Abstract
Seismic fault identification remains critical for resource exploration and geohazard prevention, yet conventional methods suffer from subjective interpretation bias and computational inefficiency. While convolutional neural networks (CNNs) enhance automation, their neglect of multiscale frequency features limits accuracy. Here, propose a novel Wavelet-Convolutional Neural Network (W-CNN) and its variants (W-CNN R1, W-CNN R2 and W-CNN R3) that architecturally fuses discrete wavelet transforms (DWT) with CNNs, establishing a spatial-frequency learning paradigm. By embedding Haar wavelet filter banks with cross-scale residual connections, W-CNN achieves explicit decoupling of high-frequency fault details from low-frequency structural contexts, reducing parameters by 21% versus conventional CNNs. Evaluated on coal mine datasets, W-CNN R3 achieves 90.0% accuracy (F1-score 90.3%), surpassing mainstream CNNs (LeNet-5, AlexNet, VGG16) by 0.6-12.3%, with the highest recall (95.5%) and faster convergence. The model successfully resolves 30 out of 32 exposed complex micro-faults (93.8% detection rate), demonstrating strong consistency with roadway-exposed faults in geologically complex zones, which significantly enhances its predictive capability for small-scale discontinuities. The frequency selection mechanism effectively suppresses noise interference, while the optimized architecture enables orders-of-magnitude acceleration in 3D processing. This framework provides an extensible solution for intelligent geological interpretation, with critical applications in mine safety monitoring.