Improving Beam Training and Tracking With Oversampled-CNN-BiLSTM in mmWave Communication

Abstract

Millimeter-wave (mmWave) wireless communication has several hurdles, including the overhead associated with beam training, the limitations of low-power phased-array topologies, and the problems caused by phase-less power measurements due to oscillator phase noise. Accuracy in beamforming is impacted by traditional beam tracking’s difficulties with high mobility and large arrays. To solve these issues, a novel oversampled convolutional neural network bidirectional long short term memory (CNN-BiLSTM) model is proposed in this paper to train and track the beam. To normalize data and reduce overfitting, synthetic minority over sampling technique (SMOTE) is used. The CNN-BILSTM architecture presented uses batch normalization, max-pooling, ReLU activation, convolution, and normalization layers to extract spatiotemporal features from location and power metrics. This improves the effectiveness of data processing and assists in developing databases for predicting the angle of arrival/angle of departure (AoA/AoD). Lastly, a fully connected layer offers a reliable solution for accurate beam alignment in mmWave communications by predicting AoA/AoD. The results obtained show that the suggested technique achieves accuracy in AoA and AoD estimates while having reduced mean squared error (MSE) as compared to baseline methods. The future work to enhance mmWave beam tracking and training may focus on dynamic adaptation, deep reinforcement learning, multiobjective optimization, hardware optimization, robustness analysis, and integration with 5G and beyond technologies.