Dynamic Channel Charting: An LSTM-AE-based Approach

With the development of the sixth-generation (6G) communication system, Channel State Information (CSI) plays a crucial role in improving network performance. Traditional Channel Charting (CC) methods map high-dimensional CSI data to low-dimensional spaces to help reveal the geometric structure of wireless channels. However, most existing CC methods focus on learning static geometric structures and ignore the dynamic nature of the channel over time, leading to instability and poor topological consistency of the channel charting in complex environments. To address this issue, this paper proposes a novel time-series channel charting approach based on the integration of Long Short-Term Memory (LSTM) networks and Auto encoders (AE) (LSTM-AE-CC). This method incorporates a temporal modeling mechanism into the traditional CC framework, capturing temporal dependencies in CSI using LSTM and learning continuous latent representations with AE. The proposed method ensures both geometric consistency of the channel and explicit modeling of the time-varying properties. Experimental results demonstrate that the proposed method outperforms traditional CC methods in various real-world communication scenarios, particularly in terms of channel charting stability, trajectory continuity, and long-term predictability.

💡 Research Summary

The paper addresses a critical limitation of existing channel charting (CC) techniques: their focus on static geometric consistency while ignoring the temporal dynamics of wireless channels. In the context of sixth‑generation (6G) communications, where massive MIMO and distributed antenna systems generate high‑dimensional Channel State Information (CSI) that changes rapidly with user motion and environmental variations, a purely static embedding leads to unstable charts and poor topological fidelity.

To overcome this, the authors propose a novel time‑series channel charting framework that integrates Long Short‑Term Memory (LSTM) networks with an Auto‑Encoder (AE), referred to as LSTM‑AE‑CC. The overall pipeline consists of:

1. Data preprocessing – Raw CSI tensors (original shape (20 827, 4, 2, 4, 13)) are corrected for carrier‑frequency offset and phase, then reshaped and segmented into overlapping windows of length 10 (SEQ_LEN = 10) to form temporal sequences. Each sequence is paired with the ground‑truth user position at the final time step, providing a supervised signal for both spatial and temporal learning.

2. LSTM encoder – A stack of three LSTM layers with 64 hidden units each processes the CSI sequence, capturing long‑range dependencies and producing a final hidden state (h_{\text{enc}}). This hidden state encodes the temporal evolution of the channel over the window.

3. Latent space projection – (h_{\text{enc}}) is passed through a fully‑connected layer to a 32‑dimensional latent vector (z), followed by a ReLU activation. A second linear layer maps (z) to a 2‑dimensional embedding (e), which directly serves as the channel chart.

4. Decoder – The 2‑D embedding is expanded back to the latent dimension, repeated (T) times (where (T) is the sequence length), and fed into an LSTM decoder that reconstructs the original CSI sequence. A final fully‑connected layer restores the high‑dimensional CSI.

5. Loss functions – Training minimizes a composite loss:

   • Reconstruction loss ((L_{\text{recon}})) – Mean‑squared error between the decoded CSI and the original CSI, ensuring the embedding retains sufficient information for accurate reconstruction.
   • Topology loss ((L_{\text{topo}})) – A pairwise distance regularizer that penalizes the squared difference between Euclidean distances in physical space ((d_{ij}^{p})) and distances in the embedding space ((d_{ij}^{e})). This enforces geometric consistency of the chart.
     The total loss is (L_{\text{total}} = L_{\text{recon}} + \alpha L_{\text{topo}}), with (\alpha) tuned experimentally (values around 0.5–1.0).

Experimental setup – The authors use a real‑world indoor massive MIMO dataset collected in a factory‑like environment. The dataset contains 20 827 samples, each with 4 antenna links, 4 sub‑carriers, and 32 CSI points per sub‑carrier. After preprocessing, 18 736 samples form the training set and 2 082 samples the validation set. The model is trained with Adam (learning rate (10^{-3})), batch size 64, for 150 epochs, and a Reduce‑LR‑On‑Plateau scheduler (factor 0.5, patience 5).

Evaluation metrics – Four metrics assess both spatial fidelity and positioning accuracy:

• CT (Correlation of Topology) – Correlation between physical and embedding distances (higher is better).
• TW (Topology Warping) – Geometric consistency measure (higher is better).
• KS (Kullback‑Leibler Divergence) – Distributional similarity (lower is better).
• MAE (Mean Absolute Error) – Average Euclidean error between predicted and true positions (lower is better).

Results – The LSTM‑AE‑CC method achieves CT = 0.9998, TW = 0.9998, KS = 0.0115, and MAE = 0.0619 m, dramatically outperforming the baseline static CC approaches (CT = 0.9966, TW = 0.9965, KS = 0.0835, MAE = 0.4847 m). The MAE reduction of roughly 87 % demonstrates that incorporating temporal modeling substantially improves positioning precision. Visualizations of error vectors show that the proposed model maintains short error vectors even in high‑curvature trajectory segments and near the chart boundaries, whereas the baseline exhibits large deviations and local distortions.

Discussion – By embedding the LSTM directly into the representation learning stage, the method enforces temporal continuity at the same time as spatial consistency, eliminating the need for post‑hoc interpolation or smoothing that previous works rely on. The pairwise topology loss effectively aligns the latent manifold with the true physical layout, preventing the rotation, scaling, or drift problems common in unsupervised embeddings. Limitations include the focus on low‑speed (≈0.3 m/s) indoor motion; performance in high‑speed vehicular or outdoor macro‑cell scenarios remains to be validated. Additionally, the weighting factor (\alpha) is manually tuned, suggesting future work on adaptive loss balancing.

Conclusion and future directions – The study presents a compelling solution for dynamic channel charting, demonstrating that a joint LSTM‑AE architecture can simultaneously capture the temporal evolution of CSI and preserve the underlying geometric structure. This capability is directly applicable to 6G tasks such as real‑time beamforming, adaptive resource allocation, and predictive handover decisions. Future research avenues include extending the framework to high‑mobility environments, automating hyper‑parameter selection (especially (\alpha)), and integrating the dynamic chart with high‑precision positioning algorithms and network‑level control loops.