Wireless Indoor Positioning Using Channel Charting: A Comparison of Deep and Graph Neural Networks
(2026) In Master’s Theses in Mathematical Sciences FMAM05 20252Mathematics (Faculty of Engineering)
- Abstract
- Indoor positioning is a challenging problem due to complex environments, signal attenuation, and multi-path effects. A promising approach to this problem is channel charting on wireless networks, where high-dimensional channel state information (CSI) is mapped into a low-dimensional latent space, which preserves spatial relationships in a self-supervised manner. Channel charting can be deep learning-based, where a Siamese neural network has demonstrated good performance. Since the Siamese neural network processes pairs of CSI data independently, it does not capture the underlying graph structure of the data. This motivates the use of a Graph neural network (GNN) instead, which is designed to operate directly on graph-structured data and... (More)
- Indoor positioning is a challenging problem due to complex environments, signal attenuation, and multi-path effects. A promising approach to this problem is channel charting on wireless networks, where high-dimensional channel state information (CSI) is mapped into a low-dimensional latent space, which preserves spatial relationships in a self-supervised manner. Channel charting can be deep learning-based, where a Siamese neural network has demonstrated good performance. Since the Siamese neural network processes pairs of CSI data independently, it does not capture the underlying graph structure of the data. This motivates the use of a Graph neural network (GNN) instead, which is designed to operate directly on graph-structured data and can in a better way capture the relationship between data points.
In this thesis, ray tracing simulations have been run in $3$D indoor environments to generate datasets describing how a signal propagates between a transmitter and a receiver. The datasets are used for channel charting-based positioning, where a Siamese neural network is used as a baseline. Several GNNs, with different architectures, are implemented to perform a systematic comparison between the different AI-models. The comparison is made in terms of positioning accuracy, generalization ability, and training efficiency.
The results show that all models achieve very similar positioning accuracy, which means the GNN does not improve performance compared to the Siamese neural network. This suggests that much of the structural information in the data is captured in the optimization of the loss, which is based on dissimilarity metrics. However, GNNs show an advantage when it comes to generalization ability in more complex environments, and some indications suggest better training efficiency, indicating that GNNs are a promising approach for wireless indoor positioning and a direction to move forward in. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9223523
- author
- Hultén, Signe LU
- supervisor
-
- Karl Åström LU
- organization
- course
- FMAM05 20252
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- publication/series
- Master’s Theses in Mathematical Sciences
- report number
- LUTFMA-3608-2026
- ISSN
- 1404-6342
- other publication id
- 2026:E8
- language
- English
- id
- 9223523
- date added to LUP
- 2026-03-16 14:21:38
- date last changed
- 2026-03-16 14:21:38
@misc{9223523,
abstract = {{Indoor positioning is a challenging problem due to complex environments, signal attenuation, and multi-path effects. A promising approach to this problem is channel charting on wireless networks, where high-dimensional channel state information (CSI) is mapped into a low-dimensional latent space, which preserves spatial relationships in a self-supervised manner. Channel charting can be deep learning-based, where a Siamese neural network has demonstrated good performance. Since the Siamese neural network processes pairs of CSI data independently, it does not capture the underlying graph structure of the data. This motivates the use of a Graph neural network (GNN) instead, which is designed to operate directly on graph-structured data and can in a better way capture the relationship between data points.
In this thesis, ray tracing simulations have been run in $3$D indoor environments to generate datasets describing how a signal propagates between a transmitter and a receiver. The datasets are used for channel charting-based positioning, where a Siamese neural network is used as a baseline. Several GNNs, with different architectures, are implemented to perform a systematic comparison between the different AI-models. The comparison is made in terms of positioning accuracy, generalization ability, and training efficiency.
The results show that all models achieve very similar positioning accuracy, which means the GNN does not improve performance compared to the Siamese neural network. This suggests that much of the structural information in the data is captured in the optimization of the loss, which is based on dissimilarity metrics. However, GNNs show an advantage when it comes to generalization ability in more complex environments, and some indications suggest better training efficiency, indicating that GNNs are a promising approach for wireless indoor positioning and a direction to move forward in.}},
author = {{Hultén, Signe}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master’s Theses in Mathematical Sciences}},
title = {{Wireless Indoor Positioning Using Channel Charting: A Comparison of Deep and Graph Neural Networks}},
year = {{2026}},
}