EUVIMEDEUVIMED entdecken
Crossref

A Dual-Branch Neural Network for Classification of Distributed Acoustic Sensing-Based Acoustic Responses during Hydraulic Fracturing

Guanghao Du, Weibo Sui, Huan Guo, Jiayi Song

SPE Journal · 01.11.2026 · 1-18

DOI: 10.2118/236994-pa ↗

Kurz gefasst

Fachartikel in der Zeitschrift SPE Journal, online erschienen am 17.11.2026. Verfasst von Guanghao Du und 3 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Real-time interpretation of distributed acoustic sensing (DAS) data is crucial for diagnosing interstage communication and optimizing stimulation strategies during hydraulic fracturing. However, downhole responses associated with fluid flow, mechanical vibration, and fracture activity overlap, making it difficult for engineers to distinguish different types of fluid-entry responses in massive DAS waterfall plots. To address this challenge, we develop the dual-branch transformer-convolutional neural network (CNN) fusion network (DTCF-Net) that automatically classifies strong fluid-entry responses within the active stage, weak or off-target responses, and background signals. DTCF-Net processes 3.2-second windows with 50% overlap. A temporal transformer branch extracts transient variations and long-range dependencies from raw 1D waveforms, while an image inception branch learns time-frequency (TF) and structural features from six 2D representations. We prevent temporal information leakage using a time-block-gap splitting strategy and evaluate the model through 12 ablation experiments, computational efficiency analysis, and validation on continuous field DAS data. DTCF-Full achieves an accuracy of 97.76% and a macro-F1 score of 0.9552, outperforming both single-branch models. Removing the two most computationally expensive 2D representations yields DTCF-Efficient, which achieves an accuracy of 97.33% and a macro-F1 score of 0.9478 while reducing the feature-generation time per window from 7.45 seconds to 0.89 seconds. On the field data set, its end-to-end processing time is 1.37 seconds per window. Field results show that the model locates strong responses near perforation clusters and identifies sustained weak responses at nonperforated locations and outside the active stage. These predictions agree with low-amplitude DAS signatures, treatment records, and independent offset-well low-frequency DAS (LF-DAS) strain-rate measurements. By revealing the time-depth distribution of strong and weak fluid-entry responses, DTCF-Net supports the near-real-time diagnosis of fluid entry and interstage communication, provides timely evidence for evaluating fracturing-treatment performance, and informs subsequent treatment adjustments aimed at improving fluid distribution and stimulation effectiveness.

Zitieren

APA

Du, G., Sui, W., Guo, H., & Song, J. (2026). A Dual-Branch Neural Network for Classification of Distributed Acoustic Sensing-Based Acoustic Responses during Hydraulic Fracturing. SPE Journal, 1-18. https://doi.org/10.2118/236994-pa

Vancouver

Du G, Sui W, Guo H, Song J. A Dual-Branch Neural Network for Classification of Distributed Acoustic Sensing-Based Acoustic Responses during Hydraulic Fracturing. SPE Journal. 2026:1-18. doi:10.2118/236994-pa

BibTeX
@article{Du2026,
  author = {Du, Guanghao and Sui, Weibo and Guo, Huan and Song, Jiayi},
  title = {A Dual-Branch Neural Network for Classification of Distributed Acoustic Sensing-Based Acoustic Responses during Hydraulic Fracturing},
  journal = {SPE Journal},
  year = {2026},
  pages = {1--18},
  doi = {10.2118/236994-pa}
}
Kostenlos starten →