| dc.description.abstract | Oil and gas drilling is an advanced process with very little instrumentation, where
drilling uid is transported through rotating drillstrings of up to several kilometers,
possibly at extreme depths with high pressure and temperature. A drilling bit is
used at the bottom of the drillstring to crush the formation, and the drilling uid is
used to carry the cuttings to the surface, as well as maintain the pressure in the well.
Drilling is a costly operation, especially o
shore. Incidents can occur that may slow
down the progress. Detecting such incidents manually, especially those occurring
down in the well, may be di cult. Early symptoms may give small variations in
pressure, temperature, and ow rates, possibly covered in measurement noise.
The push for drilling more complex wells in more remote locations demands more
from the drilling control and monitoring system. With advances in drilling control
technology such as managed pressure drilling, and sensor technology such as wired
drill pipe, the complexity of the control system greatly increases. With a high data
rate of sensor readings, as well as lower operation margins, an e cient automatic diagnosis
system is instrumental in reducing operational delays. This thesis presents
di
erent model-based methods for achieving early diagnosis of di
erent drilling incidents,
possibly distinguished from sensor bias, and with estimation of the incident
magnitude. The model-based diagnosis system consists of two parts; rst some
residuals are generated using either adaptive observers or analytical redundancy
relations, then changes to these residuals are detected using a statistical change detection
algorithm, required due to measurement noise. Univariate and multivariate
generalized likelihood ratio tests are applied, using the probability density function
that best matches the noise of the residuals. The thresholds are found using the
probability distribution of the test statistic, determined by a speci ed probability
of false alarms. The probability of fault detection is also found as a function of the
threshold, where data during the incidents are available.
Data from a medium-scale ow loop is used to test the diagnosis method, where
the noise of the residuals ts the t-distribution well. A multivariate change detection
method considering multiple residuals jointly is found to be superior over
a univariate method considering each residual separately, and is used to detect
and isolate the di
erent incidents occurring in the test data. Furthermore, the t-
distribution is shown to give an increased probability of detection compared with
assuming the more common Gaussian distribution. Simulation of a drilling incident
in the high- delity multi-phase simulator OLGA with Gaussian noise in the
measurements is also considered.
The diagnosis framework proposed in this thesis is module-based, where the
methods in each module are simple enough to be implemented in drilling monitoring software at the rig, and can be run in real-time. However, a limitation with the
proposed method is that good data during the normal operating mode is required
for reliable detection and isolation. Future work and implementations should take
this into account, and facilitate automatic acquisition of new data when changes to
the process are made. | nb_NO |
