Efficient SNR enhancement model for severely contaminated DAS seismic data based on heterogeneous knowledge distillation

Distributed acoustic sensing (DAS) is an emerging seismic acquisition technique with great practical potential. However, various types of noise seriously corrupt DAS signals, making it difficult to recover signals, particularly in low SNR regions. Existing deep learning methods address this challenge by augmenting datasets or strengthening the complex architecture, which can cause over-denoising and a computational power burden. Hence, we propose the heterogeneous knowledge distillation (HKD) method to more efficiently address the signal reconstruction under low SNR. HKD employs ResNet 20 as the teacher and student model (T-S). It utilizes residual learning and skip connections to facilitate feature representation at deeper levels. The main contribution is the training of the T-S framework with different noise levels. The teacher model that was trained using slightly noisy data serves as a powerful feature extractor to capture more accurate signal features, since high quality data is easy to recover. By minimizing the difference between the outputs of T-S models, the student that was trained using severely noisy data can distill the absent signal features from the teacher to improve its own signal recovery, which enables heterogeneous feature distillation. Furthermore, simultaneous learning of negative and positive components (PNL) has been proposed to extract more useful features from the teacher, enabling the T-S framework to learn from both the predicted signal and noise during training. Consequently, a new loss function that combines student denoising loss and HKD loss weighted by PNL was developed to alleviate signal leakage. The experimental results demonstrate that the HKD achieves distinct and consistent signal recovery without increasing computational costs.