Ablation Study: Evaluating the Contributions of Network Components for Signal Recovery

In this section, we conducted an experiment to analyze the impact of each module in our proposed network architecture. We designed four different network architectures to investigate the effects of the main network components, including the' attention mechanism', 'residual blocks', and 'feature extraction modules'. Firstly, we removed the' attention mechanism' and 'feature extraction modules' to check the effectiveness of the multi-scale strategy. Secondly, we deleted the modified 'residual blocks' from the network. Thirdly, we replaced the' attention mechanism' with common convolutional layers. Finally, we removed the 'feature extraction module' and replaced it with an' attention mechanism' to analyze its role in denoising ability. All the networks were trained with the same hyper-parameters and dataset, and we evaluated their performance on both synthetic and field records.

The results indicated that the modified networks were able to reconstruct the desired signals affected by noise to some extent, but they still had some limitations in the continuity and smoothness of the recovered signals, especially for the weak signals. The comparison results in terms of signal-to-noise ratio and computational cost showed that our proposed network architecture achieved the best performance, with an approximate 30dB improvement over other networks. The experimental results demonstrated that the network components were irreplaceable and helpful in signal recovery.