Online Failure Detection

Abstract

Well drilling is an expensive process. The cost of keeping people and equipment on the drill site cost several hundred thousand to millions each day. Some fault situations halt the operation and demand time consuming operations before drilling can be resumed. Pack off is a fault situation where the borehole is being plugged by loose rocks produced when the ground is being drilled into. The situation develops over time. And can cause the well to be abandoned if not dealt with. This paper investigates methods for predicting a pack off situation, based on sensor data measured during the drilling process. Principal component analysis is used for data compression and Gaussian process regression is used for predictions. The results on the training data set indicate that the model is very expressive. This may also cause overfitting problems, but due to late arrival of test data, which also revealed to be incompatible with the training data. Verification of the results is not conducted in great detail.