An Integrated Autoencoder-Isolation Forest and Temporal Convolutional Network-Transformer Framework for Intelligent Operating-State Recognition and Anomaly Early Warning Using Multisource Drilling Time-Series Data
SPE Journal · 01.11.2026 · 1-24
DOI: 10.2118/236996-pa ↗Kurz gefasst
Fachartikel in der Zeitschrift SPE Journal, online erschienen am 11.11.2026. Verfasst von Huohai Yang und 6 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben.
Abstract (Originalsprache)
Safe and efficient drilling operations necessitate the timely early warning of abnormal events and the reliable recognition of drilling operating states. Nevertheless, the efficacy of conventional monitoring methods in complex downhole environments is often restricted by the presence of weak precursors of abnormal conditions, severe noise contamination, and strong fluctuations in field mud-logging data. This study presents an integrated framework for intelligent drilling operating-state recognition and anomaly early warning that utilizes multisource time-series data to overcome these obstacles. The framework integrates ensemble empirical mode decomposition (EEMD)-wavelet transform (WT) hybrid denoising, sparrow search algorithm (SSA)-based optimization, autoencoder (AE)-isolation forest (IF) anomaly detection, and temporal convolutional network (TCN)-Transformer-based operating-state classification. Random forest (RF) feature-importance analysis and Pearson correlation analysis were used to identify sensitive parameters and establish an operating-state feature system using drilling data from the Sichuan-Chongqing area. The AE-IF model identifies abnormal risks using reconstruction error and anomaly scores, while the denoising module suppresses noise interference and preserves critical dynamic features. The anomaly-detection model provided early-warning evidence for pipe sticking, earlier confirmation and continuous risk tracking for kick, and retrospective warning evidence for stick/slip vibration using downhole memory data. The model achieved an accuracy of 0.93, a recall of 0.90, an F1 score of 0.91, and an area under the receiver operating characteristic curve (AUC) of 0.93. The TCN-Transformer model achieved accuracies of 92.11% on historical data and 92.92% on an independent well for multiclass operating-state recognition. This performance surpassed that of TCN, Transformer, and long short-term memory (LSTM) models and demonstrated a strong generalization capability. In general, the proposed framework unifies signal processing, anomaly detection, and operating-state classification into an integrated workflow, thereby offering technical support for real-time or near-real-time monitoring of surface-parameter-based abnormalities, post-run diagnosis of vibration-related events, and future real-time decision support when low-rate transmitted downhole vibration descriptors are available.
Zitieren
APA
Yang, H., Zheng, H., He, L., Song, H., Zhang, M., Li, R., & Li, F. (2026). An Integrated Autoencoder-Isolation Forest and Temporal Convolutional Network-Transformer Framework for Intelligent Operating-State Recognition and Anomaly Early Warning Using Multisource Drilling Time-Series Data. SPE Journal, 1-24. https://doi.org/10.2118/236996-pa
Vancouver
Yang H, Zheng H, He L, Song H, Zhang M, Li R, et al. An Integrated Autoencoder-Isolation Forest and Temporal Convolutional Network-Transformer Framework for Intelligent Operating-State Recognition and Anomaly Early Warning Using Multisource Drilling Time-Series Data. SPE Journal. 2026:1-24. doi:10.2118/236996-pa
BibTeX
@article{Yang2026,
author = {Yang, Huohai and Zheng, Haoran and He, Lei and Song, Hao and Zhang, Mingzhou and Li, Renze and Li, Fuwei},
title = {An Integrated Autoencoder-Isolation Forest and Temporal Convolutional Network-Transformer Framework for Intelligent Operating-State Recognition and Anomaly Early Warning Using Multisource Drilling Time-Series Data},
journal = {SPE Journal},
year = {2026},
pages = {1--24},
doi = {10.2118/236996-pa}
}
EUVIMED