Applications of Modelling and Simulation (Jan 2024)
Lightweight Multi-Channel Gated Recurrent Deep Neural Network for Automatic Modulation Recognition in Spatial Cognitive Radio
Abstract
Automatic modulation recognition (AMR) is a promising technology for intelligent communication receivers to detect signal modulation schemes. Recently, the emerging deep learning (DL) research has facilitated high-performance DL-AMR approaches. This research presents a novel and versatile Multi-Channel Gated Recurrent Deep Neural Network framework (MCGDNN) designed to tackle the intricate challenges of automatic modulation recognition. MCGDNN integrates two dedicated Deep Learning Networks (DLNs) to address specific signal types: one DLN specializes in classifying In-phase Quadrature (IQ) signals, overcoming limited training data with data augmentation and model optimization through pruning by Differentiable Annealing Indicator Search, resulting in a streamlined, lightweight model. The other DLN focuses on Frequency-Domain Amplitude-Phase signals, leveraging a modified Fast Fourier Transform (FFT) with data normalization which avoids the numerical distance between different features for enhancing feature extraction. Additionally, it introduces the Adaptive Moment Estimation Maximum (Adamax) Bi-directional Gated Recurrent Unit (Optimized BiGRU3) network that accurately extracts amplitude and phase spectrum features within the frequency domain. Furthermore, the research presents an innovative approach to signal classification by introducing a modified FFT technique for the extraction of amplitude and phase feature information from Amplitude Modulated-Double Sideband and Wideband Frequency Modulation signals in the frequency domain. This development culminates in the creation of a two-class dataset named DW, based on these amplitude and phase characteristics. In summary, this research signifies a significant stride in the field of AMR, offering a comprehensive framework (MCGDNN) capable of handling diverse signal types, an optimized feature extraction network (BiGRU3), and a novel dataset (DW) with enhanced classification accuracy. These advancements hold immense promise for applications in modern communication systems and signal processing.
