Elimination of False Positives in VIV Identification Using Machine Learning

Abstract

Vortex-induced-vibration (VIV) is an important consideration while drilling at sites with moderate to high current speeds. Planning for drilling operations often includes determination of limits on maximum drilling riser motion amplitude using model simulations. These limits can then be used to raise alarms in the field by comparing motions measured in the field using one or more motion sensors. The determination of such alarm limits is challenging as VIV is a highly nonlinear process, and small changes in the speed or shape of current profile can result in quite different VIV fatigue results for drilling risers, especially in deep water depths. We use feed-forward neural network, which is a powerful machine learning algorithm, to develop a classifier for distinguishing damaging and non-damaging VIV events. The neural network uses acceleration and angular rate data from only three motions sensors located on the upper flex joint, the lower flex joint and the BOP stack. To train the neural network, riser motions and fatigue damage data are generated from SHEAR7 runs on the model of the drilling riser. Thousands of current profiles measured from a current mooring at a deepwater site (water depth > 6,000 ft) are used as inputs to SHEAR7 model in order to capture full variability in VIV response from the actual field environment. Results show that the neural network classifier almost always predicts damaging and non-damaging VIV correctly. The precision, recall, and F1 score (a combination of precision and recall) for the neural network classifier are all close to 100%. A high precision, recall, and F1 score for a classifier implies that it has no false positives and no false negatives. Here, a false negative is defined as the situation when damaging VIV occurred but was identified as a non-damaging VIV event and an alarm is not raised. False positive is the situation when an alarm is raised for damaging VIV when the event was actually not so damaging. On the other hand, the baseline "constant" classifier of conservatively chosen limits (from the same data) for upper and lower flex joint angles results in very low precision and F1 scores, implying too many false positives. While the baseline classifier does not predict any false negatives, it is very expensive because of too many false positives. Furthermore, it carries the risk of being ignored by users due to too many false alarms. This work demonstrates that machine learning techniques can accurately predict damaging VIV events in the field using minimal number of sensors. Such accurate predictions were not possible using traditional methods.