VectorSync: A Deep Learning Approach for IoT Device Localisation

Explore the cutting-edge world of Machine Learning (ML) algorithms in wireless communications with our research project. Our study extensively analyzes ML algorithms, focusing on their time consumption and positional accuracy compared to traditional methods. The spotlight is on predicting the Reconfigurable Intelligent Surface (RIS) phase using channel coefficients as input.

Our standout model, the VectorSync Beamformer Model, trained on NSB, outshines conventional algorithms in antenna weight prediction, offering superior performance and reduced computation time on advanced devices. While demonstrating rapid calculations for User Equipment (UE) positioning, there's a noted trade-off with positional accuracy compared to state-of-the-art procedures. This intriguing balance sets the stage for future enhancements, aiming to refine the VectorSync Beamformer Model and strike the optimal equilibrium between speed and accuracy in detecting UE positions.

Join us on this journey of innovation and exploration!

Table of Contents

• Installation
• Features
• Project Directory
• Results
• References

Installation

Prerequisites

First make sure you have Conda installed on your system. Neither you may setup your environment using pip.

Clone The Repo

git clone https://github.com/omm-prakash/VectorSync.git

Move Into The Repository

cd ./VectorSync

Create Conda Environment

Use the provided environment.yml file to create a Conda environment named "grocery".

conda env create -f environment.yml

Activate Conda Environment

conda activate VectorSync

Features

• RIS Phase Prediction: .
• Antenna Weight Prediction: Used VectorSync to determine antenna weight inplace of NSB and Capon methods.

Project Directory

<!-- root directory -->
.
├── Deep-Learning
│   ├── autoencoder_weights <!-- trained model weights -->
│   │   ├── azimuthal
│   │   │   ├── lstmattention/
│   │   │   ├── lstmmodel/
│   │   │   ├── model1/
│   │   │   ├── model2/
│   │   │   └── model3/
│   │   ├── azimuthal_position
│   │   │   ├── lstmattention/
│   │   │   ├── lstmmodel/
│   │   │   ├── model1/
│   │   │   └── model2/
│   │   └── position
│   │       ├── lstmattention/
│   │       ├── lstmmodel/
│   │       ├── model1/
│   │       └── model2/
│   ├── dataloader_gpu.py
│   ├── loss.py
│   ├── models.py <!-- model architectures -->
│   ├── __pycache__/
│   ├── README.md
│   ├── train_checkpoint.py
│   ├── train_gpu.py
│   └── train_lstm.py
├── environment.yml <!-- environment setup -->
├── README.md
└── Simulation
    ├── beamforming.ipynb <!-- analysis of beforming -->
    ├── capon.py
    ├── data/ <!-- binary/csv files -->
    ├── DoAEstimation.py 
    ├── doa.py
    ├── models.py <!-- model architectures -->
    ├── model_states/ 
    │   ├── model_weights_best_198.pt
    │   └── model_weights_best_97.pt
    ├── music.ipynb
    ├── nsb.ipynb <!-- data generation for model training -->
    ├── simulation.ipynb <!-- testing on scenario -->
    ├── time_and_coorelation_analysis.ipynb
    ├── utils.py
    └── weight.py

Results

Model	Train Loss	Val Loss	Test Loss
VectorSync	3.55e-03	3.42e-03	3.49e-03
Base Model	6.92e-03	6.97e-03	6.93e-03
Autoencoder	8.83e-03	8.89e-03	8.91e-03
DNN	2.44e-02	2.37e-02	2.55e-02

Table 1: With Angle as Input

Model	Train Loss	Val Loss	Test Loss
VectorSync	2.92e-03	2.85e-03	2.85e-03
Base Model	3.02e-03	3.23e-03	3.19e-03
Autoencoder	7.49e-03	7.43e-03	7.15e-03
DNN	2.45e-02	2.44e-02	2.41e-02

Table 2: With Position as Input

Model	Train Loss	Val Loss	Test Loss
VectorSync	2.20e-03	2.17e-03	2.35e-03
Base Model	2.66e-03	2.75e-03	2.62e-03
Autoencoder	7.86e-03	8.44e-03	7.95e-03
DNN	2.23e-02	2.22e-02	2.26e-02

Table 3: With Angle and Position as Input

References

For more details or any clarifications please feel free to contact me @ [email protected] .

Best Regards,
Gagan Mundada
Omm Prakash Sahoo