An unsupervised intelligent warning model for drilling kick risk based on multi-temporal feature coupling

Abstract

As oil and gas exploration continues to progress into deeper and unconventional reservoirs, the likelihood of kick risk increases, making kick warning a critical factor in ensuring drilling safety and efficiency. Due to the scarcity of kick samples, traditional supervised models perform poorly, and significant fluctuations in field data lead to high false alarm rates. This study proposes an unsupervised graph autoencoder (GAE)-based kick warning method, which effectively reduces false alarms by eliminating the influence of field engineer operations and incorporating real-time model updates. The method utilizes the GAE model to process time-series data during drilling, accurately identifying kick risk while overcoming challenges related to small sample sizes and missing features. To further reduce false alarms, the weighted dynamic time warping (WDTW) algorithm is introduced to identify fluctuations in logging data caused by field engineer operations during drilling, with real-time updates applied to prevent normal conditions from being misclassified as kick risk. Experimental results show that the GAE-based kick warning method achieves an accuracy of 92.7% and significantly reduces the false alarm rate. The GAE model continues to operate effectively even under conditions of missing features and issues kick warnings 4 min earlier than field engineers, demonstrating its high sensitivity and robustness. After integrating the WDTW algorithm and real-time updates, the false alarm rate is reduced from 17.3% to 5.6%, further improving the accuracy of kick warnings. The proposed method provides an efficient and reliable approach for kick warning in drilling operations, offering strong practical value and technical support for the intelligent management of future drilling operations.