Identification of Weak Microseismic Signals in Hydraulic Fracturing Using 3D-Unet With ASPP Feature Fusion

Microseismic monitoring is crucial for guiding hydraulic fracturing in the development of unconventional oil and gas resources, but surface monitoring often captures low-energy signals with low signal-to-noise ratios (SNRs). Traditional 2-D convolutional neural networks lack the ability to enforce frequency constraints on microseismic signals in the time domain and are unable to capture essential time-frequency (T–F) correlations across different traces in the T–F domain, resulting in substantial limitations for the precise detection of weak microseismic signals. To address these limitations and to enhance signal identification and noise suppression, we propose a novel 3D-Unet deep neural network that employs 3-D atrous spatial pyramid pooling (ASPP) for comprehensive feature fusion. This 3-D approach comprehensively considers the T–F correlations inherent in multitrace microseismic data, effectively leveraging the richer feature representations within the 3-D space. We transform multitrace microseismic time-domain data into 3-D T–F domain using the S-transform (ST). The network takes the noisy T–F data as input and outputs a denoised T–F volume. Initially, the network model is pretrained on synthetic data, followed by fine-tuning on field microseismic data through transfer learning. The new network is evaluated using both synthetic and field data, and the results demonstrate effective identification of weak microseismic signals under low-SNR conditions, offering significant improvements for practical applications.