ПОКРАЩЕННЯ КЛАСИФІКАЦІЇ БПЛА ЗА РАДІОЧАСТОТНИМИ СИГНАЛАМИ ЗА ДОПОМОГОЮ ГІБРИДНОЇ ЗГОРТКОВО-РЕКУРЕНТНОЇ АРХІТЕКТУРИ

Abstract

This research aims to automate the classification of small UAVs using radio frequency signals. The object of this research is to classify UAVs by radio frequency signature using combined machine learning models. It focuses on a combination of machine learning-based models to achieve better classification results when efficient learning and high accuracy are difficult to achieve due to temporal dependency in complex signals. This study aims to enhance UAV classification models by optimizing a convolutional neural network architecture with the addition of recurrent layers. Methods described in this research propose a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model that combines convolutional and recurrent layers to extract spatially and temporally dependent features from UAV RF signals. Compared to traditional CNNs, the Convolutional Long-Short-Term Deep Neural Network (CLDNN) model reduces false positives across classes and improves classification performance (recall) for closely related or noisy signal classes. However, the complexity and computational costs of the CLDNN model are acknowledged, underscoring challenges for systems with limited processing power. The tasks of this research are as follows: 1) to review the current state-of-the-art approaches in the classification of small UAVs; 2) to outline the problem statement; 3) to create a test dataset for the proposed drone models; 4) to adapt CLDNN model for drone classification tasks; 5) to provide a comparison between CNN baseline and the proposed model. Conclusions. The experimental results show that the proposed CNN–LSTM model significantly improves classification accuracy compared with the baseline CNN architecture, especially for classes prone to misclassification. The integration of convolutional layers, which are effective at extracting local features, and recurrent LSTM layers, which are effective at modeling sequential dependencies, enables the combined architecture to reduce false positives. Accordingly, this hybrid structure is fairly effective for signal classification problems because it captures spatially localized features and temporal contextual information within the input. This study confirms that the CNN-LSTM model is efficient at classifying UAV radio-frequency signals and has great potential for practical applications. Possible future research directions include optimizing the model's computational efficiency by handling greater interference and noisier data.