Transfer Learning based Intra-Modulation of Pulse Classification using the Continuous Paul-Wavelet Transform

This paper presents the evaluation of an approach for automatic modulation classification (AMC) using continuous wavelet transform (CWT) with the Paul-wavelet as the signal input domain for a pre-trained convolutional neural network (CNN). AMC plays an elementary role in the field of electronic support measures (ESM) within the context of electronic warfare (EW) for the correct emitter and threat detection as well as classification respectively. Similarly, in the area of electronic countermeasures (ECM) for consistent jamming performance, accurate knowledge of the modulation used is necessary. The focus of the presented work is on transfer learning, an approach where a pre-trained convolutional network is applied to modulation classification by re-training the last learnable and the final classification layer. The CWT using the Paul-wavelet offers feature-visibility of transients, especially interesting for phase-modulated signals and therefore offers the ability as feature input domain to the CNN. Accordingly, the classification approach was performed and evaluated with selected modulation and pulse parameter configurations under different signal-to-noise ratios (SNR). The approach presented here shows an overall average classification rate of over 90 % for SNR values greater 5 dB.