Combining Drilling Condition Analysis With Unsupervised Time Series Models for Kick Monitoring

Abstract With the continuous development of petroleum exploration and production technology and the increasing complexity of geological conditions, the problem of narrow-margin safe drilling has become a major challenge in the field of petroleum exploration and development. Kick is also one of the high-frequency, high-hazard accidents in drilling. To reduce drilling costs and improve drilling safety, it is crucial to accurately monitor kicks and prevent their further development. A method for monitoring kick, based on the variation of logging parameters, currently exists. However, it does not take into account the impact of abnormal conditions such as pump stoppages on the outlet flow rate and the total volume of drilling fluid in the tank, which can lead to false alarms. In order to improve the accuracy of risk identification and reduce the false alarm rate, a kick monitoring method is proposed that combines drilling conditions with an unsupervised Bidirectional Long Short-Term Memory Autoencoder (BiLSTM-AE). This model simultaneously considers past and future information through forward and backward propagation, effectively extracting temporal features from sequences using bidirectional information. The proposed method was tested using kick monitoring data from 3 wells. The experimental results indicate that the recognition accuracy of the kick intelligent monitoring model based on BiLSTM-AE is 88.85%, surpassing other existing intelligent monitoring models. When combined with abnormal conditions such as pump stoppages, the model’s false alarm rate is reduced by 9.33%. This research can provide a theoretical foundation and important technical reference for accurate kick risk monitoring, especially in conditions where kick risk data labels are lacking. Moreover, it holds significant potential for practical field applications.