Audio magnetotelluric signal-noise recognition and separation based on RepVGG-DnCNN

The audio magnetotelluric (AMT) signals have a wide frequency bandwidth and are easily affected by various cultural noises. To improve the accuracy of signal-noise recognition and denoising, we propose a novel deep learning-based noise suppression method, RepVGG-DnCNN. First, an improved RepVGG network is used to distinguish between noisy and high-quality segments. Then, a denoising convolutional neural network (DnCNN) is employed to denoise the identified noisy segments. Finally, the preserved high-quality segments and denoised segments are reconstructed to form the complete AMT signal. Experimental results show that the improved RepVGG outperforms existing convolutional neural network (CNN) and temporal convolutional network (TCN) in terms of accuracy, precision, recall, and F1-score. DnCNN cleverly leverages residual learning mechanism and batch normalization (BN) module to learn the relationship between noise and noise-free signals from the training data, effectively mapping noisy signals to high-quality signals. In addition, the apparent resistivity and phase curves obtained by our method have also been significantly improved.