Signal Detection Based on Separable CNN for OTFS Communication Systems.

Area of Science:

• Wireless Communications
• Signal Processing
• Machine Learning

Background:

• Orthogonal Time Frequency Space (OTFS) modulation offers advantages in high-Doppler environments.
• Traditional signal detection methods in OTFS systems face challenges with computational complexity and performance.
• Deep learning approaches, particularly Convolutional Neural Networks (CNNs), show promise but can be resource-intensive.

Purpose of the Study:

• To develop a low-complexity and efficient signal detection method for OTFS systems.
• To enhance feature discrimination and training stability in high-Doppler conditions.
• To reduce the computational overhead of OTFS signal detection for practical deployment.

Main Methods:

• Proposed a novel separable convolutional neural network (SeCNN) architecture, termed SeCNN-OTFS.
• Integrated residual connections and a channel attention mechanism within a SeparableBlock.
• Decomposed standard convolutions into depthwise and pointwise operations to reduce complexity.

Main Results:

• SeCNN-OTFS demonstrated superior performance over Least Squares (LS) and Minimum Mean Square Error (MMSE) estimators.
• Achieved near-identical Bit Error Rate (BER) performance to 2D-CNN at Signal-to-Noise Ratios (SNR) above 12.5 dB.
• Required only 19% of the parameters compared to a standard 2D-CNN, indicating significant complexity reduction.

Conclusions:

• SeCNN-OTFS offers an effective and computationally efficient solution for signal detection in OTFS systems.
• The method is highly suitable for resource-constrained applications like satellite and Internet of Things (IoT) communications.
• A variant with conventional convolutional layers is available for scenarios demanding higher accuracy with sufficient resources.