Enhanced Modulation Recognition by Time-series Data Augmentation Based Spatiotemporal Multi-channel Framework

Automatic modulation recognition with deep learning has great prospective owing to computing power and big data. However, modulation recognition accuracy depends highly extent on massive volume of data and model applicability. Here, to overcome difficulties such as small sample dataset, manual extraction of features and low accuracy, we proposed an efficient recognition method that combined time-series data augmentation with spatiotemporal multi-channel learning framework. The results showed that the method provided positive indicators on the order of 93.5 percent for ten modulation signal types, which can be improved by at least 15 percent. Especially for QAM16 and QAM64 signals, the average recognition accuracy is increased by nearly 50 percent at SNRs as low as -2 dB, revealing remarkable recognition performance. Effectiveness of the proposed method provides an attractive approach for signal modulation recognition for wired or wireless communication fields.