Sensing-Aided Channel Estimation for Near-Field MIMO ISAC Systems via Cross-Attention Transformer

Abstract: Near-field integrated sensing and communication (ISAC) can deliver the high spatial resolution and transmission capability with the shared spectrum and hardware. Due to the partial overlap between communication scatterers and radar targets, the sensing information can provide valuable priors to enhance the channel estimation while fusing the two heterogeneous modalities remain challenging. To address this problem, a Cross-Attention Transformer based Channel Estimation Neural Network (CAT-CENet) is developed, which includes a communication pilot branch generating the the Key and Value features and a sensing information branch generating the Query feature. By elaborating the three-module structure, CAT-CENet can focus on features of overlapped targets automatically without need of identifying them in advance. The modality contribution is theoretically analyzed based on the Shapley value to verify the cross-attention gain achieved by CAT-CENet. Simulation results show that CAT-CENet outperforms the state-of-the-art schemes, especially with the higher overlapping proportion, and is robust to the model pruning.

• The paper presents CAT-CENet, a novel network that fuses radar sensing and communication pilots with a cross-attention transformer to improve channel estimation accuracy.
• It uses a modular design involving modality preprocessing, cross-modality fusion, and convolutional post-processing to extract overlapped channel features without explicit target identification.
• Extensive simulations demonstrate a significant NMSE reduction (to about 10⁻³ at 10 dB SNR) and effective model compression through weight pruning.

Sensing-Aided Channel Estimation for Near-Field MIMO ISAC Systems via Cross-Attention Transformer

Overview

The paper "Sensing-Aided Channel Estimation for Near-Field MIMO ISAC Systems via Cross-Attention Transformer" (2607.03224) presents an advanced methodology for channel estimation in near-field XL-MIMO ISAC configurations. By leveraging radar sensing information through a cross-modal fusion approach, the authors address the modality heterogeneity inherent in ISAC systems, specifically focusing on cases where communication scatterers and radar targets partially overlap. The proposed Cross-Attention Transformer-based Channel Estimation Neural Network (CAT-CENet) achieves superior estimation accuracy by integrating the structural features from both communication and sensing modalities, without requiring explicit identification of overlapped targets.

System Model and Problem Setting

The study assumes a near-field XL-MIMO ISAC system, with a base station equipped with a large antenna array operating in TDD mode. The BS simultaneously serves a mobile user and detects $K$ radar targets, some of which overlap with communication scatterers, introducing valuable prior information for channel estimation. The near-field model is used, capturing both angular and distance parameters, which increases channel dimensionality and complicates the estimation process.

Conventional far-field assumptions and sparsity-based methods, including compressed sensing and Bayesian techniques, are suboptimal in this regime due to their reliance on prior knowledge and stringent sparsity conditions. The uniqueness of the near-field scenario mandates an estimation approach capable of flexibly exploiting both modalities, especially when overlapping targets contribute significantly to the channel structure.

Proposed CAT-CENet Architecture

CAT-CENet is designed with three modules: modality preprocessing, cross-modality fusion, and convolutional post-processing. The architecture distinctly processes communication pilot and radar sensing data, employing 2D convolution to embed these inputs into feature maps. The cross-modality fusion leverages multi-head cross-attention, allocating the query generation to the sensing branch and key/value generation to the communication branch. Dual-attention mechanisms (feature map and spatial) further refine the fused representations by extracting noise and enhancing channel feature extraction.

The post-processing module connects all upstream operations through a residual structure, employing multiple convolutional layers to denoise and output the final estimated channel matrix, facilitating the direct extraction of overlapped structural information and improving robustness.

Shapley Value-Based Contribution Analysis

A significant theoretical contribution is the application of the Shapley value for quantifying the modality contributions to channel estimation. By adapting game-theoretic metrics to assess the influence of radar and communication inputs, CAT-CENet demonstrates a capacity for dynamically identifying and utilizing overlapped target features. The Shapley value analysis visualizes changing contributions under scenarios with varying degrees of overlap, confirming that only overlapped targets contribute meaningfully to estimation performance, thereby validating the network's cross-modal gain properties.

Numerical Performance and Robustness

Extensive simulations were conducted with $M=256$ BS antennas and variable numbers of communication scatterers and radar targets. CAT-CENet achieves substantial performance improvements in normalized mean square error (NMSE) across the SNR spectrum. Notably, NMSE drops to the order of $10^{-3}$ for SNR = 10 dB with only one target overlap, outperforming deep learning and conventional baseline methods.

CAT-CENet's performance advantage scales with increased overlapping targets, demonstrating the network's capacity to fully exploit cross-modality for channel refinement. Additionally, weight pruning experiments yield the pruned PCAT-CENet variant, which achieves almost identical estimation accuracy with significant model size reduction, indicating structural redundancy and efficient scalability.

Practical and Theoretical Implications

This work provides a scalable solution for robust channel estimation in near-field ISAC environments, where partial overlap between radar and communication modalities is expected. The end-to-end learning framework eliminates dependence on a priori knowledge of target overlap, thus supporting more general deployment scenarios in next-generation wireless systems.

From a theoretical perspective, the integration of game-theoretic analysis with deep learning contributes to the explainability and interpretability of cross-modal neural architectures in communication systems. The use of Shapley value quantification unveils the internal dynamics of modality fusion, enabling future research to further dissect and optimize multi-modal estimation approaches.

Future Directions

Further exploration of CAT-CENet could involve extension to multi-user ISAC environments, adaptation to non-uniform arrays, or incorporation of additional modalities such as passive sensing and semantic information. Joint optimization of pilot design, radar waveforms, and transformer architectures may enhance cross-modal synergy. Integration with model compression, distributed learning, and federated architectures could support edge deployment in large-scale MIMO systems for 6G and beyond.

Conclusion

The paper introduces CAT-CENet, a cross-attention transformer-based channel estimation network for near-field XL-MIMO ISAC systems, advancing the fusion of communication and sensing modalities in a highly heterogeneous regime. The approach demonstrates strong empirical and theoretical results, particularly as the overlap proportion between modalities increases, and offers a robust, explainable framework for practical deployment in future integrated wireless sensing and communication systems (2607.03224).