TL;DR
This paper introduces a novel two-stage framework for Integrated Sensing and Communications (ISAC) in Low-Earth Orbit (LEO) inter-satellite links (ISLs) using Affine Frequency Division Multiplexing (AFDM). It tackles the challenge of outdated channel state information (CSI) by first predicting dominant channel paths and then using this prediction to guide a pre-equalizer that jointly optimizes communication and sensing performance, achieving significant sensing gains with minimal communication degradation.
What Problem Does This Paper Address?
LEO satellites are increasingly forming constellations for global connectivity, relying on inter-satellite links (ISLs) for communication. These links face rapidly changing channels due to high orbital speeds, leading to outdated Channel State Information (CSI) at the transmitter. This outdated CSI is a major hurdle for advanced communication techniques and for Integrated Sensing and Communications (ISAC), where satellites aim to both communicate and sense their environment. Traditional methods often assume perfect CSI or focus solely on communication performance, neglecting the sensing aspect or the impact of CSI aging. Furthermore, existing ISAC approaches typically use grid-based CSI representations and don't explicitly address the unique challenges of doubly selective channels in AFDM waveforms, which are well-suited for these high-mobility scenarios. This paper aims to bridge this gap by developing an ISAC framework that leverages predicted CSI for AFDM-based LEO ISLs, enabling a joint optimization of communication and sensing.
Methods
The paper proposes a two-stage approach:
- Stage I: Path-Level CSI Prediction and AFDM Kernel Reconstruction: This stage focuses on predicting the channel's dominant multipath components. Instead of predicting the entire CSI grid, it tracks a small set of consistent delay-Doppler paths. A GRU-based predictor is used to learn these path parameters (gain, delay, Doppler) from historical data. These predicted path parameters are then used to reconstruct the AFDM delay-Doppler (DD) domain kernel, which serves as the sole CSI input for the next stage. This approach is designed to be robust to phase errors and maintain path identity.
- Stage II: CRLB-Guided Pre-Equalization under Predicted CSI: This stage takes the predicted AFDM kernel from Stage I and designs a sensing-aware pre-equalizer. It starts with a standard Minimum Mean Square Error (MMSE) equalizer and augments it with a regularization term derived from Cramér-Rao Lower Bound (CRLB) considerations. This regularization term is based on a sensitivity map that identifies bins in the DD domain where allocating more power is beneficial for sensing. The resulting pre-equalizer has a closed-form solution and a single tuning parameter (λ_ISAC) that controls the trade-off between communication performance (symbol MSE) and sensing performance (CRLB).
Datasets
The paper uses simulated data for three disjoint LEO ISL passes (Channel 1, Channel 2, Channel 3) to evaluate the predictor's accuracy. These passes represent different geometries and scattering conditions. For the ISAC experiments, the authors use these simulated channels and the predictor's output as the predicted CSI. Ground truth channel parameters are used to generate the "outdated CSI" baseline and for evaluation metrics. No new datasets are introduced; rather, the focus is on evaluating the proposed methods on representative LEO ISL scenarios.
Key Insights and Ideas
- Path-level prediction is crucial for AFDM-ISAC: Directly predicting dominant delay-Doppler paths, rather than the entire CSI grid, is more effective for AFDM-based systems and enables a more structured approach to ISAC.
- Predicted CSI can enable effective ISAC: The framework demonstrates that even with predicted CSI (rather than perfect or outdated CSI), significant improvements in sensing performance can be achieved without substantial degradation in communication quality.
- A single parameter for a smooth trade-off: The proposed CRLB-guided pre-equalizer elegantly balances communication and sensing with a single parameter (λ_ISAC), allowing for a tunable trade-off across different operating regimes.
- Sensing benefits from structured channel shaping: The sensing-aware pre-equalizer reshapes the effective channel in a way that enhances the observability of dominant paths for sensing, going beyond simple power allocation.
- Robustness to prediction errors: The proposed method shows resilience to moderate prediction errors, maintaining stable performance across different LEO ISL trajectories.
Key Results
- Prediction Accuracy: The path-level predictor achieves a mean complex normalized MSE (CNMSE) of about 6% across three test channels, with a 90th percentile CNMSE below 20%. Amplitude SMAPE is around 2%, and mean absolute phase error is around 25 degrees, which is sufficient for the downstream ISAC tasks.
- Communication-Sensing Trade-off: At moderate to high SNRs (10-15 dB), increasing the sensing weight (λ_ISAC) leads to a significant reduction in CRLB (up to an order of magnitude) with only minor increases in MSE and SER, demonstrating a favorable trade-off.
- Outdated CSI vs. Predicted CSI: The predicted CSI consistently outperforms outdated CSI, especially at higher SNRs, leading to better sensing gains and slightly lower communication errors for the same sensing weight.
- Effective Channel Shaping: Visualizations show that the sensing-aware pre-equalizer reshapes the effective channel to create a denser, more structured footprint in the delay-Doppler domain, enhancing sensing observability.
- Ablation Study: The proposed structured sensing map (Sf) is shown to be critical for achieving a controllable trade-off, outperforming constant or random regularization schemes.
Limitations and Future Work
- Single-Antenna Assumption: The current work focuses on single-antenna ISLs. Future work could extend the framework to multi-antenna systems.
- Simplified Channel Model: The paper uses a sparse multipath model. More complex scattering models could be investigated.
- Independent Stages: While Stage I informs Stage II, the prediction and pre-equalization stages are trained independently. Future work could explore joint learning of both stages for potentially greater gains.
- Error Propagation Analysis: While the paper shows robustness to prediction errors, a more systematic analysis of how prediction errors propagate and affect the ISAC performance is suggested for future work.
Summary updated