Unsupervised Learning–Based Kick Risk Warning with Dynamic Thresholds and Knowledge Embedding
Authors:
- Zhaopeng ZhuZZZhaopeng Zhu
- Gensheng LiGLGensheng Li
- 28 August 2025
- journal article
- Published by Society of Petroleum Engineers (SPE) in SPE Journal
- Vol. 30 (11) , 6584-6602
Abstract
Summary: Kick risk is one of the most common and safety-critical issues in drilling operations, posing substantial threats to operational safety. Therefore, accurate kick risk warning is essential for ensuring drilling safety and operational efficiency. However, traditional data-driven models are often affected by the quality and variability of real-time logging data and lack integration with formation characteristics and domain knowledge, leading to high false alarm rates, instability, and limited effectiveness in kick risk warning. To address these challenges, we introduce two key innovations—a dynamically adjusted threshold mechanism, informed by predrilling kick risk assessment, and knowledge embedding to enhance the performance of data-driven models. The dynamic threshold mechanism enables the model to adaptively adjust its warning criteria based on predrilling assessments, reducing the false alarm rate. Knowledge embedding further optimizes the model by incorporating domain knowledge, improving the model’s accuracy and reliability. In this study, we propose an unsupervised temporal kick risk warning method based on bidirectional gated recurrent unit (BiGRU) autoencoder (AE)(BiGRU-AE), which leverages the temporal modeling capabilities of BiGRU to extract key features from the logging data. The proposed method is validated through a comprehensive evaluation using field drilling data sets, comparing temporal vs. nontemporal models, as well as supervised vs. unsupervised approaches, and assessing model performance before and after dynamic threshold optimization. The unsupervised model is trained using risk-free data, eliminating the need for labeled kick risk samples and effectively addressing the challenge of limited risk sample availability. The results show that BiGRU-AE outperforms other models, with the combined optimization of dynamic threshold and knowledge embedding reducing the false alarm rate by 14.67%, significantly enhancing stability and reliability. The proposed method organically integrates domain knowledge with intelligent technology, greatly improving the accuracy and practicality of kick risk warning. This provides strong technical support for ensuring drilling safety and demonstrates broad engineering application potential in the field of intelligent drilling.Keywords
This publication has 21 references indexed in Scilit:
- Dynamic risk assessment of deep-water dual gradient drilling with SMD system using an uncertain DBN-based comprehensive methodOcean Engineering, 2021
- Risk analysis of well blowout scenarios during managed pressure drilling operationJournal of Petroleum Science and Engineering, 2019
- Overflow remote warning using improved fuzzy c-means clustering in IoT monitoring system based on multi-access edge computingNeural Computing and Applications, 2019
- Intelligent early warning model of early-stage overflow based on dynamic clusteringCluster Computing, 2017
- An early intelligent diagnosis model for drilling overflow based on GA–BP algorithmCluster Computing, 2017
- A DBN-based risk assessment model for prediction and diagnosis of offshore drilling incidentsJournal of Natural Gas Science and Engineering, 2016
- Probability distributions conditioned by the available information: Gamma distribution and momentsComputers & Mathematics with Applications, 2006
- Probabilistic Structural Mechanics Handbook1995
- On the stability of the geostatistical methodMathematical Geology, 1992
- Theoretical Soil Mechanics1943