Speaker
Description
Automatic Modulation Classification (AMC) is an important part of next-generation cognitive radio
networks. However, real-time deployment at the network edge remains challenging because of channel
impairments and overlapping signals. This paper presents an efficient multi-modal deep learning
framework integrated into GNU Radio for robust real-time AMC.
Using TorchSig, we generate a comprehensive dataset that includes realistic wireless channel impairments
for both single-signal and overlapping multi-signal scenarios. The proposed framework combines two
feature extraction models: a 1D Convolutional Neural Network (CNN) that learns temporal features from
raw I/Q samples and a 2D CNN that extracts spatial features from signal spectrograms. To improve
classification under changing channel conditions, a decision-level fusion classifier combines the
probability outputs of both models together with real-time Signal-to-Noise Ratio (SNR) estimates for
adaptive multi-label classification.
The complete framework is optimized using ONNX Runtime and deployed as a custom GNU Radio block
for real-time over-the-air testing with Software Defined Radios (SDRs). Experimental results show that
combining temporal and spatial features with an SNR-aware fusion classifier improves classification
accuracy, especially in low-SNR and overlapping signal scenarios. By keeping the individual models
lightweight, the proposed framework also achieves high-throughput, resource-efficient inference suitable
for edge devices.
| Talk Length | 15 Minutes |
|---|---|
| Acknowledge | Acknowledge In-Person |