Enhancing Modulation Classification Through Lightweight Dyadic down-Sampling Schemes and CNN Layer Fusion

Automatic Modulation Classification (AMC) plays a crucial role in dynamic spectrum management and interference avoidance. It is pertinent to both civilian applications, where it enhances bandwidth and link quality, and military applications, including Electronic Warfare (EW). Furthermore, AMC is a critical component of advanced radio systems like Cognitive Radios (CR). Indeed, recognizing digital modulation schemes without prior knowledge is essential for the advancement of cognitive radios. The primary aim of this paper is to introduce, for the first time, an approach to Automatic Modulation Recognition (AMR) that combines dyadic down-sampling decomposition with an Artificial Intelligence (AI) architecture, utilizing the fusion of Convolutional Neural Network (CNN) layers. The proposed lightweight methodology enhances the modulation recognition rate without imposing significant pre-processing constraints on the overall architecture as shown by a complexity study. This paper emphasizes the impact of dyadic down-sampling on IQ-based signals in comparison to other classification methods and examines its implications for CNN-based modulation classification. The influence of various parameters is analyzed using AI hypermodels. Additionally, the modularity of the applied AI fusion architecture allows for the exploration of straightforward explainable AI (XAI) concepts, as the impact of each dyadic scale on the decision of classification is visible. For this analysis, the publicly accessible "Dataset for the Machine-Learning Challenge [CSPB.ML.2018]" is used. In this paper, we report an average absolute increase of 6% in classification accuracy across all SNR levels in the database, with a peak improvement of 8.3% observed at an SNR of 4.5 dB.