A Hybrid Physics/Data Driven Modeling Approach for Virtual Sensors

Abstract

Sensor issues arise quite often in many fields such as data showing anomalous behavior or data being corrupt. This involves finding either faulty, noisy and malfunctioning sensors or anomalous behavior of the physical system deviating from the normal behavior indicating either new physics or the assumptions of the current model are being violated. This necessitates integration of domain-specific reduced form physics-based engineering models with data-driven modeling techniques to model effectively by covering wider data space. There could be sensors on the order of thousand but not every sensor is relevant and useful to the system modeling. Real-time drilling modeling is used as a prototype for demonstrating the new algorithm to deal with modeling efficiently virtual sensors. This paper provides a new real-time model with deep neural network (DNN) using a new hybrid physics/data driven algorithm that can intelligently pick the models to retrain and predict accurately for virtual sensing. This approach offers an improved and efficient methodology to arrive at the decision of whether the sensors are malfunctioning, or the physics models needs to be updated to model the new behavior. This method was applied to predict rate of penetration (ROP) with automatic sensor value predictions of hookload (HL), rotations per minute (RPM), pressure (P) and How rate (Q) for drilling. The physics model is obtained from the engineering models produced from domain insight. Thus, the modeling integrates reduced form physics-based engineering models into DNN framework. The generated data from the engineering model are needed to fill the void space in the surface not covered by the real-time measured data. The hybrid physics/data driven algorithm is fast, as the training is performed whenever the deviation occurs either between the model predictions and sensor values or ROP predictions deviate or both occur. The hybrid model uses the DNN framework to speed up the predictions and improve the accuracy of the ROP. The new hybrid modeling approach developed in this paper for virtual sensors can be applied to any real-time modeling system.