Robust Computationally-Efficient Wireless Emitter Classification Using Autoencoders and Convolutional Neural...

Area of Science:

• Wireless Communications
• Machine Learning
• Signal Processing

Background:

• Intelligent spectrum monitoring is vital for emerging wireless networks with multiple radio-access technologies (RATs).
• Existing deep learning methods like Convolutional Neural Networks (CNNs) for emitter classification struggle with accuracy degradation and high computational costs in harsh channel conditions.

Purpose of the Study:

• To develop a novel Deep Learning (DL)-based approach for accurate and computationally efficient radio-access technology (RAT) classification.
• To improve emitter classification performance under harsh propagation conditions, such as non-line-of-sight, fading, and Doppler shifts.

Main Methods:

• A Denoising Autoencoder (DAE) was employed to preprocess channel-corrupted spectrograms, reducing dimensionality and noise.
• The denoised representations were then fed into a CNN classifier for RAT identification.
• Simulations included various RATs like LTE and Wi-Fi standards under diverse, challenging channel conditions.

Main Results:

• The proposed DAE-CNN approach significantly outperformed standalone CNNs and other machine learning models in harsh channel conditions.
• The method achieved a maximum classification accuracy of 100% and an average accuracy of 91% across all tested propagation scenarios.
• The solution demonstrated improved computational efficiency compared to existing methods.

Conclusions:

• The novel DAE-CNN approach offers a robust and efficient solution for radio-access technology classification in challenging wireless environments.
• This method is suitable for resource-constrained network edge scenarios, enabling effective intelligent spectrum monitoring.
• The findings contribute to optimizing spectral utilization, interference management, and regulatory enforcement in future wireless networks.