NLOS) condition.
  2. Range Estimation Challenge: Predicting the true distance between devices, especially in NLOS conditions where measured ranges are typically biased.

1.2 Importance and Applications

Accurate LOS/NLOS classification and range estimation are essential for:

  • Indoor Positioning Systems: Enabling centimeter-level accuracy in indoor environments
  • Robotics and Autonomous Navigation: Providing reliable positioning in complex indoor settings
  • Asset Tracking: Improving inventory management and logistics operations
  • Smart Buildings: Enhancing occupancy detection and room-level positioning
  • AR/VR Applications: Supporting precise spatial mapping and user positioning

1.3 Project Objectives

This project aims to:

  1. Develop an accurate classifier to distinguish between LOS and NLOS UWB measurements
  2. Identify the most significant features that affect LOS/NLOS conditions
  3. Create a range estimation model to correct for biases in NLOS measurements
  4. Evaluate model performance using rigorous metrics and cross-validation

2. Problem Analysis

2.1 Dataset Overview

The UWB LOS/NLOS dataset contains:

  • 42,000 samples (21,000 LOS, 21,000 NLOS)
  • Data collected from 7 different indoor environments
  • 15 metadata features + 1016 Channel Impulse Response (CIR) values per sample

2.2 Initial Data Exploration

Our exploratory analysis revealed:

  • A balanced dataset (50% LOS, 50% NLOS)
  • Significant differences in signal characteristics between LOS and NLOS conditions
  • Strong correlations between certain features and LOS/NLOS classification
  • Variations in CIR patterns indicative of multipath effects in NLOS scenarios

2.3 Key Challenges

The main challenges identified include:

  • High dimensionality due to the 1016 CIR values
  • Potential redundancy in features
  • Variations across different indoor environments
  • Need for robust models that generalize to new environments

2.4 Novelty of Our Approach

Our solution is novel in several ways:

  • We apply feature engineering to extract key characteristics from the CIR
  • We implement a two-stage pipeline: classification followed by range estimation
  • We utilize ensemble learning to improve robustness and accuracy
  • We incorporate domain knowledge about UWB signal propagation into our models

3. Methodology

3.1 Data Preparation and Preprocessing

3.1.1 Data Cleaning

  • Identification and handling of outliers
  • Normalization of CIR values by RXPACC (received preamble count)
  • Feature scaling to ensure algorithm convergence

3.1.2 Feature Selection and Engineering

  • Statistical analysis of feature importance
  • Extraction of key CIR characteristics (peak amplitude, rise time, etc.)
  • Creation of new engineered features (ratios, statistical moments, etc.)

3.1.3 Dimensionality Reduction

  • PCA analysis to reduce CIR dimensionality
  • Selection of top features based on their importance scores

3.2 Data Mining Algorithms

3.2.1 Classification Models

  • Random Forest Classifier
  • Gradient Boosting Classifier
  • Support Vector Machine
  • Neural Network
  • K-Nearest Neighbors

3.2.2 Range Estimation Models

  • Random Forest Regressor
  • Gradient Boosting Regressor
  • Linear Regression with polynomial features

3.2.3 Hyperparameter Tuning

  • Grid search for optimal hyperparameters
  • Cross-validation to prevent overfitting

3.3 Evaluation Methodology

3.3.1 Classification Metrics

  • Accuracy, Precision, Recall, F1-score
  • ROC curve and AUC
  • Confusion matrix

3.3.2 Range Estimation Metrics

  • RMSE (Root Mean Square Error)
  • MAE (Mean Absolute Error)
  • R² score

4. Results and Analysis

4.1 Classification Results

4.1.1 Model Performance Comparison

[Results table comparing the performance of different classification models]

4.1.2 Feature Importance Analysis

[Analysis of the most important features for classification]

4.1.3 ROC Curves

[ROC curves for different classification models]

4.2 Range Estimation Results

4.2.1 Model Performance Comparison

[Results table comparing the performance of different regression models]

4.2.2 Error Distribution Analysis

[Analysis of error distributions in range estimation]

4.3 Detailed Analysis

4.3.1 Environment-Specific Performance

[Analysis of model performance across different environments]

4.3.2 Feature Correlation Analysis

[Correlation analysis between features and target variables]

5. Discussion

5.1 Interpretation of Results

Our results show that:

  • Feature X, Y, and Z are the most significant indicators of NLOS conditions
  • Ensemble methods (Random Forest, Gradient Boosting) outperform other algorithms
  • CIR characteristics provide valuable information beyond metadata features
  • Range estimation accuracy varies with distance and environmental complexity

5.2 Practical Implications

The practical implications of our findings include:

  • Recommendations for UWB system deployment in complex environments
  • Guidelines for anchor placement to minimize NLOS conditions
  • Strategies for integrating our models into real-time positioning systems
  • Suggested enhancements to UWB hardware and signal processing

5.3 Limitations

Our approach has the following limitations:

  • Dependence on the environments represented in the training data
  • Computational complexity of processing full CIR data
  • Challenges in real-time implementation on resource-constrained devices

6. Conclusion and Future Work

6.1 Conclusion

We successfully developed and evaluated models for UWB LOS/NLOS classification and range estimation. Our approaches achieved high accuracy in distinguishing between LOS and NLOS conditions and provided significant improvements in range estimation for NLOS scenarios.

6.2 Future Work

Potential directions for future research include:

  • Integration with sensor fusion approaches (IMU, camera, etc.)
  • Online learning algorithms that adapt to new environments
  • Deep learning architectures for end-to-end CIR processing
  • Transfer learning to adapt models to new UWB hardware

7. References

  1. [List of references]

Appendix

A. Algorithm Pseudocode

[Pseudocode for key algorithms]

B. Additional Visualizations

[Additional plots and visualizations]

C. Implementation Details

[Details on implementation and software used]