5GHz Machine Learning Device Positioning Prediction Model

Project Description

The Global Positioning System (GPS) is widely used for outdoor positioning and navigation. However, GPS relies on satellite signals that are significantly attenuated by buildings and obstacles, making it unreliable indoors. As a result, alternative approaches are required to enable accurate device positioning in scenarios with multipath interference and signal attenuation [4].

This project aims to address the limitations of GPS by developing a machine learning (ML)-based positioning prediction model that operates independently of satellite signals and instead leverages communication between devices within the same 5 GHz localised network. To achieve this, the project will evaluate and compare signal-based, direction-based and time-based indoor positioning methods [2], under emulated conditions. A key assumption in this project is that multi-path interference will not significantly affect localisation performance. Therefore, while multi-path propagation will be considered, the system will focus only on the strongest received signal when estimating position. Moreover, the specific problem addressed in this project is whether a machine learning model can outperform conventional indoor positioning methods by learning patterns and compensating for environmental conditions that arise in both indoor and outdoor propagation environments, such as noise, interference, non-line-of-sight (NLOS), and line-of-sight (LOS) [1].

This problem provides a sufficient challenge for an undergraduate dissertation due to the mathematical complexity of using device positioning methods and the intricacies of generating realistic network behaviours. Furthermore, the project requires evaluating multiple positioning methods, as well as designing and training a machine learning model. Hence, this project will demonstrate appropriate technical depth, independent problem-solving, and critical analysis for a computer science dissertation.

This project is sponsored by Clear-Com, a telecommunications company that provides hardware and embedded software solutions for a range of professional sectors, including entertainment, nuclear facilities, military operations, and space exploration. Since Clear-Com devices are used in both indoor and outdoor environments, where accurate device positioning can be crucial, they can particularly benefit from this project. However, many organisations face the same issues and constraints, allowing for the outcomes of this project to be applicable beyond a single company or use case.

Aims & Objectives

The overall aim of this project is to investigate whether a machine learning model can outperform conventional indoor positioning methods by learning patterns and compensating for variations caused by network constraints and environmental conditions.

The project will be structured into four main stages, each targeting a specific objective which forms a sequential pipeline in which the output of each stage is used as the input to the next.

1) Generate network specifications & constraints:

Multiple two-dimensional network environments will be emulated by randomly generating scenario parameters and constraints. Each environment scenario will include all data required to perform positioning estimations. This includes generating the exact (ground-truth) position of the target device and other devices in the network, obstacles, material attenuation coefficients, noise, interference, signal strength measurements, etc.

• Risk: Failing to represent realistic network conditions or introducing bias by lacking enough variation in the generated data.

• Mitigation: Generating data that offers a representative snapshot of diverse network reception conditions, including varying levels of noise, interference, and propagation conditions, to ensure sufficient diversity and relevance [1].

2) Calculate position estimations:

The generated network environments will be used to produce multiple position estimates. For each network scenario, the position of the same target device will be calculated using three indoor positioning methods: Received Signal Strength Indicators (RSSI), Time of Arrival (TOA), and Angle of Arrival (AOA).

• Risk: Emulated environments may oversimplify network conditions affecting the performance of some positioning methods more than others.
• Mitigation: Attenuation, noise, and LOS/NLOS conditions will be incorporated to improve dataset realism [1].

3) Construct a labelled dataset and train the ML model:

The results from the positioning estimation phase will be mapped back to their corresponding network environments to form a labelled training dataset. This dataset will be used to train the ML model to analyse patterns between the ground-truth device position, the different estimated positions, and the related environmental constraints for each network scenario.

• Risk: The dataset size may be insufficient to support a reliable machine learning learning performance, leading to high variance, overfitting, or inflated accuracy estimates [3].
• Mitigation: The dataset size will be incrementally increased until the model performance stabilises, and regularisation techniques will be used to prevent overfitting [3].

4) Evaluate the ML model:

The trained ML model will be assessed to determine whether it outperforms individual conventional indoor positioning methods.

• Risk: The model may not achieve significant improvements.
• Mitigation: If this occurs, the model may instead be repurposed to assist conventional positioning methods by refining existing measurements or correcting systematic errors, rather than directly predicting the device positions [1].

References

1. Alawieh, M. and Kontes, G. 5G positioning advancements with AI/ML. arXiv preprint, arXiv:2401.02427, 2023. Available at: https://arxiv.org/abs/2401.02427 [accessed March 2023].
2. Rathnayake, R.M.M.R., Maduranga, M.W.P., Tilwari, V. and Dissanayake, M.B. RSSI and machine learning-based indoor localization systems for smart cities. Eng, 4(2), pp. 1468–1494. Available at: https://doi.org/10.3390/eng4020085 [accessed March 2023].
3. Rajput, D., Wang, W.J. and Chen, C.C. Evaluation of a decided sample size in machine learning applications. BMC Bioinformatics, 24, p. 48, 2023. Available at: https://doi.org/10.1186/s12859-023-05156-9 [accessed March 2023].
4. Xie, T., Jiang, H., Zhao, X. and Zhang, C. A Wi-Fi-based wireless indoor position sensing system with multipath interference mitigation. Sensors, 19(18), p. 3983. Available at: https://doi.org/10.3390/s19183983 [accessed March 2023].

---

Authors

Giuliana E, Victor Romero Cano, Juan Hernandez Vega & Arnold Chau