Research on Deep Learning-Based Recognition Method for LEO Satellite Terminal Signals

With LEO satellite technologies widely applied in communication and military fields, accurate identification of terminal signal modulation types is critical for spectrum management and information security. To address traditional methods’ limited feature extraction in complex scenarios, this paper develops deep learning-based high-performance recognition models. Key work includes two improved models and optimized training on a self-built dataset:EnhancedCNN integrates residual and channel attention with 5 × 5 convolutions for local feature extraction; KAN-ResNet101 introduces a kernel attention module optimized via Kolmogorov–Arnold theory, adopts 7 × 7 convolutions with SiLU activation for initial feature extraction, and improves low-SNR recognition through residual block bottlenecks. The self-built dataset contains 12,150 samples of 9 signal types across 9 SNR levels (-20 dB to 20 dB), and training is optimized via data augmentation and Focal Loss. KAN-ResNet101 achieves 75% validation accuracy, 5% higher than basic ResNet, with over 98% accuracy for key signals. This study provides a technical solution for LEO satellite signal recognition in complex electromagnetic environments, with reference value in spectrum monitoring and electronic countermeasures.