A Novel Training Strategy for Deep Learning Model Compression Applied to Automatic Modulation Classification
Mateus A. S. de S. Goldbarg, Micael Balza, Sérgio N. Silva, Lucileide M. D. da Silva, Marcelo A. C. Fernandes · IEEE Open Journal of the Communications Society · 2024
Deep learning techniques, such as deep neural networks (DNNs), have proven highly effective in addressing various automatic modulation classification challenges. However, their computational demands pose a significant hurdle for real-time modulation detection. To tackle this issue, a novel training strategy is proposed in this study. This strategy aims to minimize both pruning and quantization losses during the training of compressed models, thereby reducing the computational complexity of DNNs. The effectiveness of this approach was demonstrated through experiments involving the classification of 24 modulations: OOK, ASK4, ASK8, BPSK, QPSK, PSK8, PSK16, PSK32, APSK16, APSK32, APSK64, APSK128, QAM16, QAM32, QAM64, QAM128, QAM256, AM SSB WC, AM SSB SC, AM DSB WC, AM DSB SC, FM, GMSK and OQPS. Remarkably, the results showed a substantial reduction in both DNN weights and operations, while maintaining a high level of classification accuracy. By streamlining the computational demands of deep learning models, this strategy opens up new possibilities for real-time modulation detection applications, particularly in scenarios where computational resources are limited. This research represents a significant advancement towards the practical deployment of deep learning in modulation classification systems, paving the way for enhanced efficiency and performance in wireless communication technologies.