Scholarlink

In the Sixth-generation wireless communication systems, low-latency and ultra-reliable communications (URLLC) require very energy-efficient beam alignment techniques, especially in high-frequency MIMO. Traditional technologies, such as exhaustive beam sweeps and hierarchical codebook search, have significant latency and signalling overhead, making them impractical for real-time heterogeneous systems. These issues are further exacerbated in 6G environments, which extend across a wide range of use cases, including smart cities, self-driving vehicles, industrial automation, and drone-based communication networks. In these settings, the user activity pattern of high mobility, instability of crisp channel behaviours, and application-specific and changeable deployment circumstances make traditional methods based on learning unsuitable, since they will demand large volumes of domain-specific annotated samples and retraining for each scenario. To curb this, we introduce TL-BeamPredict, a Transfer Learning-based framework to predict beam alignment in heterogeneous MIMO 6G scenarios. TL-BeamPredict utilises a pre-trained hybrid CNN-LSTM model, trained on abundant source domain data (e.g., mmWave channels in urban environments), and fine-tunes it with a few labelled examples of the target domain, such as indoor, motor vehicle, or factory setups. The CNN module derives the spatial channel feature of CSI maps, whereas the LSTM will derive the temporal changes because of movement. Cross-domain generalisation can be achieved using a lightweight adaptation layer in a way that retraining can be achieved quickly without affecting latency or accuracy. Large-scale experiments on datasets DeepMIMO and 6G-Sim have shown that TL-BeamPredict attains an 87.4% beam prediction accuracy that is significantly ahead of conventional models, and can cut down beam alignment delay by 65%. It is highly resilient to dynamic environments, such as those where users move at high speeds and in different topological layouts. The proposed system outperforms existing models by high numbers: the system needs only 60% of labelled samples in the target domain to achieve similar performance. The proposed framework, therefore, provides scalable, dynamic, and computation-compatible real-time beam alignment of future heterogeneous 6G networks to ensure dependable connectivity in highly moving wireless environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Transfer Learning Enabled Beam Alignment Prediction in Heterogeneous MIMO 6G Scenarios

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Heba Alasady

摘要

In the Sixth-generation wireless communication systems, low-latency and ultra-reliable communications (URLLC) require very energy-efficient beam alignment techniques, especially in high-frequency MIMO. Traditional technologies, such as exhaustive beam sweeps and hierarchical codebook search, have significant latency and signalling overhead, making them impractical for real-time heterogeneous systems. These issues are further exacerbated in 6G environments, which extend across a wide range of use cases, including smart cities, self-driving vehicles, industrial automation, and drone-based communication networks. In these settings, the user activity pattern of high mobility, instability of crisp channel behaviours, and application-specific and changeable deployment circumstances make traditional methods based on learning unsuitable, since they will demand large volumes of domain-specific annotated samples and retraining for each scenario. To curb this, we introduce TL-BeamPredict, a Transfer Learning-based framework to predict beam alignment in heterogeneous MIMO 6G scenarios. TL-BeamPredict utilises a pre-trained hybrid CNN-LSTM model, trained on abundant source domain data (e.g., mmWave channels in urban environments), and fine-tunes it with a few labelled examples of the target domain, such as indoor, motor vehicle, or factory setups. The CNN module derives the spatial channel feature of CSI maps, whereas the LSTM will derive the temporal changes because of movement. Cross-domain generalisation can be achieved using a lightweight adaptation layer in a way that retraining can be achieved quickly without affecting latency or accuracy. Large-scale experiments on datasets DeepMIMO and 6G-Sim have shown that TL-BeamPredict attains an 87.4% beam prediction accuracy that is significantly ahead of conventional models, and can cut down beam alignment delay by 65%. It is highly resilient to dynamic environments, such as those where users move at high speeds and in different topological layouts. The proposed system outperforms existing models by high numbers: the system needs only 60% of labelled samples in the target domain to achieve similar performance. The proposed framework, therefore, provides scalable, dynamic, and computation-compatible real-time beam alignment of future heterogeneous 6G networks to ensure dependable connectivity in highly moving wireless environments.