Ultra Lite Convolutional Neural Network for Fast Automatic Modulation Classification in Low-Resource Scenarios

TL;DR

This paper introduces a lightweight ultra lite convolutional neural network (ULCNN) for automatic modulation classification that is efficient enough for low-resource devices, achieving high accuracy with minimal parameters and fast inference times.

Contribution

The paper presents a novel ultra lite CNN architecture specifically designed for low-resource scenarios, combining data augmentation, complex-valued convolution, and other techniques to reduce complexity.

Findings

Achieves 62.47% accuracy on RML2016.10a dataset

Uses only 9,751 parameters

Operates with about 0.775 ms per sample on Raspberry Pi

Abstract

Automatic modulation classification (AMC) is a key technique for designing non-cooperative communication systems, and deep learning (DL) is applied effectively to AMC for improving classification accuracy. However, most of the DL-based AMC methods have a large number of parameters and high computational complexity, and they cannot be directly applied to low-resource scenarios with limited computing power and storage space. In this letter, we propose a fast AMC method with lightweight and low-complexity using ultra lite convolutional neural network (ULCNN) consisting of data augmentation, complex-valued convolution, separable convolution, channel attention, and channel shuffle. Simulation results demonstrate that our proposed ULCNN-based AMC method achieves an average accuracy of 62.47% on RML2016.10a and only 9,751 parameters. Moreover, ULCNN is verified on a typical edge device (Raspberry Pi), where the interference time per sample is about 0.775 ms. The reproducible code can be downloaded from GitHub\footnote{https://github.com/BeechburgPieStar/Ultra-Lite-Convolutional-Neural-Network-for-Automatic-Modulation-Classification}.