Professional Documents
Culture Documents
Giancarlo Bernasconi* and Diego Rovetta, Politecnico di Milano, Dipartimento di Elettronica e Informazione
Summary While-drilling services are effective and efficient in supporting drilling decisions. Real-time processing of seismic borehole data can help drilling operations. We study the problem of bit localization in a vertical layered media by using while-drilling measurements. The signal emitted by the bit is recorded by surface sensors and it is processed in order to obtain direct arrival traveltimes. The traveltimes are the input of a Bayesian inversion procedure for the estimation of the bit position and of the velocity model: the results are the estimated values and their associated uncertainty. We propose a parameterization of the problem and we show the conditioning analysis. We test the inversion strategy on realistic cases with significant results. While drilling data processing Eidsvik and Hokstad (2006) show a method for assessing drill-bit position, distance to look-ahead events, and parameters of the velocity model using seismic borehole data. We follow their approach in order to recover bit position and subsurface velocity parameters from whiledrilling data. We extend their model to 3D acquisition geometries and we study the conditioning of the inverse problem by analyzing the eigenvalues and eigenvectors of the linearized forward model. Starting data are the recordings of the vibrations produced by the drill-bit during perforation. Receivers are placed on the surface (geophones) and on the drillstring (pilot sensors). Basic processing includes drillstring deconvolution of the pilot signal, crosscorrelation with the geophone signals and time correction of the propagation time through the drillstring, in order to produce RVSP-like gathers (Poletto and Miranda, 2004). Then, a picking procedure obtains traveltimes from every bit level to surface locations. Figure 1 is a schematic flow chart of the basic processing, with an example of starting and final geophone gather. Direct arrival traveltimes are the input of the estimation procedure. As the perforation proceeds slowly with respect to while-drilling data processing, traveltimes can be computed repeatedly while the bit is moving down, for example at fixed step (tens of meters) of the total length of the drillstring,. Traveltime picking can be effectively automated after an initial manual training. Model parameterization Let J denote the number of geophones on the surface, (x0, y0, z0) the bit position, and (xj, yj, zj) the geophones position, with j=1, , J. We consider a 1D velocity model V(z). The traveltime from the bit to the j-th geophone can be approximated by an hyperbolic equation
tj =
( x0 x j ) 2 + ( y0 y j ) 2 + ( z0 z j ) 2 Vrms ( z 0 )
(1)
where Vrms(z0) takes into account the medium between the bit and the geophone. If the geophones are at different heights over the surface, i.e. the topography is not flat, Vrms(z0) is also a function of the geophone. For sake of simplicity we will not indicate this dependence, but in the examples we are correctly handling it. The estimation procedure aims at recovering in real time the bit position and the medium velocity along the well path. We discretize the trajectory of the bit, so that we have K bit locations in the subsurface, (x0,k, y0,k, z0,k), with k=1, , K. We model the transition from level k to level k+1 as a straight segment,
(2)
geophone
drillstring decon
geoph. raw data
equalization
geoph. proc. data
crosscorrelation
Time [s]
Time [s]
Segnale elaborato
picking
Bit depth [m]
Bit depth [m]
Figure 1. Processing of while drilling data. On the left: geophone gather, raw data. On the right: geophone gather, processed data. The arrow indicates the direct arrival curve.
The estimation procedure adopts a dynamic model which evolves from a known state mk=[k, k, Vrms,k]T to a new state mk+1. By substituting (2) in (1) we obtain a non linear function g relating the model state to the observation space of the traveltimes, dk=[t1,k, t2,k, , tJ,k]T, so that dk=g(mk).
343
x y
pk
z
Figure 2. Well geometry.
k+1
If the traveltimes are biased by some unknown quantity, i.e. the exact reference time is not available, it is possible to rephrase the problem using time differences (Haykin, 1995). Depending on the cause of the time uncertainty, it is convenient to choose one geophone as reference, or one bit level as reference, or to use traveltime differences between adjacent geophones, etc. Conditioning analysis In order to better understand the sensibility of the measurements (traveltimes) with respect to the model parameters variations, we linearize the relation between model state and observations around a reference state m0, by calculating the jacobian of g,
Figure 3. Eigenvectors in the model space. On the left: associated model parameters. Bottom: eigenvalues in decreasing amplitude.
d = g( m 0 ) + G( m m 0 ), g g g G= , , . Vrms
The singular value decomposition (SVD) of G is
G = U V T ,
where is the diagonal matrix of the eigenvalues, V is the matrix of the eigenvectors in the model space, and U contains the eigenvectors in the data space. The eigenvalues represent the energy in the data due to a unitary variation along the eigenvectors in the model space. This allows us to study the problem conditioning. As reference model we consider a source (bit) right below the well-head, with z0=800 m, and moving forward with =1, =45, p=10 m. Reference velocity is Vrms=4500 m/s. We simulate a geophone layout on a square surface with a side of
decreases for the receivers placed perpendicular (for ) and parallel (for ) to the horizontal projection of the bit direction, which can change during the perforation. We have investigated the conditioning for other reference models and for different parameterization of the data space (time differences instead of exact times), obtaining similar results. This analysis can be used as a tool for designing the acquisition geometry, in order to achieve maximum sensibility towards critical events and/or deviations of the planned well trajectory. Inversion procedure We look at the perforation as a discrete-time (or discrete drillstring length) process. The underlying variables are initialized at some first reference level, and then they evolve to increasing levels until the last level K. The variables at level k+1 depend on the variables at level k. The evolving model state is the vector mk, as defined in the previous sections. Traveltime data are acquired at each level. They are the observations dk. The level step is
344
Figure 4. Eigenvectors in the data space. The first eigenvector is associated with the well inclination, the second eigenvector with the medium velocity, the third eigenvector with the well azimuth.
controlled by the known quantity pk, the length of drillpipes inserted in the well between two data measurements. The step can be non constant. We consider a maximum step of few tens of meters. According to the extended Kalman filter theory (Haykin, 1995), we use the model state mk and some known process model function f to predict the state at level k+1,
is assumed to be independent for the three model parameters, white and with normal probability distribution. We have computed the model noise variance from experimental data, with pk10 m. We use
= 2 o ; = 20 o ; Vrms = 5 m/s .
Example We consider data coming from a realistic acquisition geometry. Data is acquired every 10-20 m (not constant step), starting at the bit depth of 700 m until the maximum depth of 2000 m. Geophones are placed on the surface along three lines departing from the well. Minimum offset from well-head is 400 m, maximum offset 1000 m. Total number of geophones is 21. The inversion procedure is initialized at the first bit level (700 m depth). Measurement noise error is assumed white, zero mean, normally distributed. In the first simulation we set a standard deviation of the measurement noise of 1 ms. Figure 5 is the result. The panel on the left is the distance between the estimated well trajectory and the real trajectory. The central panels are the estimated model state parameters (red) over the real ones (green). The figure on the right is a 3D visualization of the well, with the actual trajectory (green), and the estimated trajectory (red) with its standard deviation (grey tube). Bit localization error is less than 15 m, velocity error is less than 10 m/s. Figure 6 shows the results for a higher measurements noise (standard deviation 5 ms). Bit localization error is less than 40 m, velocity error is less than 70 m/s. Conclusions We have tested the possibility of assessing bit position and velocity model from while drilling data. Vibrations
k +1 = f ( m k ) , m
where ^ indicates a predicted value. We introduce in the function f the information and/or the constrains about the perforation process. Taking into account the rigidity of the drill-pipes, we predict the well inclination and azimuth with
k +1 = k ,
= . k +1 k
For the velocity, as Vrms is generally a smooth function of depth, we use a linear extrapolator
The difference between the predicted data and the actual data is used to update the predicted model, which serves as starting point for a new iteration. For sake of brevity we do not report here the equations governing the Kalman filter: they can be found in Haykin (1995). It is important to remember that this estimation procedure obtains also the uncertainty of the solution at each level. The user has to provide a process noise covariance, expressing the uncertainty of the prediction of the model state. This noise
345
Figure 5. Inversion results. Standard deviation of measurements noise is 1 ms. From left to right: distance between estimated and actual bit position, estimated (red) and actual (green) well inclination, estimated (red) and actual (green) well azimuth, estimated (red) and actual (green) medium velocity. On the right: well trajectory (green), estimated well trajectory (red), estimation uncertainty (grey tube)..
Figure 5. Inversion results. Standard deviation of measurements noise is 5 ms. From left to right: distance between estimated and actual bit position, estimated (red) and actual (green) well inclination, estimated (red) and actual (green) well azimuth, estimated (red) and actual (green) medium velocity. On the right: well trajectory (green), estimated well trajectory (red), estimation uncertainty (grey tube).
346
EDITED REFERENCES Note: This reference list is a copy-edited version of the reference list submitted by the author. Reference lists for the 2007 SEG Technical Program Expanded Abstracts have been copy edited so that references provided with the online metadata for each paper will achieve a high degree of linking to cited sources that appear on the Web. REFERENCES Eidsvik, J., and K. Hokstad, 2006, Positioning drill-bit and look-ahead events using seismic traveltime data: Geophysics, 71, no. 4, F79F90. Haykin, S., 1995, Adaptive filter theory: Prentice-Hall, Inc. Poletto, F., F. Miranda, 2004, Seismic while drilling fundamentals of drill-bit seismic for exploration: Elsevier Science Publ. Co., Inc.
347