ShuffleNet-inspired lightweight neural network design for automatic modulation classification methods in ubiquitous IoT cyber–physical systems

Abstract

Automatic modulation classification (AMC) is one of the most important technologies of cognitive radios and ubiquitous internet of things (IoT) cyber–physical systems, and it can be adopted to recognize unknown signals. Recently, deep learning (DL) has been applied into AMC for the advanced classification performance. However, DL-based AMC methods generally have high computation complexity and large model sizes, which means that these methods can be rarely implemented into some IoT devices. In this paper, inspired by ShuffleNet, we design a lightweight convolutional neural network (CNN), which is named as ShuffleCNN, and a ShuffleCNN-based AMC (ShffuleAMC) method is proposed for the ubiquitous IoT cyber–physical systems with orthogonal frequency division multiplexing (OFDM). Besides, we also introduce fast Fourier transform (FFT) to pre-process the OFDM signals for the classification performance improvement, and apply ${\ell }_2$ regularization to avoid overfitting. It is demonstrated by simulation results that our proposed ShuffleAMC method has little performance loss, when compared with the common CNN-based AMC methods. More importantly, our proposed ShuffleAMC method also has the strengths of low computation complexity and few model sizes.

Introduction

Advanced signal processing techniques are required to investigate to develop reliable and secure wireless communications and internet of things (IoT) [1], [2], [3], [4], [5]. Automatic modulation classification (AMC) is one of key technologies of ubiquitous internet of things (IoT) cyber–physical systems and cognitive radios [6], [7], [8], [9], which has been introduced into military and civilian scenarios for analyzing the characterizes of unknown signals. Most of traditional AMC methods are modeled as classification problems based on manmade features [10], [11], such as instantaneous feature (IF) [12], high order cumulants (HOC) [13] and cyclic spectrum-based feature [14], and the original classification algorithms rely on multiple thresholds, but with the development of machine learning (ML), ML-based classifiers [15], [16], such as decision tree (DT), random forest (RF), K-nearest neighbor (KNN) and support vector machine (SVM), are introduced into AMC for better classification performance. In recent years, deep learning (DL) has been applied to various communication technologies [17], [18], [19], [20], [21], [22], [23], [24], including AMC [25], [26], [27], [28], [29], because of its powerful feature extraction and advanced classification capability. DL-based AMC methods has been demonstrated to outperform the feature-based AMC methods, and it also has outstanding robustness in various communication scenarios [28], [29], [30], [31], [32], [33], [34].

However, the previously proposed AMC methods are based on DL models with huge model parameters and high complexity [25], [26], [27], [28], [29], and they must be supported by powerful graph process unit (GPU), which mean that it is difficult for these methods to apply into the practical ubiquitous IoT cyber–physical systems. Thus, we redesign a lightweight convolution neural network (CNN) model for AMC in the orthogonal frequency division multiplexing (OFDM) systems, which is inspired by the famous ShuffleNet [35]. In detail, fast Fourier transform (FFT) is first to pre-process the received OFDM signals. Then, we apply efficient convolution, channel shuffle and global pooling to reconstruct the CNN, which is named as ShuffleCNN. Finally, based on the pre-processed OFDM signals, the ShuffleCNN-based AMC method. i.e., ShuffleAMC, is trained with ${\ell }_2$ regularization, which can avoid overfitting and slightly improving the classification performance. The contributions of this paper are listed as follows.

• The FFT is introduced as the pre-processing method to improve the classification performance at low signal-to-noise ratios (SNRs).

• We design and implement a lightweight CNN (LCNN)-based AMC methods, i.e., ShuffleAMC, and our proposed ShuffleAMC method has the similar classification performance with the original AMC method, and its main advances are the fewer model parameters and lower computation complexity, when compared with the original AMC method.

• The ${\ell }_2$ regularization is introduced into the training processing, and the experimental results show that ${\ell }_2$ regularization with the appropriate regularization factor can prevent overfitting, improve the classification performance and accelerate the training process.