Geoelectromagnetic data

As a deep exploration method based on natural field sources, magnetotelluric sounding signals are highly susceptible to various types of noise. Noise suppression is one of the key challenges in magnetotelluric data processing. Although traditional denoising methods can improve data quality to some extent, they often lack robustness when dealing with complex noise. To address this issue, this paper proposes a convolutional autoencoder model (ResAtt-CAE) that integrates multi-scale residual learning and a temporal attention mechanism. First, a training set for the model is constructed by adding square-wave noise, harmonic noise, impulse noise, step noise, triangular-wave noise, and white noise to clean magnetotelluric data. During the dimensionality reduction and reconstruction processes of the convolutional autoencoder, complex features of the signals are captured to achieve signal fitting and signal-to-noise discrimination. By constructing a dilated convolutional residual module, the model captures electromagnetic response features across multiple time scales, while adaptive attention weights are applied to enhance the representation of effective signal segments. Experimental results demonstrate that after processing synthetic noisy data, the ResAtt-CAE model achieves a coherence coefficient (CORC) of 0.9864, a signal-to-noise ratio (SNR) of 16.161, and a normalized root mean square error (NRMSE) of 0.0224. In comparison, the traditional autoencoder yields a CORC of 0.5442, an SNR of 3.297, and an NRMSE of 0.3175. For field data, comparative analyses from multiple perspectives\u2014power spectral density (PSD), polarization direction, and apparent resistivity-phase curves\u2014demonstrate that the ResAtt-CAE method outperforms traditional autoencoders. The proposed method exhibits significant noise suppression effects on common types of magnetotelluric noise, effectively improves data processing quality, and provides reliable data support for deep inversion.