Light Weight Automatic Modulation Classification for Edge Devices in Spectrum Sensing Wireless Networks

With the aid of software-defined radio (SDR) technologies, the recent progress in automatic modulation classification of high accuracy using deep neural networks (DNNs) may play an increasingly important role in spectrum monitoring. However, the computational demand brought by the stochastic gradient descent method in the training process is unbearable in edge devices. In this paper, we proposed a lightweight trained neural network for an Internet of thing (IoT) chip by separating the learning process from the inference process. Besides, we analyzed the influence of the size of the convolution kernel and the length of the sampling samples on the calculation and memory usage of the neural network Experimental results on actual datasets show that the proposed network occupies about 700kb of memory, and requires a single sample floating-point operation per second (FLOPs) of about 400k, while achieving high accuracy of 95.58% at a signal-to-noise ratio (SNR) from -10 to 19dB.