The safe and efficient construction of a well requires close attention to how the well and the rig equipment behave as the well is drilled. This primarily involves monitoring and correctly interpreting all the data collected at the rig site. In the past, this was primarily accomplished with an experienced drilling crew. While we still continue to use mostly antiquated rig sensors that were introduced decades ago, there has also been an "explosion" in the number of new advanced sensors used on a rig. The argument put forward for the additional sensors is that they increase safety and help reduce non-productive time and invisible lost time. Many of these new sensors provide data at high(er) sampling rates. This has made human monitoring of the data very difficult, if not impossible, and has led to a variety of event detection algorithms. The primary function of these event detection algorithms is to bring the driller's attention to anomalies in the system, based on data collected from all the rig sensors. However, the frequent occurrences of false and missed alarms, when using automated event detection software, can lead to significant downtime, and also reduce safety at the rig. One reason for these false alarms is bad data due to either sensor failure or data communication failure. Having sophisticated sensors, but not knowing whether the data can be trusted, defeats the very purpose of having additional sensors. This paper explores a methodology that can greatly reduce these false and missed alarms by validating the data coming from the sensors.
The data validation methodology described in this paper utilizes a Bayesian network model of the sensed parameters, to maximize the number of analytical redundancies among the parameters. Also, data is collected and aggregated from multiple sources within and outside the rig, such as the well plan, morning reports and real-time data, to obtain increased certainty in the prediction of sensor / process anomalies. The algorithm allows for the detection and isolation of sensor faults and process faults, and does not impose unrealistic assumptions of process invariance. For any type of drilling operation, the parameters in a drilling model are constantly changing. Here, we briefly discuss how the model can be updated in real-time, depending on whether the fault is in the sensor or in the process. The approach is very flexible and easily accommodates the widely varying number of sensors from rig to rig. The methodology can also be easily scaled to be applied on a small subsystem (top drive, hydraulic unit, etc.) or the entire rig (all surface and sub-surface sensors). The proposed methodology is applied on sample scenarios that are typical of the unconventional shale drilling operations conducted in North America.