Signal Detection Method for OTFS System Based on Adaptive Wavelet Convolutional Neural Network.

Area of Science:

• Wireless Communications
• Signal Processing
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) for Orthogonal Time-Frequency Space (OTFS) signal detection face limitations in feature extraction and handling signal characteristics.
• Fixed kernels in CNNs struggle with the sparsity and non-stationarity of OTFS signals, leading to slow convergence and high training costs.

Purpose of the Study:

• To propose an Adaptive Wavelet Convolutional Neural Network (AWCNN) for enhanced OTFS signal detection.
• To improve feature extraction, convergence efficiency, and overall detection performance in OTFS systems.

Main Methods:

• Replaced fixed convolution kernels in CNNs with adaptive wavelet convolution layers, specifically using Sym4 wavelet kernels with learnable parameters.
• Integrated the original received signal and message-passing (MP) algorithm estimates as input features into the AWCNN model.
• Evaluated the AWCNN model's performance against standard CNNs in terms of convergence efficiency and bit error rate (BER).

Main Results:

• The AWCNN model demonstrated superior convergence efficiency compared to the standard CNN model.
• AWCNN achieved a bit error rate (BER) comparable to CNNs at a low signal-to-noise ratio (SNR) of 2 dB.
• The proposed method operates effectively without requiring pilot-assisted channel state information acquisition.

Conclusions:

• The AWCNN offers a more intrinsically matched approach to OTFS signal characteristics in the delay-Doppler domain.
• AWCNN provides excellent detection performance with faster convergence, making it a promising technique for OTFS systems.
• The integration of adaptive wavelet layers and enhanced input features significantly boosts detection accuracy and efficiency.