Abstract
The deep convolutional neural network has strong representative ability, which can learn latent information repeatedly from signal samples and improve the accuracy of automatic modulation classification (AMC). In this paper, a novel compressive convolutional neural network (CCNN) is proposed for AMC, where different constellation images, i.e., regular constellation images (RCs) and contrast enhanced grid constellation images (CGCs), are generated as network inputs from received signals. Moreover, a compressive loss constraint is proposed to train the CCNN, which aims at capturing high-dimensional features for modulation classification. Additionally, CCNN utilizes intra-class compactness and inter-class separability to enhance the classification and robustness performance for the different orders of modulations. The simulation results demonstrate that CCNN displays superior classification and robustness performance than existing AMC methods.
Introduction
In the era of 5G and dynamic spectrum access, the radio environment is increasingly crowded and "disordered." To manage this, systems need Automatic Modulation Classification (AMC)—the ability to identify the modulation format (e.g., QPSK, 16-QAM) of a received signal without prior information. Traditional methods usually fall into two camps: Likelihood-based (LB) methods, which are mathematically optimal but computationally expensive, and Feature-based (FB) methods, which are faster but rely on "hand-crafted" features that often miss subtle patterns.
This paper introduces a Compressive Convolutional Neural Network (CCNN) that significantly boosts the accuracy of AMC in complex wireless environments. By combining standard scatter-plot images with a novel "contrast-enhanced grid" representation and a specialized "compressive loss" function, the model achieves superior classification performance and robustness against noise and channel interference.
Methodology
The proposed CCNN architecture consists of three main stages: an Extraction Module, a Fusion Module, and a Classification Module.
The core innovations lie in the input data representation and the training objective:
Dual-Image Input
Instead of raw signal samples, the network processes two types of images:
- Regular Constellation (RC): A standard 2D scatter plot of the signal's real and imaginary parts.
- Contrast Enhanced Grid Constellation (CGC): The image is divided into a grid, and the density of signal points in each cell is calculated. A contrast enhancement rule is applied to make dense regions more prominent, helping the network perceive the underlying probability distribution of the signal.
Compressive Loss ($L_c$)
Standard networks typically use Softmax Loss, which focuses on classification accuracy but may not optimize the internal feature space effectively. The authors introduce a Compressive Loss that enforces:
- Intra-class Compactness: Pulling features of the same modulation type closer to a central point.
- Inter-class Separability: Pushing the centers of different modulation types further apart.
This geometric approach "compresses" the clusters, making the boundaries between modulation types much clearer and enhancing robustness.
Experimental Setup
The model was evaluated using simulated signals across five common modulation formats: BPSK, QPSK, 8PSK, 16-QAM, and 64-QAM. Signals were passed through a Rayleigh fading channel with Additive White Gaussian Noise (AWGN). The dataset included 40,000 training images per modulation type, with testing performed across various Signal-to-Noise Ratios (SNRs) ranging from -5 to 15 dB and varying sample lengths.
Key Results and Insights
- Accuracy Gains: CCNN achieved a 1.5 dB to 2 dB gain over RNN, DNN, and standard CNN models at a 95% correct classification threshold.
- Low SNR Performance: The use of multiple constellation images provided a 0.5 dB gain over using RCs alone at 90% accuracy, specifically aiding performance in noisy environments.
- Efficiency: Despite the dual-image input, CCNN training was faster (720s) than RNNs (810s) or standard DNNs (900s) on a GTX 1080 GPU, demonstrating that effective feature representation can be more efficient than increasing model depth.
- Robustness: The model remained stable under phase offsets and handled frequency offsets well up to a normalized value of $10^{-4}$.
Conclusion
The CCNN approach demonstrates that density-based features and geometric optimization of the feature space significantly improve AMC performance. While the model shows sensitivity to high frequency offsets and introduces some preprocessing overhead, its superior accuracy and robustness make it a promising candidate for next-generation wireless systems. Future work will focus on compensating for high-frequency Doppler shifts and validating the model with real-world signal data.
Summary updated