Attention-Enhanced Learning for Sensing-Assisted Long-Term Beam Tracking in mmWave Communications

Technical Summary: Attention-Enhanced Learning for Sensing-Assisted Long-Term Beam Tracking in mmWave Communications

Problem Statement
In millimeter-wave (mmWave) communications, beam training incurs significant overhead regarding signaling, latency, and power consumption due to fast time-varying channels and sensitivity to blockages and mobility. While sensing-assisted communications (ISAC) have emerged as a solution using modalities like GNSS, LiDAR, radar, and vision, existing approaches face two primary limitations:

1. Complexity: Effectively fusing features from multiple sensing modalities often requires high-complexity machine learning (ML) models (e.g., multiple Transformer architectures), hindering practical implementation.
2. Temporal Scope: Most existing works focus on predicting only the current beam based on current or past data. This necessitates frequent inference at every time step, maintaining high overhead. While long-term beam prediction (predicting multiple future slots) can alleviate this, it remains largely unexplored. Existing long-term prediction methods often rely on costly LiDAR or require large models to process raw images.

Methodology
The authors propose an efficient, low-complexity, attention-enhanced ML framework for long-term beam tracking using vision data. The system considers a downlink mmWave scenario where a Base Station (BS) equipped with a Uniform Linear Array (ULA) and an RGB camera serves a mobile User Equipment (UE).

• Problem Formulation: The beam tracking problem is formulated as a sequence-to-sequence classification task. Given a predefined beamformer codebook $\mathcal{V}$, the goal is to predict the optimal beam indices for the current time slot $t$ and $J$ future slots ($t+1, \dots, t+J$) based on a sequence of past and current RGB images ($L$ time steps). The objective is to maximize spectral efficiency (or received signal power in low SNR scenarios) without requiring Channel State Information (CSI).
• Model Architecture: The proposed model consists of three main components:
  1. Efficient Image Preprocessing & CNN: Instead of using heavy pre-trained models like ResNet or object detection networks (e.g., YOLOv4), the authors employ a dedicated five-layer Convolutional Neural Network (CNN). This network uses standard PyTorch modules (Conv2d, BatchNorm2d, ReLU, MaxPool2d) to extract compact spatial features from the input image sequence. Background interference is suppressed via adjacent-frame subtraction and motion masks.
  2. Temporal Modeling (GRU): The extracted feature vectors are fed into Gated Recurrent Unit (GRU) networks. The GRUs capture temporal dependencies step-by-step, updating hidden states to produce a context vector summarizing the input sequence.
  3. Temporal Attention (MHA): A residual Multihead Attention (MHA) module is integrated after the GRU. This mechanism allows the model to attend to all positions in the sequence simultaneously, capturing global temporal dependencies that GRUs might miss (which focus more on local dependencies). This enhances the model's ability to predict beams for distant future time slots.
• Training Strategy: The model is trained using a focal loss function to address class imbalance in the beam index distribution (where optimal beams are non-uniformly distributed). The training minimizes the sum of focal losses over the $J+1$ time slots.

Key Contributions

1. Long-Term Prediction Framework: The paper introduces a learning framework that directly solves the long-term beam tracking problem, predicting optimal beams for the current and six future time slots simultaneously, rather than decoupling the problem into single-step predictions.
2. Low-Complexity Architecture: By replacing heavy pre-trained networks with a custom, lightweight CNN and integrating GRUs with MHA, the authors achieve a model with approximately $1.8 \times 10^6$ trainable parameters.
3. Attention Mechanism for Temporal Dependencies: The integration of the MHA module specifically targets the improvement of prediction accuracy for future time slots, addressing the degradation of performance as the prediction horizon extends.

Numerical Results
The proposed method was evaluated using the DeepSense 6G dataset (Scenario 9), with $L=8$ past images and $J=6$ future slots.

• Accuracy: The proposed model with MHA achieved a Top-5 beam prediction accuracy exceeding 90% across both current and future time slots. Specifically, it reached 93.35% Top-5 accuracy compared to 92.31% without MHA.
• Comparison with State-of-the-Art: The model achieved 97% of the performance of the optimal scheme (and the scheme in [19]) in terms of Average Distance-Based Accuracy (ADBA), but with only 3% of the computational complexity (parameter count).
• Impact of Attention: The MHA module provided significant gains for distant future slots (e.g., $t_5$). Compared to the architecture without MHA, the attention-enhanced model improved Top-1 accuracy by 6.34 percentage points and Top-3 accuracy by 6.59 percentage points at $t_5$.
• Efficiency: The design significantly reduces the overhead associated with sensing and processing for beam training.

Significance and Claims
The paper claims that the proposed attention-enhanced design offers a practical solution for long-term beam tracking in mmWave communications. By leveraging vision data and a low-complexity ML model, it significantly reduces the signaling overhead, latency, and power consumption typically associated with frequent beam training. The authors assert that their design achieves comparable accuracy to state-of-the-art methods while drastically lowering computational requirements, making it suitable for practical implementation.

The authors note modestly that the current design is constrained by dataset limitations and low-complexity requirements, which may affect generalization under domain shifts and robustness in multi-user scenarios. These aspects are identified as subjects for future study rather than current achievements.