Automatic Modulation Recognition Based on Multi-Source Heterogeneous Networks

Automatic Modulation Recognition (AMR) is crucial to modern communications systems because it identifies the modulation scheme of incoming signals, serving as a key method to address spectrum resource constraints and enhance communication quality. Recently, multi-source models have achieved impressive results in a range of application domains. To improve AMR performance, we propose a multi-source heterogeneous network (MSHNet) that jointly processes time-series signals, images, and graph representations. Transforming the raw signals into images and graphs enables the model to capture complementary features from multiple dimensions, which markedly boosts recognition accuracy. In the time-series branch, we integrate a lifting scheme that extracts features in the frequency domain. For the image branch, the signal is converted into a Gramian Angular Field (GAF) image, supplying the network with two-dimensional spatial information while retaining temporal context. In the graph branch, we construct a graph from feature vector correlations, allowing the model to capture long range dependencies and to view the signal from a broader perspective. Features from the three branches are concatenated and then fused using fully connected layers to integrate the multi-source information. Experiments show that MSHNet outperforms state-of-the-art baselines on the RML2016.10a dataset and maintains strong performance even with a limited number of training samples.