Anomaly detection of drilling overflow using Shapelet-aware Graph Temporal Network

Drilling safety is a critical concern. Overflow is one of the most severe drilling accidents and can lead to blowouts or worse if not promptly managed. Forecasting overflows in advance is essential to prevent serious accidents. To accurately detect overflow abnormal conditions and ensure drilling safety, this paper designs and implements a method named Shapelet-aware Graph Temporal Network for anomaly detection of oil drilling operations. This method integrates Shapelet extraction for pattern recognition, temporal modeling via advanced sequence learning, and graph-based spatial learning for capturing inter-variable dependencies. Experiments conducted on real-world drilling datasets (16 wells with 73 parameters) reveal our method, which can superiority over classic models such as recurrent neural networks based models, as well as models derived from graph convolutional neural networks. Achieving an accuracy of 87.75%, recall of 83.91%, and F1-score of 61.19%, our method demonstrates its effectiveness in identifying rare overflow events while maintaining computational efficiency, which can offer a practical solution for enhancing drilling safety.