CA‐CNN‐GRU Communication Modulation Signal Classification Method Based on Multi‐Scale Feature

Automatic Modulation Classification (AMC) is critical in communication and information processing, enabling efficient identification of modulation types. However, low SNR environments challenge AMC's accuracy, impacting system reliability and anti‐jamming capability. To address this issue, a CA‐CNN‐GRU method based on multi‐scale features is proposed. Time‐domain, frequency‐domain, wavelet‐domain, and entropy features are extracted, forming a 36‐dimensional feature database to represent signals comprehensively. These multi‐scale features serve as input to a CA‐CNN‐GRU model, which integrates channel attention (CA) for feature selection, CNN for local feature extraction, and GRU for sequence modeling. Experiments classify signals from 10 modulation types across 7 SNR levels. Results show 94.25% accuracy at 0 dB SNR, exceeding 98.5% at 5 dB and achieving 100% in optimal conditions. This represents a 24.9% accuracy improvement over traditional CNN methods, enhancing the anti‐jamming ability of communication systems and optimizing signal processing efficiency. This research is expected to provide significant support for the technological development in related fields.