Suppressing Drilling Noise From Seismic Data Based on Multiscale Generator Network With Adaptive Feature Extraction

Abstract

In oilfield exploration and development, drilling noise creates significant interference, severely reducing the signal-to-noise ratio (SNR) of seismic data. Due to the complex characteristics of noise and signal, suppressing noise while recovering effective signals poses a challenge for denoising models. To address this, we propose a multiscale generator network with adaptive feature extraction for drilling noise suppression. The constructed network primarily consists of a dual-layer encoder-decoder structure. In the encoder, we designed an adaptive feature extraction module (AFEM) and a depthwise separable encoding module. The former utilizes deformable convolutions for adaptive extraction of effective features in seismic data, while the latter employs depthwise separable convolutions and an inverse bottleneck design to reduce computational complexity while maintaining effective feature extraction. The decoding module in the decoder uses two convolutional layers to reconstruct the seismic data with minimal computational cost. To prevent network degradation, residual connections are employed in both the encoding and decoding modules. The dual-layer structure extracts semantic information at different scales, preserving richer effective signal features and maximizing drilling noise suppression. Experimental results on both synthetic and field data demonstrate that the proposed method achieves higher-quality denoising results compared to denoising convolutional neural network (CNNS) (DnCNN), Unet, and MLGNet, while retaining the maximum amount of effective data.