Physics-guided self-supervised learning for monochromatic noise removal

Abstract

Monochromatic noise, such as the electromagnetic interference from power lines, is a well-known important problem in seismic signal processing. Removal of this contamination is challenging due to the time-varying nature of the fundamental frequency of monochromatic noise. We propose a physicsguided self-supervised learning method to restore the overkilled notch frequency components and improve the denoising effect facing power line interference with the time-varying fundamental frequency. This physics-guided learning method is designed to decompose a seismic record into power line noise and signal by a convolutional encoder-decoder network consisting of two decoders, which outputs the noise and signal components, respectively. The criterion of this unsupervised separation is formulated in both frequency and time domains. The former minimizes the trust-spectrum residual of the predicted signal with a sparsity-promoting constraint. The latter minimizes two residuals: First, the residual between the input seismic and the summation of the two sub-network outputs. Second, the residual between the output signal and the degradations of conventional filter results. The test results on the BP2004 seismic signal with synthetic time-varying frequency monochromatic noise demonstrate that the proposed method is computationally efficient. More importantly, the signal spectrum that is overly suppressed by the conventional notch filter can be restored to improve the denoising effect by the proposed method.