Professional Documents
Culture Documents
CITATIONS READS
59 116
1 author:
Anthony Goh
Nanyang Technological University
107 PUBLICATIONS 2,233 CITATIONS
SEE PROFILE
All content following this page was uploaded by Anthony Goh on 27 November 2015.
TECHNICAL NOTE
A. T. C. GOH*
KEYWORDS: clays; design; numerical modelling and neurons. The basic mathematical concepts and
analysis; piles; statistical analysis. the steps involved in developing back-
propagation neural network models are well
established and described by Caudill & Butler
INTRODUCTION (1990), Eberhart & Dobbins (1990), Masters
Many geotechnical problems lack a precise ana- (1993), Smith (1993) and Goh (1994a, 1994b).
lytical theory or model for their solutions. This is
usually because of an inadequate understanding
of the phenomena involved and the factors affect- EXAMPLE APPLICATION
ing them, as well as a limited quantity and The application of a back-propagation neural
inexact quality of information available network to evaluate the friction capacity f. of
(Chameau, Alteschaelll, Michael & Yao, 1983). In driven piles is presented. The data were drawn
these situations, engineers generally rely on from load test records compiled by Vijayvergiya
empirical design solutions. The information is & Focht (1972), Flaate & Seines (1977) and
usually gathered, synthesized and presented in the Semple & Rigden (1986) for driven piles in clay.
form of design charts or tables, or numerically The data from Flaate & Seines (1977) relate
using empirical formulae. This technical note uses mainly to timber piles; the others are for steel
a practical example to show the potential of pipe piles. Several of the data are averages of
neural networks to improve the synthesizing of more than one test pile at a site, in order to avoid
the information for the development of empirical giving undue weight to essentially repetitive
design aids. Problem domains in which the results from a site (Semple & Rigden, 1986). The
behaviour of the system is governed by non-linear embedded pile length L ranges from 4.7 m to
multivariate relationships, and where reliable case 96 m, the undrained shear strength s, ranges from
records are available, offer the greatest promise. 9 kPa to 335 kPa and the friction coefficient x is
in the range 0.42-1.73. Most of thef, values were
derived from pile compression load tests after
NEURAL NETWORKS allowance for end-bearing, with unit end-bearing
A neural network is a computational mecha- assumed to be 9 x s, at the pile tip. The
nism able to acquire, represent, and compute a undrained shear strength s, was determined
mapping from one multivariate space of informa- mainly from unconfined compression tests.
tion to another, given a set of data representing However, for very soft and soft clays (s, < 40
that mapping (Garrett, 1994). A typical neural kPa), s, was derived mainly from vane shear tests.
network consisting of three layers of neurons is
shown in Fig. 1. The basic architecture of neural
networks has been covered by Rumelhart &
McClelland (1986), Lippmann (1987) and Flood
& Kartam (1994). The most popular neural
network paradigm and the one adopted in this
study is the back-propagation learning algorithm
(Rumelhart, Hinton & Williams, 1986). Example
patterns are presented to the neural network and
learning essentially takes place through the adap-
tation of the connection weights between the
709
710 GOH
Altogether 45 patterns were used for training the patterns that only broadly resemble the training
neural network and 20 patterns (randomly data. The data are summarized in Tables 1 and 2.
selected) for testing the ability of the neural The values of s, and ai listed are the averages for
network to interpolate the correct responses for the embedded pile lengths.
14.1 15.0 96.0 26.0 27.0 23.5 Flaate & Selnes (1977)
13.0 15.0 102.0 15.0 26.0 19.7
11.7 20.0 54.0 23.0 14.0 19.6
17.5 14.3 87.0 23.0 26.0 21.0
15.9 15.0 49-o 17.0 12.0 16.2
8.1 13.5 37.0 13.0 11.0 14.8
7.7 16.5 32.0 15.0 9.0 15.5
10.0 13.5 33.0 10.0 12.0 13.4
12.0 15.5 39.0 12.0 10.0 14.2
10.2 22.0 19-o 15.0 8.0 14.5
24.2 15.0 146.0 19.0 29.0 23.2
17.1 15.0 109.0 57.0 24.0 38.4
12.7 23.2 38.0 19.0 17.0 16.8
10.0 17.0 82.0 36.0 28.0 27.9
14.3 26.0 89.0 22.0 22.0 21.9
22.5 47.0 60.0 45.0 23.0 27.0
5.5 30.5 44.0 30.0 38.0 24.4
19.2 61.0 142.0 31.0 30.7 33.8 Mansur & Focht (1956)
15.2 35.6 448.0 104.0 109.2 109.8 Cox, Kraft & Verner (1979)
12.2 35.6 718.0 162.0 162.0 167.9
43.9 30.5 162.0 38.0 30.0 29.3 Semple & Rigden (1986)
96.0 61.0 354.0 80.0 44.0 44.4
73.8 61.0 273.0 67.0 47.6 37.6
22.6 76.7 651.0 170.0 192.1 180.7 Pelletier & Doyle (1982)
30.5 32.5 153.0 45.0 29.3 33.4 Darragh & Bell (1969)
45.7 32.5 148.0 52.0 21.8 32.9 Peck (1961)
13.7 32.5 112.0 45.0 42.3 35.1 Woodward, Lundgren &
Boitano (1961)
5.5 16.9 51.6 129.5 76.7 68.2 Tomlinson (1970)
29.0 33.0 105.0 39.0 39.8 26.8 Ravmond International (1982)
12.2 16.8 33.0 16.0 9.9 15.3 Semple & Rigden (1986)
14.0 35.1 59.0 30.0 23.4 23.2 Hutchinson & Jensen (1968)
39.6 27.4 297.0 165.0 80.9 80.2 Peck (1958)
30.5 61.0 91.0 52.0 30.7 29.7 McCammon & Golder (1970)
25-9 32.5 99.0 61.0 34.2 35.7 Peck (1958)
13.1 27.4 80.0 110.0 53.9 59.1 ONeill, Hawkins &
Mahar (1982)
20-4 61.0 105.0 208.0 91.5 92.0 Heerema (1979)
9.1 45.0 54.0 144.0 73.4 75.1 R&den. Petit. St John
k P&k& (1979)
16.8 61.0 87.0 100.0 55.0 58.4 United States Army Corps
of Engineers (1979)
13.7 32.5 112.0 137.0 64.4 72.2 Woodward et al. (1961)
18.3 76.2 115.0 335.0 154.1 154.5 Fox, Sutton &
Oksuzler (1976)
4.6 16.9 43.0 120.5 84.6 64.3 Tomlinson (1970)
33.6 32.5 121.4 35.4 30.4 26.0 Vijayvergiya & Focht
(1972)
33.6 32-5 108.0 48.8 27.1 29.9
20.3 32-5 158.2 112.8 53.0 63.9
30.5 51-o 102.8 24.4 23.5 20.9 Blessev (1970)
NEURAL NETWORKS 711
Details of the back-propagation program the most reliable neural network model. The
adopted in this study are described by Goh scatter of the predictedf, values versus the mea-
(1994a, 1994b). Training was carried out until the sured f, values were assessed using regression
average sum squared error over all the training analysis as well as by computing the error rate.
patterns was minimized. This occurred after The error rate (Yeh, Kuo & Hsu, 1993) is defined
about 30000 cycles of training. Training time on as
a 80486 33 MHz personal computer was about
ten minutes. The numbers of input and hidden error rate = 1 error,/N, for i = 1, 2, . . . , N,
neurons were varied to determine the most reli-
able model. A single output neuron was used to (1)
representf, errori = J[c (Tj - O,)/N,,J for j
= 1, 2, , N,,, (2)
Results
Experiments were carried out using a number where errori is the error of the ith pattern, Tj is
of combinations of input parameters to determine the target (expected) output value, Oj is the
8.0 13.5 27.0 9.0 9.0 13.0 Flaate & Selnes (1977)
9.4 29.3 52.0 29.0 18.0 23.1
14.6 16.0 67.0 29.0 16.0 22.3
11.6 17.5 57.0 27.0 18.0 21.3
9.6 19.2 42.0 15.0 13.0 16.0
21.6 45.7 147.0 31.0 28.8 30.2 Mansur & Focht (1956)
36.9 30.6 149.6 28.2 30.5 25.2 Vijayvergiya & Focht (1972)
66.4 32.5 223.0 60.0 31.2 39.9 Peck (1961)
11.6 11.4 44.0 21.0 13.4 17.6 Kirby & Roussel (1979)
22.9 32.5 91.0 52.0 27.0 32.3 Peck (1961)
13.8 19.0 21.0 21.0 13.0 16.4 Flaate & Selnes (1977)
253 27.4 244.0 185.0 88.8 88.6 Peck (1958)
14.9 52.8 66.0 53.0 27.6 35.4 Togrol (1973)
18.3 32.5 51.0 33.0 32.0 22.3 American Railway Engineering
Association (1950)
48.2 61.0 152.0 64-O 37.8 32.4 McCammon & Colder
(1970)
32.0 21.4 141.0 115.0 59.8 51.1 Endley, Ulrich &
Gray (1979)
13.4 27.0 81.0 22.0 20.0 21.5 Flaate & Seines (1977)
24.2 15.0 147.0 19.0 30.0 23.3
15.5 17.5 80.0 72.0 35.0 42.6
12.8 32.5 110.0 96.0 54.7 58.0 Stermac, Selby &
Devata (1969)
$j
na, 100
a
0
0 100 200 300
Measured skw ir~ct~on: kPa Measured skm frictm kPa
Fig. 3. Comparison of predicted @ method, Burland, Fig. 4. Comparison of predicted (Semple & Rigden
1973) and measured skin friction (1986) method) and measured skin friction
NEURAL NETWORKS 713
clay from actual field records. As with any empiri- experience and tests. In Design and construction of
cal or statistical regression technique, the neural ofihore structures, pp. 61-73. London: Institution of
Civil Engineers.
network predictions are safe to apply only in the
Garrett, Jr, J. H. (1994). Where and why artificial neural
context for which they were formulated.
networks are applicable in civil engineering. J. Com-
purng Civ. Engng Am. Sot. Civ. Engrs 8, No. 2, 129-
130.
NOTATION Goh, A. T. C. (1994a). Seismic liquefaction potential
pile width assessed by neural networks. J. Geotech. Engng Am.
r pile skin friction Sot. Ciu. Engrs 120, No. 9, 1467-1480.
i pile embedded length Goh, A. T. C. (1994b). Some civil engineering applica-
N, number of patterns tions of neural networks. Proc. Instn Civ. Engrs
N out number of output neurons Structs & Bldgs 104, Nov., 463-469.
Oj predicted output value Heerema, E. P. (1979). Pile driving and static load tests
s undrained shear strength on piles in stiff clay. Proc. IIth Offshore Tech. Con&
s,/o: strength ratio Houston 2, 1135-1146.
expected output value Hutchinson, J. N. & Jensen, E. V. (1968). Loading tests
7 friction coefficient on piles driven into estuarine clays at Port-of-Khor-
gv average effective vertical stress adjacent to pile ramshahr, Iran, and observations on the effect of
shaft bitumen coatings on shaft bearing capacity. Norw.
Geotech. Inst. 78, 1-12.
Kirby, R. C. & Roussel, G. (1979). ESACC project field
model pile load test, Hamilton Air Force Base Test
REFERENCES Site Norato, California. Amoco.
American Railway Engineering Association (1950). Steel Lippmann, R. P. (1987). An introduction to computing
and timber pile tests, West Atchafalaya Floodway- with neural nets. Acoust. Speech & Signal Processing
New Orleans, Texas and Mexico Railway. Bull. Am. 4, No. 2,4-22.
Rly Engng Ass., 489, 149-202. Mansur, C. I. & Focht, J. A. (1956). Pile loading tests,
Blessey, W. E. (1970). Allowable pile capacity, Mis- Morganza floodway control structure. Trans. Am.
sissippi River Deltaic Plain. Design and installation Sot. Ciu. Engrs 121, 555-576.
of pile foundations and cellular structures, pp. 87-06. Masters, T. (1993). Practical neural network recipes in
Lehigh University. C +. San Diego : Academic.
Burland, J. B. (1973). Shaft friction of piles in clay-a McCammon, N. R. & Golder, H. Q. (1970). Some
simple fundamental approach. Ground Engng 6, No. loading tests on long pipe piles. Gtotechnique 20,
3, l-15. No. 2, 171-184.
Caudill, M. & Butler, C. (1990). Naturally intelligent ONeill, M. W., Hawkins, R. A. & Mahar, L. J. (1982).
systems. Cambridge: Massachusetts Institute of Load transfer mechanisms in piles and pile groups.
Technology Press. J. Geotech. Engng Am. Sot. Civ. Engrs 108, No. 12,
Chameau, J. L. A., Alteschaem, A., Michael, H. L. & 1605-1623.
Yao, J. T. P. (1983). Potential applications of fuzzy Peck, R. B. (1958). A study of the comparative behavior
sets in civil engineering. Int. J. Man-Machine Stud. of friction piles. Highw. Res. Spec. Rep. 36, l-78.
19, 9-18. Peck, R. B. (1961). Records of load tests on friction
Cox, W. R., Kraft, L. M. & Verner, E. A. (1979). Axial piles. Highw. Res. Spec. Rep. 67, 1-418.
load tests on 14-inch pipe piles in clay. Proc. 21th Pelletier. J. H. & Dovle. E. H. (1982). Tension cauacitv
Offshore Technol. Cot&, Houston 2, 1147-l 158. in s&y clay-beia pile tes;. P&c. 2nd Int. Coni:
Darragh, R. D. & Bell, R. A. (1969). Load tests on long Numer. Meth. O#shore Piling, Texas, 163-182.
bearing piles. Performance of deep foundations, pp. Raymond International. (1982). Raymond step taper
41-67. STP 444. Philadelphia: American Society of piles. Brochure ST-82,25. Raymond International.
Testing and Materials. Rigden, W. J., Petit, J. J., St John, H. D. & Poskitt, T. J.
Eberhart, R. C. & Dobbins, R. W. (1990). Neural (1979). Developments in piling for offshore struc-
network PC tools: a practical guide. San Diego: Aca- tures. In Behaviour of offshore structures, pp. 29-43.
demic Press. London: Institution of Civil Engineers.
Endley, S. N., Ulrich, E. J. & Gray, J. B. (1979). A study Rumelhart, D. E., Hinton, G. E. & Williams, R. J.
of axial pile load tests. Am. Sot. Civ. Engrs Natn. (1986). Learning internal representation by error
Conv., Atlanta, Preprint 3766. propagation. In Parallel distributed processing, vol. 1
Flaate, K. & Selnes, P. (1977). Side friction of piles in (eds D. E. Rumelhart & J. L. McClelland). Cam-
clay. Proc. 9th Int. Co@ Soil Mech., Tokyo 1, 517- bridge: Massachusetts Institute of Technology
522. Press.
Flood, I. & Kartam, N. (1994). Neural networks in civil Rumelhart, D. E. & McClelland, J. L. (1986). Parallel
engineering. I: Principles and understanding. J. distributed processing-explorations in the micro-
Computng Civ. Engng Am. Sot. Civ. Engrs 8, No. 2, structure of cognition, ~01s. 1 and 2. Cambridge:
131-148. Massachus&ts institute of Technology Press. _
Fox, D. A., Sutton, V. J. R. & Oksuzler, Y. (1976). Semple, R. M. & Rigden, W. J. (1986). Shaft capacity of
North Sea platform piling development of the driven pipe piles in clay. Ground Engng 19, No. 1,
Forties Field piles from West Sole and Nigg Bay 11-17.
714 Gl
Smith, M. (1993). Neural networks for statistical model- Vijayvergiya, V. A. & Focht, J. A., Jr. (1972). A new way
ing.
_, New York: Van Nostrand Reinhold. to predict capacity of piles in clays. Proc. 4th Ofi
Stermac, A. G., Selby, K. G. & Devata, M. (1969). shore Technol. Conf, Houston 2, 865-874.
Behavior of various types of piles in stiff clay. Proc. Woodward, R. J., Lundgren, R. & Boitano, J. D. (1961).
7th Int. Conf Soil Mech., Stockholm 2, 239-245. Pile loading tests in stiff clay. Proc. 5th Int. Conf:
Togrol, E. (1973). Bearing capacity by load test. Proc. Soil Mech., Paris 2, 177-184.
8th Int. Conf Soil. Mech., Moscow 2, 231-236. Yeh, Y. C., Kuo, Y. H. & Hsu, D. S. (1993). Building
Tomlinson, M. J. (1970). Some effects of pile driving on KBES for diannosine. PC nile with artificial neural
skin friction. In Behaviour of piles, pp. 107-114. network. J. Computng Ciu. Engng Am. Sot. Ciu.
London: Institution of Civil Engineers. Engrs 7, No. 1, 71-93.
United States Army Corps of Engineers (1979). Test pile
report for Lock D Project. Tennessee-Tombigbee
Waterway. USACE.