Siamese Neural Networks for Wireless Positioning and Channel Charting

Eric Lei, Oscar Castañeda, Olav Tirkkonen, Tom Goldstein, Christoph Studer

TL;DR

This paper introduces a unified Siamese neural network architecture that can perform both supervised wireless positioning and unsupervised channel charting using Channel State Information (CSI). This novel approach achieves competitive or superior performance compared to existing methods across various scenarios, including semisupervised learning with limited location data.

What Problem Does This Paper Address?

Accurately determining the location of user equipment (UE) in wireless systems is challenging, especially in complex environments with non-line-of-sight (NLOS) propagation, which is becoming more prevalent with technologies like millimeter-wave. Traditional methods struggle in these conditions. While deep learning has shown promise for CSI-based positioning, it typically requires extensive, high-resolution datasets of CSI and corresponding locations, which are costly and time-consuming to acquire. Channel Charting (CC) offers a way to extract relative positional information from CSI without needing explicit location data, but it doesn't provide absolute positioning. Existing CC methods, like those using autoencoders, can be difficult to train and don't always preserve desirable distance properties. This paper aims to bridge this gap by proposing a single, flexible architecture that can handle supervised positioning, unsupervised channel charting, and even semisupervised positioning (where only a small subset of location data is available).

Methods

The core of this paper is a Siamese neural network architecture. This architecture is inspired by Sammon's mapping, a dimensionality reduction technique. It consists of two identical neural networks (sharing parameters) that process pairs of high-dimensional CSI feature vectors. The network is trained to minimize the difference between the distances of the original high-dimensional CSI vectors and the distances of their corresponding low-dimensional representations.

Key aspects of the method include:

  • Parametric Sammon's Mapping: By using neural networks as the mapping function, the approach becomes parametric, allowing it to learn a function that can map new CSI data to low-dimensional representations.
  • Unified Architecture: The same Siamese network architecture is adapted for different tasks:
  • Supervised Positioning: By adding terms to the loss function that penalize deviations from known ground-truth locations for some data points (anchor points).
  • Unsupervised Channel Charting: By solely minimizing the distance preservation loss between pairs of CSI vectors.
  • Semisupervised Positioning: By incorporating a trainable scaling factor that balances the pairwise distance loss with the anchor point loss, allowing the network to leverage both known locations and the inherent structure of the CSI data.
  • Smooth Approximation: To ensure stable training, Euclidean norms in the loss function are replaced with a smooth approximation to avoid issues with gradients.

Datasets

The paper uses simulated datasets for evaluation.

  • Synthetic Test Dataset: A dataset with 2000 UEs placed randomly in a 200x200 m² area, with an additional test set of 400 points (200 random, 200 on a square shape) for visualization.
  • T-intersection Simulator: A more realistic scenario using a Unity game engine to simulate car movement and generate CSI traces from observed UE positions. This dataset is used to demonstrate performance with limited training data (20 traces).

The paper does not introduce new benchmark datasets but rather uses existing simulation environments and models (Quadriga, "Berlin UMa" scenario) to evaluate its proposed method against established baselines.

Key Insights and Ideas

  • Unified Framework Power: The most significant insight is the demonstration that a single Siamese network architecture can effectively handle supervised positioning, unsupervised channel charting, and semisupervised positioning. This unification simplifies the system design and reduces the need for multiple specialized models.
  • Semisupervised Advantage: The paper suggests that semisupervised learning, even with a small set of annotated data, can outperform fully supervised learning when dealing with limited location information, especially in challenging (NLOS) scenarios. This is attributed to the network learning from the structure of the unlabeled CSI data.
  • Siamese Networks as Regularizers: The authors hypothesize that the Siamese network's training on pairs of data points (effectively learning from (N^2 - N)/2 distinct pairwise distances) acts as a strong regularizer, leading to improved performance over traditional feedforward networks trained on N data points.
  • Parametric CC: The Siamese network provides a parametric approach to channel charting, which is a significant advantage over non-parametric methods like traditional Sammon's mapping, as it allows for mapping new, unseen CSI data.

Key Results

  • Supervised Positioning: The Siamese network consistently outperformed a reference fully-connected neural network (FCNN) in terms of Mean Distance Error (MDE) and Kruskal's Stress (KS) for both Line-of-Sight (LoS) and Non-Line-of-Sight (NLOS) scenarios. For example, in Q-LoS, MDE was 6.69m for Siamese vs. 7.31m for FCNN.
  • Channel Charting (Unsupervised): The Siamese network performed comparably to autoencoders (AE) in unsupervised channel charting, with a slight performance gap in test-set performance for NLOS channels. It also showed similar performance to traditional Sammon's mapping, but with the advantage of being parametric.
  • Semisupervised Positioning: The Siamese network achieved comparable performance to a representation-constrained autoencoder (AE) in semisupervised scenarios, even when trained with only 10% of the UE locations. Interestingly, in some cases, semisupervised learning with the Siamese network showed advantages over fully supervised learning with the same limited anchor points.
  • T-intersection Scenario: The Siamese network demonstrated accurate positioning in a realistic T-intersection scenario with very limited training data (20 traces), achieving an MDE of 2.95m for Q-LoS.

Limitations and Future Work

  • Limited Training Data for Semisupervised: While semisupervised learning showed promise, the paper notes that it did not always show a significant advantage over limited supervised training, suggesting that further improvements are needed to fully leverage semisupervised learning for challenging scenarios.
  • Real-World CSI Evaluation: The current evaluation is based on simulations. The authors mention that an evaluation of Siamese networks with real-world CSI is ongoing.
  • Multi-BS Scenarios: Extensions to scenarios with simultaneous connections to multiple base stations (multi-point CC) are left for future work.
  • Challenging Propagation Scenarios: Further research is needed to develop methods that can further improve semisupervised training for challenging propagation scenarios.

Summary updated

BibTeX
@inproceedings{lei2019siamese,
  author = {Eric Lei and Oscar Castañeda and Olav Tirkkonen and Tom Goldstein and Christoph Studer},
  title = {{Siamese Neural Networks for Wireless Positioning and Channel Charting}},
  booktitle = {Allerton},
  year = {2019},
  doi = {10.1109/allerton.2019.8919897},
}

Loading related content...

Figure