Introduction
As the wireless industry pivots toward 6G, Integrated Sensing and Communication (ISAC) has emerged as a cornerstone technology. To meet the demands for ultra-high data rates and centimeter-level sensing precision, Extra-Large MIMO (XL-MIMO) systems are being deployed. However, traditional compact antenna arrays with hundreds of elements are costly and face significant challenges in the "near-field" region, where spherical wave propagation couples a user's physical position with the wireless channel state.
Sparse Arrays (SAs) offer a cost-effective alternative by increasing the physical aperture with fewer antennas. Yet, this sparsity introduces new complexities in channel estimation due to reduced spatial correlation. This paper addresses these challenges by proposing CP-Mamba, a deep learning framework that treats positioning and channel estimation as a unified, cooperative task.
The CP-Mamba Framework
The core innovation lies in a two-stage deep learning architecture that leverages the geometric relationship between a user's location and the resulting radio channel. Instead of treating these as isolated problems, the framework uses positioning data to "warm-start" the channel estimation process.
The framework operates in two distinct phases:
- Positioning Stage: The model processes uplink pilot signals to predict the user's 2D coordinates.
- Channel Estimation Stage: The predicted coordinates are used to analytically determine the Line-of-Sight (LoS) component of the channel. This LoS information is then fed back into a second CP-Mamba block, which focuses on reconstructing the complex Non-Line-of-Sight (NLoS) reflections to produce the final full-channel estimate.
The architecture itself combines a multi-scale U-Net backbone with Mamba blocks (Selective State Space Models). This hybrid design allows the network to capture both local spatial features and long-range dependencies across the large antenna array with high computational efficiency.
Experimental Setup and Array Configurations
The researchers evaluated the framework using synthetic datasets modeled at 28 GHz. They compared four distinct antenna configurations: Conventional Compact Arrays (CA), Uniform Sparse Arrays (USA), and advanced sparse configurations like Modular Arrays (MOA) and Nested Arrays (NA). The study utilized 20,000 training samples per configuration, with users randomly positioned in the near-field zone (0.1m to 10m from the base station).
Performance Analysis
The results demonstrate a clear "cooperative gain." By solving for position first, the model effectively denoises the channel estimation task, allowing the neural network to focus on the stochastic NLoS components.
A primary finding is that aperture size is the dominant factor in positioning accuracy. As shown in the performance metrics, Sparse Arrays (NA and MOA) significantly outperform Compact Arrays (CA) in positioning, despite having the same number of antennas. The larger physical footprint provides the high angular resolution necessary for precise near-field sensing.
Furthermore, CP-Mamba proves to be exceptionally efficient. Compared to heavy-duty benchmarks like Wide Residual Networks (WRN), CP-Mamba is approximately 6x smaller in parameter count and 45x faster in terms of floating-point operations (FLOPs), all while achieving superior accuracy in both positioning and channel estimation.
Conclusion
The CP-Mamba framework represents a significant step toward efficient ISAC in 6G systems. By exploiting the inherent coupling between position and channel in XL-MIMO near-field environments, the authors achieve high-precision results with a fraction of the computational overhead of traditional deep learning models. Future work will look to extend these findings to high-mobility, multi-user scenarios.
Summary updated