Ulf Jakob Flo Aarsnes

Drilling operations invite failure situations that cause harmful environmental consequences and high monetary losses. We consider mechanical faults, mud loss, and insufficient cuttings transport, which can happen at unexpected moments during the operation. We develop a novel real time fault detection and diagnosis (FDD) algorithm that employs multitask neural networks (MTL-NNs), trained with synthetic data using a simulator that integrates four coupled hyperbolic partial differential equations which describe the distributed flow and pressure dynamics across the drill string and annulus. Due to the uniqueness of each new well being drilled and the scarcity of historical data representing faulty operations, the utilization of synthetic data to create models is the only realistic avenue. By utilizing only topside measurements, we cater to the practical limitation that down-hole measurements are typically not available in real-time. The predictions of the NN are required to be robust with respect to small temporal shifting, since the onset of the fault and the onset of the successful identification of the fault are typically unknown. To fulfill this requirement, we train the NN with timeseries taken at different temporal window slices. In addition, we test the NN's robustness on meaningful modelling errors. The trained NN can accurately detect faults and accurately diagnose washout location, reservoir pressure in mud loss, and blockage coefficient for pack-off and consistently obeys the temporal shift robustness requirement. We concluded that robustification with respect to meaningful modeling errors is required after perturbing the friction profiles, density profiles, and bulk modulus.

Purpose: The purpose of this article is to test whether floor plan image segmentation can be used to improve automated valuation model (AVM) accuracy and whether image segmentation provides the opportunity to assess single aspects of property floor plans. Design/methodology/approach: Using a dataset comprising floor plans of 5,498 apartments sold in Oslo, we estimate balcony sizes by image-segmenting the rooms in floor plans using our machine learning model FloorPlanNet and extracting the size of the balcony from the segmentation. We also extract balcony size using text recognition. Then we utilize two models for AVM estimation – hedonic regression (linear OLS) and the non-linear XGBoost model – before measuring feature importance using SHAP. Findings: Our experiments show that including balcony size as a feature in AVMs enhances model performance. We also find that balcony size has a positive but diminishing impact on property price. Research limitations/implications: Demonstrating that image segmentation can be used for valuation in AVMs opens up the possibility to value numerous other aspects of dwelling floor plans. Practical implications: The use of floor plans in AVMs can provide a more objective valuation of single apartment floor plan aspects, giving architects and developers better insights into how homes should be designed. Originality/value: To our knowledge, this is the first attempt to extract features for use in AVMs from floor plans.

In underbalanced drilling (UBD), the bottom hole pressure is kept below pore pressure, causing pressure-dependent influx of reservoir gas into the wellbore that makes the system unstable at low drawdowns. In this article, we propose a feedback controller which stabilizes the system, represented by the “reduced drift flux model (DFM),” around an arbitrary pressure setpoint, while using only topside measurement and assuming unknown reservoir parameters. A particular challenge with this problem is the distributed and highly nonlinear nature of the system dynamics, where the “reduced DFM” models gas–liquid flow as a nonlinear transport equation with a nonlocal integral source term. An observer estimates the distributed gas concentration, downhole pressure, and reservoir parameters by solving the system dynamics backward relative to how the gas rises in the well. The control inputs are then constructed by designing target states over the next sampling period and again solving reversed dynamics to obtain the required topside pressures. The resulting controller is implemented with a 2-min zero-order hold to accommodate the actuation limitation situation on an actual drilling rig. Finally, the results are illustrated in simulations with an industry standard drift flux formulation as the plant model.

In this letter, we design a robust stabilizing controller for a system composed of two sets of linear heterodirectional hyperbolic PDEs, with actuation at one boundary of one of the PDEs, and couplings at the middle boundary with ODEs in a PDE-ODE-PDE configuration. The system is underactuated since only one of the PDE systems is actuated. The design approach employs a backstepping transformation to move the undesired system couplings to the proximal boundary (where the actuation is located). We can then express this target system as a time-delay neutral system for which we can design an appropriate control law to obtain an exponentially stable target system.

Stick slip is usually considered a phenomenon of bit-rock interaction, but is also often observed in the field with the bit off bottom. In this paper we present a distributed model of a drill string with an along-string Coulomb stiction to investigate the effect of borehole inclination and borehole friction on the incidence of stick-slip. This model is validated with high frequency surface and downhole data and then used to estimate static and dynamic friction.A derivation of the torsional drill string model is shown and includes the along-string Coulomb stiction of the borehole acting on the string and the ‘velocity weakening’ between static and dynamic friction. The relative effects of these two frictions is investigated and the resulting drillstring behavior is presented. To isolate the effect of the along-string friction from the bit-rock interaction, field data from rotational start-ups after a connection (with bit off bottom) is considered. This high frequency surface and downhole data is then used to validate the surface and downhole behavior predicted by the model.The model is shown to have a good match with the surface and downhole behavior of two deviated wellbores for depths ranging from 1500 to 3000 meters. In particular, the model replicates the amplitude and period of the oscillations, in both the topside torque and the downhole RPM, as caused by the along-string stick slip. It is further shown that by using the surface behavior of the drill-string during rotational startup, an estimate of the static and dynamic friction factors along the wellbore can be obtained, even during stick-slip oscillations, if axial tension in the drillstring is considered. This presents a possible method to estimate friction factors in the field when off-bottom stick slip is encountered, and points in the direction of avoiding stick slip through the design of an appropriate torsional start-up procedure without the need of an explicit friction test.

We detail in this article the importance of a change of strategy for the delay-robust control of systems composed of two linear first order hyperbolic equations. One must go back to the classical trade-off between convergence rate and delay-robustness. More precisely, we prove that, for systems with strong reflections, canceling the reflection at the actuated boundary will yield zero delay-robustness. Indeed, for such systems, using a backstepping-controller, the corresponding target system should preserve a small amount of this reflection to ensure robustness to a small delay in the loop. This implies, in some cases, giving up finite time convergence.