Abstract
Subsampling of received wireless signals is important for relaxing hardware requirements as well as the computational cost of signal processing algorithms that rely on the output samples. We propose a subsampling technique to facilitate the use of deep learning for automatic modulation classification in wireless communication systems. Unlike traditional approaches that rely on pre-designed strategies that are solely based on expert knowledge, the proposed data-driven subsampling strategy employs deep neural network architectures to simulate the effect of removing candidate combinations of samples from each training input vector, in a manner inspired by how wrapper feature selection models work. The subsampled data is then processed by another deep learning classifier that recognizes each of the considered 10 modulation types. We show that the proposed subsampling strategy not only introduces drastic reduction in the classifier training time, but can also improve the classification accuracy for the considered dataset. An important feature herein is exploiting the transferability property of deep neural networks to avoid retraining the wrapper models and obtain superior performance through an ensemble of wrappers over that possible through solely relying on any one of them.
Introduction
Automatic Modulation Classification (AMC) is a critical component in modern wireless communications, enabling Software Defined Radios (SDR) to identify signal types for efficient spectrum management and electronic warfare. Despite the success of deep learning in this domain, three primary challenges persist: high hardware costs associated with high-rate sampling, significant computational overhead for training on massive datasets, and the "confusion" problem where models struggle to distinguish between similar modulations like QAM16 and QAM64.
Methodology
This research introduces a data-driven subsampling strategy inspired by wrapper feature selection. Rather than relying on fixed filters, the system utilizes neural networks to identify the most informative signal samples.
The Ensemble Approach
The methodology leverages an ensemble of three distinct deep learning architectures: a Convolutional Neural Network (CNN), a Convolutional Long Short-term Deep Neural Network (CLDNN), and a Residual Network (ResNet). Each architecture captures unique signal features, ensuring a robust selection process.
Subsampler Nets and Ranking
For each architecture, a "Ranker Model" is established. The system systematically sets individual samples to zero and measures the resulting drop in classification accuracy. Samples that cause the most significant performance degradation are prioritized. To handle dependencies between samples, a modified $\epsilon$-greedy tree search explores optimal sample combinations rather than simple top-$k$ selection.
Transfer Learning
To maintain efficiency, the authors exploit the transferability of deep learning features. Pre-trained models are used to rank samples, which avoids the computationally expensive requirement of retraining the wrapper models during the selection phase.
Experimental Setup
The proposed method was evaluated using the RadioML2016.10b dataset, which includes 160,000 vectors across 10 modulation types (8 digital, 2 analog). Each vector consists of 128 complex (I/Q) samples captured at approximately 6 times the Nyquist rate, with SNR ranging from -20 dB to +18 dB.
Results and Discussion
Accuracy and Regularization
Counter-intuitively, subsampling was found to improve classification accuracy. At high SNR (18 dB), the proposed method achieved over 99% accuracy, significantly outperforming models trained on the full dataset (~91-92%). This suggests that projecting signals onto a lower-dimensional subspace acts as a form of regularization, clarifying decision boundaries and resolving confusion between similar modulation schemes.
Efficiency and Sub-Nyquist Performance
The technique demonstrated remarkable efficiency; a subsampling rate of 1/16 (using only 8 of 128 samples) reduced training time by two-thirds while maintaining 92.6% accuracy. Furthermore, the model remained effective at sampling rates as low as 37.5% of the Nyquist rate, indicating that hardware requirements for AMC can be substantially relaxed without sacrificing performance.
Conclusion
Ensemble Wrapper Subsampling provides a powerful framework for reducing the data requirements of deep learning-based AMC. While the method shows sensitivity to SNR and involves initial computational overhead for sample ranking, its ability to improve accuracy while slashing training time makes it a promising candidate for real-time SDR applications. Future work will focus on optimizing the selection process for low-power devices and extending these data-driven subsampling insights to other wireless signal processing tasks.
Summary updated