Deep learning-based low-SNR signal detection for passive space debris detection

The increasing amount of space debris poses a growing risk to satellite operations. While active radar-based detection provides high precision, it is constrained by high cost and frequency limitations. Passive detection using existing radio signals offers a more cost-effective alternative but faces difficulties in low Signal-to-Noise Ratio (SNR) conditions. This study proposes a deep learning-based non-image signal detection approach for passive space debris monitoring. The method employs convolutional neural network (CNN) trained on features derived from time-domain radio signals as well as frequency-domain representations, including power spectral density and magnitude spectrum, without relying on image datasets. Compared with the conventional power-threshold method, the proposed approach achieved a 4 dB improvement in the SNR required to reach a 90% detection probability. Furthermore, comparative experiments were conducted with anomaly detection methods such as One-Class SVM and Autoencoders. The results show that, although these techniques outperform the CNN-based approach when SNR exceeds 0 dB, the proposed method remains more effective under challenging low-SNR conditions. These findings demonstrate that deep learning can enhance the performance of passive detection while reducing computational costs.