10 October 2025 in IEEE Transactions on Wireless Communications
Accurate channel state information (CSI) prediction is essential for 6G massive multiple-input multiple-output (m-MIMO) IoT systems. Deep learning models, such as Transformers, exhibit quadratic computational complexity, resulting in significant efficiency bottlenecks when processing high-dimensional, long sequences channel data in high-mobility scenarios. The Mamba architecture, distinguished by its unique selective state-space model (SSM), presents a promising solution which combines linear computational complexity with robust capabilities for modeling long-range dependencies. Building on this foundation, we propose ChannelMamba, an end-to-end model specifically designed for channel prediction. First, the model employs a dual-domain input module that captures comprehensive channel features by concurrently processing frequency-domain CSI and delay-domain channel impulse response (CIR) data. Sub-sequently, we develop a cross-path parameter-sharing strategy for the Mamba modules to efficiently capture temporal channel dynamics while enhancing model generalization. Furthermore, to address the multi-dimensional dependencies and global context inherent in channel data, we design a bidirectional Mamba module for cross-feature modeling, enhanced with a lightweight attention mechanism. Finally, extensive experimental evaluations across various standard scenarios demonstrate the significant advantages of ChannelMamba over baseline methods in terms of prediction accuracy, robustness, generalization and computational efficiency, achieving new state-of-the-art performance in channel prediction tasks.
2 July 2025 in IEEE Transactions on Cognitive Communications and Networking
In the context of Industry 4.0, industrial wireless communication faces significant challenges due to the complex and dynamic nature of the channel and the stringent requirements for real-time performance. Existing channel prediction methods exhibit inherent limitations in maintaining prediction accuracy and prediction task latency in heterogeneous environments characterized by both Line-of-Sight (LoS) and Non-Line-of-Sight (NLoS) scenarios. To address these concerns, we propose a Multi-Time-Scale Attention (MTSA) model for channel prediction and implement it within a novel collaborative cloud-edge architecture. Specifically, the cloud prediction model leverages a multi-task learning framework which simultaneously optimizes multiple prediction objectives, enabling high-precision heterogeneous modeling of LoS and NLoS channels. The edge prediction model incorporates a cross-layer parameter sharing mechanism and a dynamic updating strategy in response to environmental changes. To fully exploit the benefits of cloud-edge collaboration, we propose a Multi-Objective Dynamic Task Scheduling (MODTS) algorithm based on double deep Q-network, which effectively optimizes the trade-off between prediction accuracy and prediction task latency. Experimental results demonstrate that the proposed architecture significantly outperforms existing works in terms of channel prediction accuracy, real-time performance, and model robustness.