[society of exploration geophysicists seg technical program expanded abstracts 2011 - ()] seg...
TRANSCRIPT
Well log data inversion using radial basis function network Kou-Yuan Huang* ,National Chiao Tung University, Liang-Chi Shen, University of Houston, Li-Sheng Weng, National
Chiao Tung University
Summary
We adopt the radial basis function network (RBF) for well log data
inversion. We propose the 3 layers RBF. Inside RBF, the first layer
is the K-means clustering method and PFS-test. Then the 1-layer
perceptron is replaced by 2-layer perceptron. It can do more
nonlinear mapping. The gradient descent method is used in the
back propagation learning rule at 2-layer perceptron. The input of
the network is the apparent conductivity (Ca) and the output of the
network is the true formation conductivity (Ct). 25 simulated well
log data are used in the training. From experimental results, the
network with 10 input data, first layer with 27 nodes, second layer
with 9 hidden nodes and 10 output nodes can get the smallest
average mean absolute error in the training. After training in the
network, we apply it to do the inversion of the 6 simulated well log
data and 1 real field well log data to get the inverted Ct. Result is
good. It shows that the RBF can do the well log data inversion.
Introduction
The inversion of the well log is a difficult problem to get the
properties of the true formation around the well. The well log
inversion were ever done using the method of least-square error by
Lin et al. (1984), the maximum entropy method by Dyos (1987), and
the neural network with Levenberg-Marquardt (L-M) optimization
algorithm by Martin et al. (2001). They were computationally
intensive for the well logs with high relative dip angle. Goswami et
al. (2004) used the evolution algorithm to invert the well log, but it
was a less efficient method. As compared with these methods, the
neural networks could obtain a faster operational speed and a
robustness result. Powell (1985) proposed the radial basis function
network (RBF), it had computational efficiency. Here we adopt the
RBF to do the inversion of well log data.
We modify the two-layer RBF to the three-layer RBF in the well
log data inversion. There are two steps in the network. The first
step is the K-means clustering algorithm and Pseudo F-statistic
(PFS) test to determine the number of nodes in the first layer. The
second step is a 2-layer perceptron. The 1-layer perceptron is
replaced by 2-layer perceptron. It has a more non-linear mapping.
RBF Network and Learning
Conventional 2-layer RBF Network
Figure 1 is a 2-layer RBF network. It is a supervised training
model (Huang, 2003). First, classify all training samples to I
clusters by unsupervised K-means clustering algorithm (Tou and
Gonzales, 1974). Center and variance of each cluster is
recorded in the first layer.
∑
(1)
where denotes each input sample.
Figure 1: Conventional RBF network.
The first layer is to calculate the statistic distance between input
sample and each cluster center, and then calculate the response of
Gaussian basis function:
[
] (2)
The second layer is the 1-layer perceptron. The output of the first
layer becomes the input of the second layer. Then do the linear
combination in inputs and weights, and get the output of node :
∑
Output of node is ( ) is the activation function. We use
sigmoidal activation function, and the output of perceptron is:
( )
(4)
where is the output of node of the second layer. The error
function E of network:
∑
(5)
where denotes the node number of second layer, denotes the
desired output of node of second layer, and is the actual
output of node of the second layer. The weighting adjustment is
by gradient descent method (Rumelhart et al., 1986).
(6)
where is the learning rate of network, and is the count of
execution. And the formula of network adjustment:
© 2011 SEGSEG San Antonio 2011 Annual Meeting 499499
Dow
nloa
ded
09/1
4/13
to 1
28.8
3.63
.20.
Red
istr
ibut
ion
subj
ect t
o SE
G li
cens
e or
cop
yrig
ht; s
ee T
erm
s of
Use
at h
ttp://
libra
ry.s
eg.o
rg/
(( ) ( ) )
( ) ( )=( )
(7)
(8)
To determine the node number of first layer, we use K-means
clustering algorithm and Pseudo F-statistic (PFS) test (Vogel and
Wang, 1979). The PFS is:
(9)
where is the number of samples, is the number of clusters; tr is
the trace of a matrix, i.e., the summation of diagonal elements. Do
K-means clustering algorithm for each K and calculate PFS value.
Find the maximum value of PFS and the corresponding K as the
optimal number of clusters.
Three-layer RBF Network
Figure 2: Three-layer RBF.
To have more nonlinear mapping, we propose three-layer RBF.
Figure 2 is a three-layer RBF network. The first layer is the same
as the conventional RBF in Figure 1. Then it is a 2-layer
perceptron. The weights at the second and the third layer are
adjusted by back-propagation learning rule with gradient descent
method and momentum term (Rumelhart et al., 1986).
The output of first layer is the same as (2):
[
] (10)
where denotes each input sample. Activation function of network
is sigmoidal function, the output of the second layer , and the
output of third layer are:
( ) ∑
∑
(11)
(12)
The formula of weighting adjustment between second and third
layer:
( )
(13)
(14)
The formula of weighting adjustment between first and second
layer:
( (∑
) ( ) )
(∑
) ( )
(15)
(16)
Besides, we adopt the momentum term to consider the effect of
the previous step. Then the final adjustment are:
(17)
(18)
is the learning rate of network, is the momentum
coefficient.
Determination of the Number of Hidden Nodes
From the following experimental results of conventional 2-layer
RBF, the network has the minimum error when we use 10 input
features, 27 nodes in the first layer, and 10 output nodes. We
expand this kind of 2-layer RBF to the three-layer RBF. And we
use the theorem developed by Michandani and Cao (1989) to
determine the number of the hidden nodes in the two-layer
perceptron.
When the inputs of the network are d dimensions and the number
of hidden nodes is H, then the maximum number of linear
separable regions M is:
∑
(20)
If d>H, M=H2 . In the experiment, the input nodes are d=10, the
number of the training patterns is 500. In order to have one pattern
almost in one region, M=500. According to the theorem, we can
get the maximum regions M as follows:
Each region has almost one training pattern, so we may have 8 or
9 hidden nodes (H) in the network.
Training Process
In our experiments, the inverse of apparent resistivity (Ra), which
is called apparent conductivity (Ca), is the input of the network.
And the desired output is the true formation conductivity (Ct). The
transition diagram is shown in Figure 3.
Figure 3: Transition of apparent properties and true formation
properties in the processing.
© 2011 SEGSEG San Antonio 2011 Annual Meeting 500500
Dow
nloa
ded
09/1
4/13
to 1
28.8
3.63
.20.
Red
istr
ibut
ion
subj
ect t
o SE
G li
cens
e or
cop
yrig
ht; s
ee T
erm
s of
Use
at h
ttp://
libra
ry.s
eg.o
rg/
There are 31 simulated well log datasets and 1 real field log. We
use the 1st to the 25th well logs for training, and others are for
testing. The depth of simulated well is from 490 to 589.5 feet and
the sample interval is 0.5 feet. Each well log has 200 Ca data and
200 Ct data. We divide the 200 data into every 1, 2, 4, 5, 10, 20, 40,
50, 100, and 200 data as the input features to the networks.
Because we do the inversion, the number of output nodes is the
same as the number of input nodes. We compare the mean absolute
error of each network in the training process to determine the best
network. The mean absolute error (MAE) is defined:
∑ ∑|( )|
(19)
where P is the number of training patterns, K is the number of
output nodes, and and are the desired output and real
output of pattern p at the output node k. The number of iterations is
set to 20,000. is set to 0.6. is set to 0.4. The error threshold for
stopping training is set to 0.002.
Experimental Results
Results of Conventional 2-layer RBF
We use the 2-layer RBF on the well log inversion. The numbers of
input nodes are 1, 2, 4, 5, 10, 20, 40, 50, 100, and 200 respectively.
We have 25 training well logs. The training results are shown in
Table 1. From Table 1, the network has the minimum inversion
error when we use 10 input features, 27 nodes in the first layer, and
10 nodes in output layer. Figure 4 shows 27 is the maximum in
PFS value versus cluster number in 10 input features. After
training, we use the network to do inversion on the 26th-31st well
log data and the results are shown in Table 2. Figure 5 shows the
error MAE versus iteration in 10-27-10 RBF. Figure 6 shows the
inverted Ct and desired Ct of the 26th well log with 10-27-10 RBF.
Network
size
Number of
training patterns
Error MAE at
20,000 iterations
Error MAE of 6 well
log data inversion
1-2426-1 5000 0.018645 0.123119
2-14-2 2500 0.045808 0.071876
4-7-4 1250 0.049006 0.069823
5-44-5 1000 0.032716 0.058754
10-27-10 500 0.031394 0.048003
20-7-20 250 0.048767 0.073768
40-2-40 125 0.164247 0.174520
50-2-50 100 0.160658 0.165190
100-2-100 50 0.185587 0.188436
200-4-200 25 0.277741 0.252023
Table 1: Experimental results of conventional RBF.
Well Log Data Error MAE of well log data inversion
#26 0.051753
#27 0.055537
#28 0.041952
#29 0.040895
#30 0.047587
#31 0.050294
Table 2: Inversion error of RBF (10-27-10).
Figure 4: Result of PFS value versus cluster number at 10 input features.
Figure 5: Mean absolute error vs. iteration in 10-27-10 network learning.
Figure 6: Inverted Ct and desired Ct of the 26th well log with 10-27-10
network.
© 2011 SEGSEG San Antonio 2011 Annual Meeting 501501
Dow
nloa
ded
09/1
4/13
to 1
28.8
3.63
.20.
Red
istr
ibut
ion
subj
ect t
o SE
G li
cens
e or
cop
yrig
ht; s
ee T
erm
s of
Use
at h
ttp://
libra
ry.s
eg.o
rg/
Results of Three-layer RBF with Original Features
In the experiments, we adopt the three-layer RBF with 7, 8, 9,
10, 11, and 12 hidden nodes in the 2-layer perceptron respectively.
We have 25 training well logs. The inversion errors in training are
in Table 3. When the number of iteration is set to 20,000, we find
that the 10-27-9-10 network can get the smallest error MAE than
that of the conventional RBF. The error versus iteration in the
training process is shown in Figure 7. We try the hidden nodes
from 12 to 24 in the 2-layer perceptron, but the errors do not
decrease. After training, we use the network to do the 26th-31st
well log data inversion and the results are shown in Table 4. The
inverted real output Ct of the 26th well log with 10-27-9-10
network is shown in Figure 8. Because the MAE error between the
inverted Ct and the desired Ct of the 26th well log is 0.041526 that
is the smallest error in Table 4, the inverted Ct and the desired Ct
on the Figure 8 are very close.
Network size Number of
training
patterns
Error MAE at 20,000
iterations
Error MAE of 6 well log data
inversion
10-27-7-10 500 0.017231 0.050430
10-27-8-10 500 0.017714 0.050313
10-27-9-10 500 0.015523 0.046625
10-27-10-10 500 0.015981 0.048452
10-27-11-10 500 0.019848 0.048173
10-27-12-10 500 0.021564 0.053976
Table 3: Inversion error of each kind of three-layer RBF.
Well Log Data Error MAE of well log data inversion
#26 0.041526
#27 0.059158
#28 0.046744
#29 0.043017
#30 0.046546
#31 0.042763
Table 4: Inversion error of three-layer (10-27-9-10) RBF.
Figure 7: Mean absolute error vs. iteration in 10-27-9-10 network learning.
Figure 8: Inverted Ct and desired Ct of the 26th well log with 10-27-9-10
network.
Result of Real Well Log Data Inversion
We adopt the 10-27-9-10 RBF to do the real well log data
inversion. The well depth is from 5,577.5 to 6,722 feet, and the
sampling interval is 0.5 feet. So, there are total 2,290 input features
from this well. Every 10 data are the inputs. After the previous 25
well log training, the real data are fed to the network for inversion.
Figure 9 shows the output, the inverted Ct.
Figure 9: Result of real well log data inversion.
Conclusions
From experiments we find that the three-layer RBF can get the
less mean absolute error between the actual output and the desired
output. The reason is that the 2-layer perceptron has the property of
more non-linear mapping. The three-layer RBF with 10 input
features, 27 nodes in the first layer, 9 hidden nodes in the second
layer, and 10 output nodes can get the smallest average mean
absolute error on simulated well log data. In the final step, the
network with the smallest error in training is applied to the
inversion of real field log. The result is good. It shows that the
RBF network can do the well log data inversion.
© 2011 SEGSEG San Antonio 2011 Annual Meeting 502502
Dow
nloa
ded
09/1
4/13
to 1
28.8
3.63
.20.
Red
istr
ibut
ion
subj
ect t
o SE
G li
cens
e or
cop
yrig
ht; s
ee T
erm
s of
Use
at h
ttp://
libra
ry.s
eg.o
rg/
EDITED REFERENCES
Note: This reference list is a copy-edited version of the reference list submitted by the author. Reference lists for the 2011
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
Dyos, C. J., 1987, Inversion of induction log data by the method of maximum entropy: 28th Annual
Logging Symposium, paper T, 1–13.
Goswami, J. C., R. Mydur, P. Wu, and D. Heliot, 2004, A robust technique for well-log data inversion:
IEEE Transactions on Antennas and Propagation, 52, 717–724, doi:10.1109/TAP.2004.825158.
Huang, K.-Y., 2003, Neural networks and pattern recognition: Weikeg Publishing Co.
Lin, Y. Y., S. Gianzero, and R. Strickland, 1984, Inversion of induction logging data using the least
squares technique: 25th Annual Logging Symposium Transactions, paper AA, 1–14.
Martin, L. S., D. Chen, T. Hagiwara, R. Strickland, G. Gianzero, and M. Hagan, 2001, Neural network
inversion of array induction logging data for dipping beds: Society of Professional Well Log
Analysts: 42nd Annual Logging Symposium, paper U, 1–11.
Mirchandani, G., and W. Cao, 1989, On hidden nodes for neural nets: IEEE Transactions on Circuits and
Systems, 36, 661–664, doi:10.1109/31.31313.
Powell, M. J. D., 1985, Radial basis functions for multivariate interpolation: A review, technical report
DAMPT 1985/NA 12, Department of Applied Mathematics and Theoretical Physics: Cambridge
University.
Rumelhart, D. E., G. E. Hinton, and R. J. Williams, 1986, Learning internal representations by error
propagation, in D. E. Rumelhart and J. L. McClelland, eds. Parallel distributed processing:
Explorations in the Microstructure of Cognition: MIT Press.
Tou, J. T., and R. C. Gonzalez, 1974, Pattern Recognition Principles: Addison-Wesley Publishing Co.
Vogel, M. A., and A. K. C. Wong, 1979, PFS clustering method: IEEE Transactions on Pattern Analysis
and Machine Intelligence, PAMI-1, 237–245, doi:10.1109/TPAMI.1979.4766919.
© 2011 SEGSEG San Antonio 2011 Annual Meeting 503503
Dow
nloa
ded
09/1
4/13
to 1
28.8
3.63
.20.
Red
istr
ibut
ion
subj
ect t
o SE
G li
cens
e or
cop
yrig
ht; s
ee T
erm
s of
Use
at h
ttp://
libra
ry.s
eg.o
rg/