Abstract: The microseismic data containing diverse noises with extremely low signal-to-noise ratio cannot be effectively processed using traditional denoising methods. This paper proposes a deep convolutional auto-encoder network with optimized loss function constraints fused with residual attention (RADNet). The proposed method employs a deep convolutional auto-encoder structure for local feature extraction from noisy data and fuses global features to assign weights to different features using the attention mechanism. The optimized loss function is used to guide network training, and denoised signals are finally reconstructed based on the residual network. The application of RADNet and other denoising methods to simulated and real data demonstrates that RADNet improves peak signal-to-noise ratio (PSNR) by 2.783 and 8.099 dB and structural similarity (SSIM) by 0.031 and 0.065, respectively, compared to denoising convolutional neural network (DnCNN) and deep convolutional auto-encoder networks. Furthermore, RADNet decreases mean square error (MSE) and better preserves effective signals and texture details in microseismic waveform.