Designing low-power hardware for high-precision cellular localization using attention-based machine learning algorithms
(2025) EITM02 20242Department of Electrical and Information Technology
- Abstract
- This thesis investigates the feasibility and performance of implementing attention-based machine learning models for high-precision cellular localization on low-power hardware. While such models, particularly those using the self-attention mechanism, have demonstrated impressive accuracy in extracting spatial information from wireless signals, they typically rely on GPU acceleration, limiting their applicability in embedded and mobile environments due to power and resource constraints.
To address this challenge, a quantized version of a self-attention model was developed and deployed on an FPGA using VHDL. The study explores multiple quantization strategies to reduce data and weight precision to 8-bit and 4-bit formats. Computationally... (More) - This thesis investigates the feasibility and performance of implementing attention-based machine learning models for high-precision cellular localization on low-power hardware. While such models, particularly those using the self-attention mechanism, have demonstrated impressive accuracy in extracting spatial information from wireless signals, they typically rely on GPU acceleration, limiting their applicability in embedded and mobile environments due to power and resource constraints.
To address this challenge, a quantized version of a self-attention model was developed and deployed on an FPGA using VHDL. The study explores multiple quantization strategies to reduce data and weight precision to 8-bit and 4-bit formats. Computationally intensive operations such as softmax are evaluated and a modified sigmoid activation function suitable for hardware use is introduced. A modular hardware architecture was designed and validated through functional simulation and comparison with a Python-based reference model.
Experimental results show that the hardware implementation achieved comparable accuracy to the original model, with a mean localization error increase of only 0.04m when using 8-bit quantization. Furthermore, the FPGA design consumed over six times less energy than a conventional CPU, demonstrating the significant potential of reconfigurable hardware for power-efficient, real-time AI inference.
This work highlights the viability of deploying transformer-inspired models in resource-constrained systems and contributes to the broader field of hardware-aware machine learning. The methods and insights presented serve as a foundation for future research in low-power, high-accuracy localization and embedded AI acceleration. (Less) - Popular Abstract
- Have you ever tried to find your way inside a huge airport or shopping mall and realized that your phone’s GPS suddenly doesn’t work? That’s because GPS relies on satellites, and inside buildings—or in dense city streets, it often gets confused. But what if we could use the mobile network itself to help locate you, precisely and reliably?
That’s what this thesis is about: using the signals your phone already receives from nearby cell towers to figure out exactly where you are. And not just roughly. With an accuracy down to a few centimeters. Making it work on devices that are small, low-cost, and energy-efficient.
To do this, a machine learning model was trained to "read" the radio signals in the air, much like how a bat navigates... (More) - Have you ever tried to find your way inside a huge airport or shopping mall and realized that your phone’s GPS suddenly doesn’t work? That’s because GPS relies on satellites, and inside buildings—or in dense city streets, it often gets confused. But what if we could use the mobile network itself to help locate you, precisely and reliably?
That’s what this thesis is about: using the signals your phone already receives from nearby cell towers to figure out exactly where you are. And not just roughly. With an accuracy down to a few centimeters. Making it work on devices that are small, low-cost, and energy-efficient.
To do this, a machine learning model was trained to "read" the radio signals in the air, much like how a bat navigates using echoes. These signals change subtly depending on your location, and with enough training data, the AI can learn to recognize the unique patterns, like footprints in the air.
These AI models exist already, but are usually so big and power-hungry that they need powerful graphics cards (GPUs) to run. That works fine for a server farm, but not for something like a drone or a smartwatch. They need to be fast, small, and energy-efficient.
So instead of running the AI on a computer, I built it into a tiny custom chip (FPGA), a reprogrammable piece of hardware that can act like a brain but use way less energy than a computer. I also simplified the model, reducing how much data it needs to work with, and swapped out complicated math with easier, hardware-friendly alternatives.
The result is a smart, efficient chip that can tell where you are by analyzing cellular signals—all in real time, and with amazing precision. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9207189
- author
- Ohlsson, Henrik LU
- supervisor
-
- Liang Liu LU
- organization
- alternative title
- Design av lågenergihårdvara för mobilnät-högprecisionslokalisering med hjälp av uppmärksamhetsbaserade maskininlärningsalgoritmer
- course
- EITM02 20242
- year
- 2025
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Machine Learning, FPGA
- report number
- LU/LTH-EIT 2025-1082
- language
- English
- id
- 9207189
- date added to LUP
- 2025-07-02 08:43:54
- date last changed
- 2025-07-02 08:43:54
@misc{9207189,
abstract = {{This thesis investigates the feasibility and performance of implementing attention-based machine learning models for high-precision cellular localization on low-power hardware. While such models, particularly those using the self-attention mechanism, have demonstrated impressive accuracy in extracting spatial information from wireless signals, they typically rely on GPU acceleration, limiting their applicability in embedded and mobile environments due to power and resource constraints.
To address this challenge, a quantized version of a self-attention model was developed and deployed on an FPGA using VHDL. The study explores multiple quantization strategies to reduce data and weight precision to 8-bit and 4-bit formats. Computationally intensive operations such as softmax are evaluated and a modified sigmoid activation function suitable for hardware use is introduced. A modular hardware architecture was designed and validated through functional simulation and comparison with a Python-based reference model.
Experimental results show that the hardware implementation achieved comparable accuracy to the original model, with a mean localization error increase of only 0.04m when using 8-bit quantization. Furthermore, the FPGA design consumed over six times less energy than a conventional CPU, demonstrating the significant potential of reconfigurable hardware for power-efficient, real-time AI inference.
This work highlights the viability of deploying transformer-inspired models in resource-constrained systems and contributes to the broader field of hardware-aware machine learning. The methods and insights presented serve as a foundation for future research in low-power, high-accuracy localization and embedded AI acceleration.}},
author = {{Ohlsson, Henrik}},
language = {{eng}},
note = {{Student Paper}},
title = {{Designing low-power hardware for high-precision cellular localization using attention-based machine learning algorithms}},
year = {{2025}},
}