A Multi-Feature Time-Series Data Generation Method for LEO Electromagnetic Monitoring Systems

Low Earth Orbit (LEO) electromagnetic monitoring systems face challenges in signal acquisition and data scarcity in typical scenarios. To address this issue, we propose an electromag-netic signal generation model based on Generative Adversarial Networks (GANs). This model synergistically combines UNet, SENet, and BiLSTM architectures, approaching multi-feature data as a multitask learning challenge. By leveraging a mask vector to regulate the training process, it enables the parallel generation of high-fidelity electromagnetic signal data, ensuring both efficiency and precision in data synthesis. We validated the model's effectiveness by simulating LEO satellite signals reception in the constructed virtual environment. Compared with other conditional GAN architectures for both generators and discriminators, the proposed model exhibits superior performance in both qualitative and quantitative analyses. Experimental results demonstrate that the signals generated by our model outperform those of existing models in terms of similarity and fidelity.