Bad quality data is ubiquitous in oil well drilling operations. The hostile and uncertain environment in which sensors are used and the lower priority/importance usually given to regular calibration of these sensors to ensure accuracy are the main reasons for the lack of good quality and reliable data. Today, online streaming of sensor data from at least all the primary sensors on a rig to a real-time data monitoring facility is becoming commonplace. There also exists software to analyze this data, in real-time, to detect trends and identify potential drilling problems long before they occur. However, these software applications require good quality data to perform accurate analysis. When data is not validated, false and missed alarms are the norm whereby supposedly autonomous software applications require continuous human supervision. High quality data will also be required in the near future when adopting control algorithms for rig automation, which will use this data for making autonomous, closed-loop decisions.
In this paper, a novel technique for real-time sensor data validation is proposed that improves the quality of data collected from the sensors on a rig. This technique uses a model-based approach and the principle of conditional independence. Here, a sensor probabilistic graphical model is created and the relational redundancies in the model are exploited to differentiate between a sensor fault and a process fault as well as to identify the faulty sensor. The approach described in this paper allows for model update in real-time to account for process degradation/changes, thereby alleviating problems associated with the use of a purely analytical model. The application of the algorithm on a generated data set and the results of the experiment are presented in this paper. The adoption of such techniques is expected to greatly enhance the safety and the performance of well drilling and pave the path forward for more automated rig operations.