Frequency-learning adversarial networks based on transfer learning for cross-scenario signal modulation classification

Abstract

Automatic modulation classification (AMC) serves a challenging yet crucial role in wireless communications. Despite deep learning-based approaches being widely used in signal processing, they are challenged by signal distribution variations, especially in various channel conditions. In this paper, we introduce an adversarial transfer framework named frequency-learning adversarial networks (FLANs) based on transfer learning for cross-scenario signal classification. This method uses the stability in the frequency spectrum by introducing a frequency adaptation (FA) technique to incorporate target channel information into source-domain signals. To address the unpredictable interference in the channel, a fitting channel adaptation (FCA) module is used to reduce the difference between the source and target domains caused by variations in the channel environment. Experimental results illustrate that FLANs outperforms state-of-the-art transfer approaches, demonstrating an improved top-1 classification accuracy by about 5.2 percentage points in high signal-to-noise ratio (SNR) scenes on a cross-scenario real collected dataset CSRC2023.

1. Introduction

Automatic modulation classification (AMC), serving as an integral component of signal direction and demodulation in wireless communications, plays a pivotal role in a range of civil and military applications including spectrum monitoring, interference identification, and spectrum management. The importance and application of modulation classification have been extensively explored over several decades. More recently, deep learning (DL) techniques have shown remarkable success in various complex tasks. Particularly, end-to-end models have been recognized for their exceptional performance (O'Shea et al., 2016), enabling a DL system to learn suitable features through a data-driven approach, without needing hand-crafted features. This advancement in DL has been effectively used in communications, particularly in the area of modulation classification. The use of convolutional neural networks (CNNs) in complex-valued radio signal classification has displayed competitive results when compared to traditional approaches based on expert features. With a substantial number of labeled signals, models can “study” the characteristics of the signals autonomously. Another effective tool is the recurrent neural network (RNN), designed specifically for processing time-sequenced signals and their interconnections; despite requiring large volumes of data, RNN has proved effective in signal modulation recognition (Rajendran et al., 2018).

However, wireless radio signals can drastically vary across different channel conditions. The existence of a line of sight (LOS) between the transmitter and the receiver can significantly affect wireless communications, leading to deviations in the statistical characteristics of the received radio samples from the training data. DL algorithms necessitate large volumes of training data that adhere to the independent and identically distributed (i.i.d.) condition with the test data (Pan and Yang, 2010). Unfortunately, obtaining sufficient training data for all potential radio scenarios is not practical, which can compromise the effectiveness of the model. The impact of scenario variation, particularly noticeable in the time domain, can also compromise model effectiveness. Conversely, the frequency domain offers higher stability. For instance, when collecting signals indoors using universal software radio peripheral (USRP), the presence of walls and obstacles introduces multipath interference, which leads to the presence of strong high-frequency components in the received signal. However, significant variations in these components do not affect the identification of the modulation format used in the signal. Additionally, due to the numerous factors in real channel conditions, accurately estimating the variation in the characteristics of the received signals from different channels is a formidable task.

To address this challenge, transfer learning (TL) is introduced to AMC. This relaxes the previous constraint, allowing the use of knowledge from existing channel conditions to solve issues in unknown scenarios. Fine-tuning, an extensively used method, shows promising performance.