A Physics Model Embedded Hybrid Deep Neural Network for Drillstring Washout Detection

Abstract

Abstract One of the practical challenges in the oil and gas industry is the lack of quality data for applying machine learning techniques. A way to tackle this problem is to build a hybrid system that combines physics models with machine learning workflows. To demonstrate the applicability, the proposed hybrid model has been applied to drillstring washout detection which is relatively a common but severe and very expensive failure in drilling. We propose a hybrid deep neural network (hybrid-DNN) composed of three components – Parameter Network (PNet) for estimating model parameters, Residue Network (RNet) for predicting regression or classification results, and a physics model appropriate for the problem at hand. PNet learns the system behavior based on the embedded physics model, which it controls through adjusting model parameters. RNet utilizes the outputs from the PNet and physics model as input and is being trained for predicting the residual. Once trained, the hybrid system can control the parameters of the physics model and predict the desired results in real-time. The proposed hybrid system has been applied to drillstring washout detection. Traditionally, drillstring washouts have been detected by monitoring the trends of hydraulic coefficients between standpipe pressure and flow rate of the circulated drilling fluid. Several data-driven methods based on statistical change detection have been tried to automatically detect the occurrence of abnormal trends. However, most current applications struggle to identify true washout events since they largely overlap with normal drilling patterns in noisy measurements. The hybrid-DNN has been verified by augmented drillstring washout cases and actual field events. When fed with a real-time data stream, it is able to extract the optimized model parameters and identify normal drilling conditions and potential washout situations. It also demonstrates that the hybrid-DNN predicts more reliable results without false-positives compared to the data-driven approach without a physics model. The proposed physics model embedded hybrid-DNN provides a general-purpose framework for many different domain problems based on physics.