Drilling Anomalies Identification Using Near Bit Vibrations

Various anomalies while drilling are a major hazard both for safety and economic efficiency. Additional measurements combined with Bayesian approaches allow to identify anomalies and their precursors with high accuracy. In this work, we utilize a high-frequency accelerometer located downhole with Bayesian networks to avert twist-off events. Statistical analysis was performed to identify the most effective method to train the structure of Bayesian network. It was demonstrated that the usage of K2 score-based algorithm produced the most favorable results. Moreover, the model's performance was further enhanced by incorporating domain expertise at the learning and validation stages. The results indicated significant improvement in the performance of the model, with increased accuracy and faster processing speed. Our study achieved an F1 score of 0.84 for the task of anomaly identification. The results of the proposed approach are in good agreement with those obtained by unsupervised machine learning techniques (based on an isolation forest and principal component analysis).