Intelligent Modulation Recognition via Semantic Attribute Embedding

The patent introduces semantic attributes as an intermediary between raw spectrum data and classification decisions. These attributes (e.g., signal bandwidth, cyclic prefix length, subcarrier spacing) serve as mediators that capture essential characteristics of modulation schemes without requiring extensive training data. The attribute learning model extracts these semantic features, which then guide the classification process, reducing dependence on large training datasets while maintaining high recognition accuracy.

The patent performs preliminary action by pre-training the attribute learning model on semantic attribute extraction before the main classification task. This preliminary step establishes a knowledge base of modulation scheme characteristics that can be reused during recognition, reducing the need for extensive task-specific training data. The model learns attribute representations in advance, which are then applied to new signals with minimal additional training.

The patent segments the modulation recognition process into two distinct components: an attribute learning module that extracts semantic features, and a classification module that makes decisions based on these features. This segmentation allows the attribute learning part to be pre-trained once, and then reused for classification with minimal computational overhead. The separation enables the system to achieve low-SNR performance through learned attributes while keeping real-time classification computationally efficient.

The patent changes the parameter representation from raw signal parameters to semantic attribute parameters. By transforming the problem from direct signal analysis to attribute-based analysis, the system achieves better robustness to noise. The semantic attributes (such as detected signal structure characteristics) remain stable even at low SNRs, enabling accurate classification without requiring complex real-time computations on noisy signals.

The patent extracts the essential feature learning function into a separate attribute learning model that operates independently from the classification network. Instead of using a huge end-to-end deep network, the system extracts semantic attributes (bandwidth, cyclic prefix, subcarrier spacing) as intermediate representations. This extraction approach maintains strong feature learning capability while dramatically reducing the computational burden, as the attribute learning model is much smaller and can be pre-trained efficiently.

The patent replaces the need for huge, resource-intensive deep networks with a lightweight attribute learning model. The semantic attributes serve as simple, computationally inexpensive representations that capture the essential features needed for classification. These attribute representations are 'cheap' in terms of computational resources required to process, while still providing the feature learning capability needed for accurate modulation recognition.