Microseismic signals in complex geological environments are frequently contaminated by multiple non-stationary noise. Current denoising models often struggle to effectively balance noise suppression with accurate signal preservation, compromising the reliability of early warning systems. A conditional denoising U-Net (CDU-Net) model is proposed in this study to address this challenge. The model augments the U-Net backbone with a Transformer block to capture long-range temporal dependencies and a feature-wise linear modulation (FiLM) layer to adaptively adjust feature representations in different signal-to-noise ratio (SNR) environments. Microseismic monitoring data obtained through field measurements at the Hanjiang-to-Weihe River Diversion Project were employed to establish the research dataset. The denoising performances of five methods, wavelet transform (WT), variational mode decomposition (VMD), convolutional neural network (CNN), U-Net, and CDU-Net, are systematically compared from two perspectives: quantitative evaluation metrics and qualitative analysis. Significant advantages were demonstrated by the CDU-Net model across all quantitative evaluation indicators and waveform reconstruction quality, as shown by the experimental results. Particularly excellent performance was achieved, with the peak signal-to-noise ratio (PSNR) exceeding 30 dB and the structural similarity index measure (SSIM) reaching 0.82. At the same time, the denoising performance of the CDU-Net model was verified through visual analysis. Furthermore, ablation studies further confirm the complementary benefits of the U-Net backbone, Transformer, and FiLM components, reducing loss and synergistically improving output SNR. This study provides a solid theoretical foundation and reliable technical support for the long-term stable identification and high-precision early warning of microseismic signals in deep engineering.
Li et al. (2026) studied this question.