Complex-Valued Transformer for Short-Wave Signal Recognition

Short-wave signal recognition provides valuable information for spectrum management and holds significant promise in various applications. Existing studies have predominantly relied on convolutional neural network models for signal recognition. However, these models exhibit a limitation in capturing a comprehensive contextual understanding, necessitating the incorporation of multiple layers for a broader perspective. Moreover, the prevailing neural network architectures are optimized for identifying real-valued signals, proving suboptimal for recognizing the intricate complexities of short-wave signals with complex values. Addressing this gap, we present a novel approach that integrates complex-valued neural network architectures with global attention mechanisms, resulting in complex-valued Transformer designed explicitly for short-wave signal recognition. This model exhibits a remarkable capacity to capture long-term dependencies inherent in complex signals. To enhance its robustness, we introduce and explore signal augmentation techniques tailored to the unique characteristics of short-wave signal patterns. Through comprehensive experimentation, we demonstrate the compelling effectiveness of our complex-valued Transformer when paired with these purpose-designed signal augmentations.