Towards Real-Time Bad Hole Cleaning Problem Detection Through Adaptive Deep Learning Models

Abstract

Monitoring of Equivalent Circulating Density (ECD) may improve assessment of potential bad hole cleaning conditions if calculated and measured sufficiently accurately. Machine learning (ML) models can be used for predicting ECD integrating both along-string and surface drilling measurements and physics-based model (PBM) results, even though their generalization is often challenging. To remediate this generalizability issue, we present an adaptative predictive deep-learning model that is retrained with new measurements in real-time, conditionally that the new measurements are not detected as anomalies. Past ECD measurements, corresponding values predicted by a 1D PBM and other drilling measurements are used as input to a deep learning model, which is pretrained on historical drilling data without any hole cleaning problem. This model has two components: an anomaly detector, and a predictor. In this paper, both components are based on combinations of Long Short-Term Memory (LSTM) cells that allow (1) to account for data correlations between the different time series and between the different time stamps, and (2) generate future data conditioned to past observations. As drilling progresses, new data is proposed to the anomaly detector: if the network fails to reconstruct them correctly, an alarm is raised. Otherwise, the new data is used to retrain the models. We show the benefits of such an approach on two real examples from offshore Norway with increasing complexity: For the first one, with no major drilling issue, we simply use ECD from the PBM to predict ECD ahead of the bit. The second example had multiple issues linked with mud loss and poor hole cleaning. For this latter case, we used additional topside measurements to better constrain the ECD prediction.