Abstract: To address the challenges of low signal-to-noise ratio and difficulty in phase picking of microseismic signals induced by coal fracturing during hydraulic fracturing in coal seams, a deep neural network model based on a masking strategy is proposed (Mask Residual Attention Denoiser, termed MRAD). The model is built upon the classical U-Net architecture, incorporating a mask-guided strategy to enable the network to learn the feature distributions of effective microseismic signals and noise separately. It outputs corresponding signal and noise masks, and applies the effective signal mask to the time-frequency matrix of the raw microseismic signal via Hadamard product to achieve mask-weighted filtering for denoising. A total of 18 670 microseismic training samples were constructed by iterative superposition of manually labeled clean signals and random noise. Structurally, residual blocks replace standard convolution and downsampling layers in U-Net to mitigate gradient vanishing and prevent local minima; spatial attention mechanisms are further introduced in skip connections to enhance the network's focus on valid signal regions. Experimental results show that, after denoising with MRAD, the average S_NR of the test signals improves to 18.22 dB, the root mean square error is reduced to 0.042 4, the normalized cross-correlation reaches 0.969 9, and the energy ratio is 1.028 6. The denoising performance is particularly significant for signals with original S_NR in the range of 0～10 dB. Moreover, the model processes a single microseismic signal in under 30 ms, with low computational demands, meeting the requirements for real-time microseismic monitoring and data processing in hydraulic fracturing operations. To validate the effectiveness of MRAD, 30 synthetic mixed signals were denoised using both MRAD and traditional methods. Comparative results demonstrate that MRAD outperforms in terms of signal quality improvement and distortion reduction. Field application in a coal mine in Ningxia, China, shows that after denoising the microseismic signals from three fractured boreholes, the S_NR was concentrated in the range of 10～25 dB, with an average increase of 6.90 dB. Noise suppression was effective, leading to improved P-wave arrival picking accuracy, and the number of detected microseismic events increased from 487 to 653, a growth of approximately 1.34 times. Source localization analysis of the denoised signals indicates that the unidirectional extent of induced fractures ranges from 12 to 37 meters, aligning well with the construction parameters of the borehole fracturing sections. These results confirm that the proposed method offers strong technical support for real-time monitoring and effectiveness evaluation of microseismic signals in coal seam hydraulic fracturing.