channel estimation

To address the challenges of limited pilot overhead, low channel state information (CSI) estimation accuracy, and poor generalization in massive MIMO systems, this paper proposes the first prediction-based foundation model framework for CSI estimation. Our method innovatively adopts a vision transformer architecture pretrained across domains, enabling joint spatiotemporal-frequency modeling and adaptive prior-measurement fusion to achieve efficient pilot encoding and strong cross-scenario transferability. This work pioneers the application of prediction-based foundation models to wireless channel estimation. Extensive evaluations demonstrate significant improvements over conventional and state-of-the-art AI-based methods across diverse deployment scenarios: CSI estimation error is reduced by 42%, pilot overhead is supported below 5%, and robustness to channel noise is substantially enhanced.

To address the challenges of untimely and inaccurate CSI acquisition in wireless communications—caused by high pilot overhead and channel aging—this paper proposes a diffusion-based probabilistic CSI prediction framework. The framework decouples the task into two stages: temporal encoding and diffusion-based generation, marking the first application of diffusion models to channel prediction; it explicitly captures CSI’s stochasticity and multimodal distribution. It supports both autoregressive and sequence-to-sequence inference modes and explores a lightweight architecture eliminating the need for an explicit temporal encoder. Leveraging a U-Net/Transformer hybrid backbone and DDIM-accelerated sampling, it balances generation fidelity with computational efficiency. Experiments across multiple channel datasets demonstrate significant improvements over state-of-the-art methods, effectively mitigating the trade-off between pilot overhead and channel aging, thereby enhancing system reliability and spectral efficiency.

This work addresses the scarcity of MIMO channel measurement data under extreme weather conditions, which hinders reliable coverage assessment for 5G/6G networks. To overcome this limitation, the authors propose a conditional diffusion model that, for the first time, incorporates both weather type and intensity as conditioning inputs. Leveraging only pilot-based channel state information (CSI) estimates collected under mild weather, the model generates realistic MIMO channels across three distinct weather types and multiple intensity levels. The synthesized channels demonstrate strong performance in key metrics such as downlink bit error rate and outage probability, confirming the model’s generalization capability and scalability in harsh environments. This approach offers an effective solution for channel modeling in scenarios where empirical measurements under adverse weather are unavailable.

Offline-trained neural networks for channel estimation suffer from poor generalization, reliance on prior channel knowledge, and difficulty adapting to unknown time-varying channels. To address this, we propose a network-agnostic synthetic data design principle that models channel statistics and delay-spread dynamics to construct a robust training set covering the boundaries of the channel distribution. Our approach requires neither online fine-tuning nor real-time channel feedback, significantly enhancing model generalization to unseen channels. Experiments demonstrate that the proposed method satisfies prescribed robustness requirements in terms of mean squared error under both fixed and variable delay-spread scenarios. Moreover, it achieves low latency and low computational overhead, making it suitable for deployment in resource-constrained real-time wireless communication systems.

To address the challenges of scarce CSI time-series data, high annotation costs, heterogeneous formats, and short coherence times—limiting prediction and classification performance in integrated sensing and communication (ISAC) systems—this paper proposes CSI-BERT2, the first pretraining-finetuning framework tailored for CSI time-series modeling. Methodologically, it introduces an Adaptive Reweighting Layer (ARL) and a subcarrier-time dual-sensitive MLP architecture to overcome permutation invariance inherent in CSI sequences, and incorporates Masked Prediction Fine-tuning (MPM) to enhance few-shot generalization. Built upon the BERT architecture, CSI-BERT2 unifies self-supervised pretraining with joint time-frequency representation learning. Evaluated on multi-task CSI prediction and fine-grained activity classification, it achieves state-of-the-art (SOTA) performance, notably improving accuracy by 12.7%–23.4% in ultra-low-data regimes (<1k labeled samples).

In 5G link adaptation, inaccurate and delayed channel state information (CSI) arises from channel aging, user mobility, and feedback latency. To address this, this paper proposes two CSI prediction frameworks tailored for TDD and FDD systems, innovatively modeling prediction in the effective SINR domain to jointly optimize accuracy and computational efficiency. We systematically compare Wiener filtering against deep learning methods—including GRU, LSTM, and delay-aware DNN—under dual metrics: mean squared error (MSE) and floating-point operations (FLOPs), evaluating prediction accuracy, complexity, and generalization across diverse channel conditions. Results show that when second-order channel statistics are known, Wiener filtering achieves near-GRU accuracy with significantly lower computational cost; conversely, GRU demonstrates superior generalization under unknown or time-varying channel statistics. The study recommends deep learning (e.g., GRU) for TDD systems, while advocating lightweight classical methods (e.g., Wiener filtering) for FDD systems, thereby revealing architecture-dependent optimal paradigms for CSI prediction.

Traditional AWGN channel simulation fails to account for the multiplicative compression effect introduced by the receiver’s automatic gain control (AGC), leading to a severe mismatch between simulated and real-world CSI amplitude distributions. This work is the first to identify and characterize this nonlinear distortion and proposes the M_QTC calibration framework, which jointly models CSI amplitude statistics through quantile mapping, temporal filtering, and Copula-based subcarrier reordering. Experimental results demonstrate that M_QTC reduces amplitude estimation error by a factor of eight and closes 89% of the fidelity gap relative to empirical measurements. Furthermore, classifiers trained on data generated with M_QTC recover 93% of their true performance in interference detection tasks, substantially outperforming those trained using conventional AWGN-simulated data.

To address the bottlenecks of low pilot efficiency and limited channel estimation accuracy in large-scale OFDM systems, this work pioneers the integration of diffusion models into wireless channel estimation, proposing an AI-native receiver paradigm. Methodologically, it formulates channel estimation as a generative inverse problem and establishes an end-to-end trainable joint framework for channel estimation and signal recovery, enabling synergistic optimization between generative AI and classical communication signal processing. Experimental results demonstrate that the approach significantly enhances channel reconstruction fidelity—even from coarse initial estimates—thereby surpassing the performance ceilings of conventional methods. This work establishes a novel, interpretable, high-performance, and generalizable technical pathway for 6G intelligent transceiver design.

This study addresses a critical gap in communication-aware robotic planning, where existing approaches commonly rely on channel-level metrics to predict end-to-end 5G throughput—a practice lacking empirical validation in private 5G deployments. Conducted in a shielded underground industrial environment, the work integrates commercial ray-tracing simulations, Gaussian process regression with a rational quadratic kernel, a mobile robotic platform, and off-the-shelf 5G user equipment to perform real-world measurements. It reveals for the first time that dynamic adaptation of MIMO spatial layers is the primary cause of systematic overestimation of throughput by conventional channel models, with ray tracing significantly overpredicting performance even in line-of-sight conditions. In contrast, a data-driven approach that directly learns end-to-end throughput reduces prediction error by approximately two-thirds and exhibits near-zero bias, demonstrating clear superiority over traditional channel-centric modeling.