Real-Time Automatic Modulation Recognition Based on FPGA

This paper presents a real-time automatic modulation recognition scheme that can implement deep learning algorithm on Field Programmable Gate Array (FPGA) to meet the performance and power requirements of real-time signal processing in edge computing scenarios. The proposed convolution neural network (CNN), using very few parameters and multiplications/additions, can recognize the modulation types of the wireless signals with high accuracy according to the In-phase and Quadrature component. Then the CNN model is compressed by Int8 quantization and deployed on FPGA fabric for real-time and low-power embedded computing, while maintaining recognition accuracy close to the software performance. The measurement results show that the performance in terms of latency and energy consumption greatly improved compared with other embedded platforms. When compared to other FPGA implementation, the proposed scheme achieves the same performance with fewer hardware resources.