Lightweight Asymmetric Convolutional Networks with Grid Constellation Matrices for Hardware-Efficient Automatic Modulation Classification in Harsh Environments
Lightweight Asymmetric Convolutional Networks with Grid Constellation Matrices for Hardware-Efficient Automatic Modulation Classification in Harsh Environments
Moein Sefidi1
1) Amirkabir University of Technology (Tehran Polytechnic)
IOI Code
:
ABGG-ACZFBK
Publication :
11th.International Conference on Researches in Science & Engineering & 8th.International Congress on Civil, Architecture and Urbanism in Asia(11icrsie.ir)
Abstract :
Automatic Modulation Classification (AMC) plays a pivotal role in intelligent spectrum management and interference mitigation in next-generation cognitive radio networks and dynamic spectrum access networks. Although deep learning models have demonstrated superior classification capabilities compared to traditional feature extraction-based methods, their deployment on resource-constrained edge devices faces significant limitations due to high computational complexity and vulnerability to severe channel disturbances. In this paper, a framework named Lightweight Asymmetric Convolutional Networks with Network Constellation Matrices (LACN-GCM) is proposed, which is designed to optimize hardware utilization while maintaining accuracy. To create a robust representation of the data, the raw sequences of in-phase and orthogonal (I/Q) components are transformed into Grid Constellation Matrices (GCMs). These matrices effectively capture the underlying 2D probability distributions of the modulation symbols while naturally filtering out the added noise. Next, an asymmetric convolutional architecture is developed using 3×1 and 1×3 decomposed kernels, along with Skip Connections. The goal of this architecture is to extract deep spatiotemporal features by reducing the number of parameters. Simulations performed on benchmark datasets show that the proposed LACN-GCM framework achieves an overall classification accuracy of over 95% under a challenging set of modulation formats with a signal-to-noise ratio (SNR) of 10 dB. Importantly, the proposed architecture reduces computational complexity compared to standard residual networks and is therefore suitable for real-time deployment on edge devices in noisy communication environments.
Keywords : automatic modulation classification grid based formulation matrices asymmetric convolution deep learning cognitive radio