First arrival picking of microseismic signals based on nested U-Net and Wasserstein Generative Adversarial Network

• Published in: Journal of petroleum science & engineering 2020-12, Vol.195, p.107527, Article 107527
• Main Authors: Zhang, JingLan, Sheng, GuanQun
• Format: Article
• Language: English
• Subjects: First arrival picking; Low SNR environment; Microseismic; Nested U-Net wasserstein generative adversarial network; Residual blocks
• ISSN: 0920-4105; 1873-4715
• DOI: 10.1016/j.petrol.2020.107527

Picking the first arrival of microseismic signals, quickly and accurately, is the key for real-time data processing of microseismic monitoring. The traditional method cannot meet the high-accuracy and high-efficiency requirements for the firstarrival microseismic picking, in a low SNR environment. Concentrating on the problem of relatively low microseismic SNR, this paper proposes the Residual Link Nested U-Net Network (RLU-Net), which can not only retain the spatial position information of input signal and profile, but also realize the first arrival picking from end-to-end. The added number of layers with the residual block is important to extract more features and better distinguish the signal and noise in low SNR environments, which improves the accuracy and efficiency of first arrival picking. We also use the improved Wasserstein Generative Adversarial Network (WGAN) generated adversarial sample sets and conducted adversarial training on the network to enhance the generalization ability of the model. The resulting model then accurately and efficiently identified the first arrival of effective signals. The testing of forward modeling microseismic signals and measured microseismic data show that, compared with the Short Term Average/Long Term Average (STA/LTA) algorithm, the proposed algorithm not only has a high accuracy, but also can pick the first arrival time of microseismic signals in a low SNR environment. [Display omitted] •Designed the RLU-Net, which uses residual blocks to optimize the layer depth, making it easier to distinguish between signals and noise in low SNR environment.•The improved WGAN generates adversarial sample sets and effective sample sets.•Analyzed the impact of hyperparameter optimization and the number of training sets, on the performance of the network model.•The proposed method can quickly and accurately pick up the first arrival of microseismic signals in a low SNR environment.