Type Recognition of Frequency Synthesizer Based on Convolutional Neural Networks

This paper mainly proposes a method of frequency synthesizer type recognition based on Convolutional Neural Networks (CNN). The unique influence of the internal structure of the frequency synthesizer on the output signal is proposed in this paper, the phase truncation model of the direct digital frequency synthesizer (DDS) and the model of fractional-N charge pump phase-locked loop (Fractional-N PLL) are constructed respectively. Two models are simulated to obtain signals with 40 frequency. The time-frequency images are obtained by Short-Time Fourier Transform (STFT), as the input of the CNN. The trained networks model could identify which frequency synthesizer the signal comes from. In this experiment, signals with 6 different signal-to-noise ratios (SNR) were used to verify the feasibility of the method. The results show that the CNN-based frequency synthesizer type recognition has a higher accuracy and can realize the type recognition of frequency synthesizer effectively.