Supervised Contrastive CSI Representation Learning for Massive MIMO Positioning

Junquan Deng, Wei Shi, Jianzhao Zhang, Xianyu Zhang, Chuan Zhang

TL;DR

This paper introduces a novel method for precise indoor positioning using the Channel State Information (CSI) from massive MIMO wireless systems. By employing a deep convolutional neural network trained with a supervised contrastive learning approach, the system learns a robust CSI similarity metric that significantly outperforms existing methods, even in complex radio environments.

What Problem Does This Paper Address?

Accurate location information is crucial for many modern applications, from smart cities to autonomous driving. While massive MIMO (Multiple-Input Multiple-Output) systems in 5G and beyond offer enhanced sensing capabilities, precisely determining a user's location using their CSI remains a challenge. Existing methods often rely on geometrical assumptions, require extensive calibration, or struggle with the complexity and variability of real-world radio environments (like multipath and non-line-of-sight conditions). Fingerprinting and channel charting methods, which leverage machine learning on CSI data, are promising but heavily depend on a reliable similarity metric to determine how "close" two CSI samples are. This paper aims to develop a more robust and universally applicable CSI similarity metric for positioning.

Methods

The core of this paper is a novel approach to learning a CSI similarity metric using supervised contrastive learning and a Deep Convolutional Neural Network (DCNN).

Here's a breakdown:

  • DCNN Encoder: A DCNN is designed to act as an encoder, transforming raw CSI matrices into a lower-dimensional feature representation. This encoder is designed to be robust to environmental factors and doesn't require specific knowledge of the antenna array structure. A key preprocessing step involves computing the autocorrelation matrix of the CSI and then extracting its real and imaginary parts from the upper and lower triangular portions, which is more stable than raw CSI.
  • Contrastive Learning: This machine learning paradigm aims to learn representations where similar samples are pulled closer together in a feature space, and dissimilar samples are pushed apart.
  • Positive and Negative Samples: For a given "anchor" CSI sample, positive samples are CSI measurements taken from nearby locations, while negative samples are from distant locations. These are generated using the available location information from a training dataset.
  • Supervised Contrastive Loss: The authors propose a novel contrastive loss function that considers multiple positive and multiple negative samples for each anchor. This loss encourages the DCNN to map CSI samples from nearby locations to close points in the learned feature space and CSI samples from distant locations to far-apart points.
  • Similarity Metric: Once the DCNN encoder is trained, the learned similarity metric is simply the Euclidean distance between the feature representations of two CSI samples. This learned metric is then used with a k-Nearest Neighbors (kNN) algorithm for positioning.

Datasets

The experiments were conducted using a real-world outdoor massive MIMO dataset collected in a residential area. This dataset consists of CSI measurements from a moving single-antenna transmitter and a fixed 8x8 patch antenna array receiver. The CSI data was collected at a carrier frequency of 1.27 GHz. The dataset contains 4979 labeled CSI measurements with corresponding GPS locations. The authors split this dataset into training and testing sets, using 4481 samples for training and either 498 (10%) or 996 (20%) for testing.

Key Insights and Ideas

  • Universality of CSI Similarity: The learned similarity metric is designed to be universally applicable across different wireless systems and CSI data types, as long as spatial consistency holds for the wireless channels. This is a significant advantage over methods tied to specific hardware or environmental assumptions.
  • Robustness to Environment: The contrastive learning approach, by focusing on relative distances between positive and negative samples, inherently learns to be robust to complex radio environments, including non-line-of-sight (NLOS) and multipath conditions, which are difficult to model deterministically.
  • No Hardware Specifics Needed: The method does not require any specific knowledge of the antenna array structure or calibration, making it more practical for deployment.
  • Contrastive Learning for Positioning: Applying contrastive learning, typically used in computer vision, to learn CSI representations for positioning is a novel application. The key is how they construct positive/negative samples based on spatial proximity.

Key Results

The proposed supervised contrastive learning (SupCon) method demonstrated superior positioning accuracy compared to several state-of-the-art methods:

  • Mean Positioning Error: The SupCon method achieved a mean positioning error of 35.2m on the test set.
  • Comparison to Other Metrics: This significantly outperformed other CSI similarity metrics when used with kNN:
  • Correlation Matrix Distance (CMD): 48.3m
    • Singular Value Decomposition (SVD)-based: 50.8m
    • Channel Magnitude-based: 56.6m
  • Comparison to Direct Mapping (DM): Direct mapping using a similar DCNN architecture achieved a mean error of 54.3m, highlighting the benefit of learning an intermediate similarity space rather than directly mapping CSI to location.
  • CDF Performance: The Cumulative Distribution Function (CDF) showed that 66% of test samples achieved a positioning error less than 25m with SupCon, compared to less than 47% for other methods.
  • Training Efficiency: The DCNN encoder could be trained in less than 10 minutes, making the approach computationally efficient.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the performance of the learned similarity metric is influenced by the number of positive and negative samples used during training (as shown in Fig. 7).
  • While the method is more robust than others, the presence of a long tail in the error distribution is attributed to very low Signal-to-Noise Ratios (SNRs) in certain regions of the measurement area.
  • Future Work:
    • The authors suggest exploring trainable regressors instead of a static kNN estimator in the final positioning step for potentially further accuracy improvements.
  • Combining the learned similarity metric with semi-supervised channel charting methods is proposed, acknowledging that neighborhood relationships in the learned feature space might be non-linear.
  • Investigating the impact of different numbers of positives and negatives on performance is also suggested.

Summary updated

BibTeX
@article{deng2022supervised,
  author = {Junquan Deng and Wei Shi and Jianzhao Zhang and Xianyu Zhang and Chuan Zhang},
  title = {{Supervised Contrastive CSI Representation Learning for Massive MIMO Positioning}},
  journal = {IEEE Communications Letters},
  year = {2022},
  doi = {10.1109/lcomm.2022.3181008},
}

Loading related content...

Figure