On-line drill system parameter estimation and hazardous event detection

The project aims to develop statistical methods for automatic detection of hazardous events in oil and gas drilling operations. During such operations, drilling fluid is pumped into the system at a pressure which can be observed (subject to noise) where it travels down through the pipe and towards the drill bit. In addition to powering the drilling bit, the fluid also flushes the borehole from the cuttings as it flows back up the system's annulus towards the earth's surface where its exit velocity is observed. The fluid is then collected in a tank to be reused by pumping it back into the system.
Hazardous events within this process are undesirable for both safety and financial reasons. In extreme cases these events may result in costly operations to recover the defected system or loss of human lives due to a malfunctioning drill. Initially we limit the study to simple events, but as the project progresses, more complex scenarios may also be considered.
Because the observable quantities are subject to noise, the problem requires statistical methods to estimate the state of the hidden system. In order to develop the statistical framework to perform these calculations, a PDE model is required to describe the physical behaviour of the drill fluid within the system. For computational efficiency, a simplified surrogate model of Schlumberger's current model is likely to be needed as the most accurate models may be computationally too expensive to be used. One of the initial goals of the project is to develop a surrogate model that also retains a sufficient level of accuracy. Initial studies suggest this may be achieved through an electrical circuit analogue model.
In addition to the development of a surrogate model, there are multiple statistical problems to be studied. One problem is the estimation of parameters that describe the drill system in the absence of hazardous events. Preliminary studies suggest that this could be done using some sophisticated Hidden Markov Model techniques.
Provided that the parameters have been estimated accurately enough, it should be possible to detect the hazardous events. Because the detection of the hazardous events should occur as soon as possible, it is likely that methods suited to non-linear problems, such as sequential Monte Carlo, are needed. An alternative framework that could be studied is the so-called 4D-VAR that has been applied in meteorological applications. As the project progresses, the use of multi-level Monte Carlo methods may also form a focus of the research in order to compensate for the errors propagated by the simplified PDE model.
Within the first six months, a working algorithmic framework that utilises an appropriate surrogate model is sought. This framework will aim to provide a platform for the successful detection of at least one hazardous event. Towards the twelve month mark the aim is to have a better understanding of how different methodologies should be prioritised to make the computations feasible in practice (i.e., the necessity of multi-level methods) and some preliminary implementations of some of these methods.