Вы находитесь на странице: 1из 10

Advances in Engineering Software 34 (2003) 559568 www.elsevier.

com/locate/advengsoft

Estimation of consolidation settlement caused by groundwater drawdown using articial neural networks
T. Kerh*, Y.G. Hu, C.H. Wu
Department of Civil Engineering, National Pingtung University of Science and Technology, Pingtung 91207, Taiwan, ROC Received 12 December 2002; accepted 20 March 2003

Abstract The method of back-propagation neural networks was employed in this study to develop a model for estimating the consolidation settlements caused by transient or long-term groundwater drawdown along the main Red line sections of Kaohsiung mass rapid transit, Taiwan. The available on-site boring test data including soil void ratio, groundwater drawdown depth and total unit weight of soil were taken as the input parameters. Three neural networks models with different combinations of these inputs were examined, which showed that the groundwater drawdown depth was the dominating factor to affect the consolidation settlement. The estimated results were compared with theoretical results, and statistical t-tests were performed to enhance the reliability of neural networks model. From the overall estimated results, the potential hazardous regions were identied along the Red line sections. It was found that there exists either a higher initial groundwater level or a thicker low compressibility clay layer, which might be the major reasons to cause the severely settlements, and must be carefully evaluated during the construction in these regions. q 2003 Elsevier Ltd. All rights reserved.
Keywords: Kaohsiung mass rapid transit; Back-propagation neural networks; Groundwater drawdown; Consolidation settlement

1. Introduction Natural disasters such as earthquake, typhoon and ooding sometimes are inevitable as they usually occur in a huge scale and involve in many uncontrolled reasons, which are difcult to predict correctly. In contrast, manmade disasters such as over development of land, different kinds of environmental pollution, and overdraft of groundwater may occur in many situations, but they are predictable and the negative impacts can be reduced by available up to date methods. Therefore, the disaster prevention technology is become one of popular research areas in the eld of civil engineering in recent years. For instance, the consolidation settlement caused by a transient or a long-term groundwater drawdown is a very important issue, which must be considered during the construction of engineering projects, as it may cause serious damage on the structures and may endanger the safety of human lives in the neighborhood of construction region. As shown in Fig. 1, Taiwan is an island of about 400 km long from tip to tip and 130 km wide at its broadest points,
* Corresponding author. Tel.: 886-9333-25118; fax: 886-877-40122. E-mail address: tfkerh@mail.npust.edu.tw (T. Kerh). 0965-9978/03/$ - see front matter q 2003 Elsevier Ltd. All rights reserved. doi:10.1016/S0965-9978(03)00053-X

where the second largest modern city, Kaohsiung, is ideally situated on the southern coastline. This city is the foremost industrial center and largest international port in Taiwan. There are over 1.5 million populations lived in this city, and many of traditional architectures and modern high-rise buildings are built in the metropolitan area [1]. Now in Kaohsiung city, the major engineering project in progress is the construction of Kaohsiung mass rapid transit (KMRT) system. The original plan of this transportation system consists of four lines, with a total of 77.7 km, where the main line and under construction is the Red line. Through the region with the highest density of population and building in the city, this main Red line has a length of 28.3 km in north south direction, that including 19.8 km underground works and 8.5 km trestle works. The Red line is expected to accomplish all the construction works in the year of 2007, which will have a signicantly inuence on the people living in this city. Since this major engineering project includes a large portion of underground works, the estimation of consolidation settlement resulting from groundwater drawdown during the period of construction is a necessary task to prevent unwanted manmade disasters. Based upon the on-site boring test report on major sections of Red line from local government and engineering

560

T. Kerh et al. / Advances in Engineering Software 34 (2003) 559568

Fig. 1. Sketch of the Red line sections in the investigation area.

consulting company [2], the groundwater levels (GWL) along the line sections are mostly between 2 and 7 m in average under the ground surface. Because the engineering foundation excavation in this area must be able to reach 17 20 m in depth, thus a temporary or a long-term withdrawing of groundwater is the requirement to smooth the construction. To solve this problem, the scientic methods such as coupled and uncoupled consolidation models, and numerical simulations may be used to evaluate the consolidation settlement due to groundwater drawdown in local area of Taiwan [3,4]. However, regarding the application of using articial neural networks (ANN) in this important problem is still rarely seen up to the present time. Additionally, because the neural computing has the advantage of using eld or experimental data directly without simplifying assumptions, and the nonalgorithm method is capable of executing massive computation in a parallel environment. This approach has been extensively developed in recent years, and applied to solve various civil engineering problems [5 13]. Therefore, the purpose of this study is to develop a model for estimating the consolidation settlements from ground

level to the excavation surfaces along the main Red line sections of KMRT by using the method of back-propagation neural networks. The available on-site boring test data including soil void ratio, groundwater drawdown depth and total unit weight of soil are taken as the input parameters, and three combinations of these inputs are used to obtain a relatively better estimating model. The estimated results and available theoretical results from underground 1 3 m, with statistical t-tests, are compared to enhance the reliability of neural networks models. The distribution of surface settlements along the Red line sections may be seen clearly from the overall estimating results and post-plots, and the obtained information may provide a useful reference for evaluating the potential hazardous regions due to construction works.

2. Approach of articial neural networks The well-known back-propagation neural networks, which uses a specic learning law for updating

T. Kerh et al. / Advances in Engineering Software 34 (2003) 559568

561

the weightings of each layer in accordance with the errors from the network output, is frequently applied to solve various types of engineering problems due to its simplicity [14 16]. The multi-layered neural networks may include input layer, hidden layer and output layer. For the basic algebraic equation of each layer, it may be written as X  Yj F 1 Wij Xi 2 uj where Yj is the output of neuron j; Wij represents the weight from neuron i to neuron j; Xi is the input signal generated for neuron i; uj is the bias term associated with neuron j; and the nonlinear activation function F is assumed to be a sigmoid function as Fx 1=1 e2x ; which will make the operating process continuous and differentiable. The back-propagation neural networks are basically a gradient descent method, and two parameters called as the learning rate h and the momentum factor z; are usually introduced in the iterative calculation process as the following equation
n n21 n21 Wij Wij hdj Xi zDWij

input layer, two hidden layers and one output layer. Furthermore, the learning rate h 0:9 and the momentum factor z 0:9 are chosen in the training process due to a relatively better tendency of convergence based upon computational experiments in this case study.

3. Theoretical background of estimating consolidation settlement Regarding the fundamental theoretical background of consolidation in soil mechanics, the Terzaghi theory is properly the most common used in today [17]. The basic concept of consolidation is that when the void water in saturated soil is decreased for some reasons, but the reducing space between soil particles is not able to replace by the air. Usually, This behavior is occurred in a cohesive soil such as clay which has a low permeability and slow drain ability. Originally, the increment of stress due to transient loading is supported by the void water pressure, once the void water pressure no longer exists, because the external loading will increase the effective stress in soil particle, as that it makes the soil particles become more condense locally, and at last it causes the phenomenon of consolidation settlement globally. Suppose that the consolidation coefcient of soil and the changing stress caused by groundwater drawdown are known in advance, and then the amount of settlement may be calculated by applying the following equation S Cc H =1 e0 logp0 Dp=p0 4

n21 denotes the where dj is the error signal for neuron j; DWij adjusting weights between neurons i and j; meanwhile the symbols n and (n 2 1) are the current and the most recent training step, respectively. In general, the learning rate and momentum parameters may accelerate convergent speed and may smooth oscillations in weight corrections during training. But these two parameters require some computational experiments to determine their better values for tting in the case study. Furthermore, to evaluate the effectiveness of neural networks model, the coefcient of correlation R may be used and dened as follows " #1=2 m m m X X X 2 2 R xi 2 x xi 2 x yi 2 y 3  i yi 2 y  i = i i i1 i1 i1

where xi and x  i are the theoretical results and its averaged values, respectively, yi and y  i are the estimated and its averaged values, respectively, and m denotes the number of data in the analysis. Followed the success of using back-propagation neural networks to estimate peak ground acceleration due to strong motion in main lines of KMRT [1], in this study, the similar neural networks models are trained according to the actual boring test data around the sections divided by train stations. There are three input parameters including soil void ratio, groundwater drawdown depth and total unit weight of soil are used for the models. For these models, the model 1 (M1) uses each of the parameters, the model 2 (M2) uses the combinations of two parameters, and the model 3 (M3) takes the whole parameters as the inputs, respectively. Additionally, in neural networks models M1 and M2, there is one input layer, one hidden layer and one output layer, but in model M3, there is one

where S is the amount of consolidation settlement, Cc the consolidation coefcient; H the height of soil layer before consolidation, e0 the initial void ratio, p0 the effective stress of cover soil, and Dp the increment of stress. In each parameter, the consolidation coefcient Cc represents the slope of original soil consolidation curve, which may be determined from laboratory tests. However, according to the on-site boring data in the present investigation area, the one-dimensional consolidation test is not sufcient to reect the soil characteristics, and thus a modication based on the empirical formulas [18,19] is used as follows Cc 0:009LL 2 10 5

where LL is the liquid limit for clay layer. But for the clay layer mixing with silt sand, the formula is Cc 0:3e0 2 0:27 6

where e0 can be taken directly from on-site test as it is nearly the same as the laboratory result in this study. Since the present Red line of KMRT is to be excavated by using tunneling boring method or shield driving method, the withdrawing of groundwater is inevitable in some sections, and will cause the variation of stress in soil. In general, the following equation may be used to evaluate

562

T. Kerh et al. / Advances in Engineering Software 34 (2003) 559568

the increment of stress as Dp gw h 7

where gw is the water unit weight, and h is the groundwater drawdown depth. Furthermore, the deeper the initial GWL, the more the initial effective stress of the cover soil, and that will affect the amount of settlement. By inserting Eq. (7) into Eq. (4), the relationship of consolidation settlement and groundwater drawdown depth then becomes h p0 =gw 10S1e0 =Cc H 2 1 8

where the denition of each parameter are mentioned as above. Theoretically, the above equations may be used to estimate the amount of consolidation settlement, but it requires detailed soil and groundwater parameters, which may not be found available from boring tests for all cases. In addition, the groundwater drawdown depth needs to be evaluated according to some of on-site geological conditions such as soil permeability coefcient, radius of water well, underground water pressure, tunneling radius and excavation area. Therefore, the determination of using effective coefcients and post calculations may require some assumptions and may be a tedious process, which makes the use of neural networks become an applicable tool in this important engineering problem, and the details will be discussed in Section 4.

4. Estimation of consolidation settlement by neural networks model The ANN have the advantage of learning past records characteristics to classify a model for predicting some of

important engineering problems. In this study, three neural networks models using different combination of input parameters are compared to choose a better model for generating the output under consideration. The three training parameters include soil void ratio e; groundwater drawdown depth h and soil total unit weight g can be found available from boring test data. To check the performance of neural networks model, three theoretical calculations of consolidation settlement for groundwater drawdown depth 1, 2 and 3 m, respectively, are taken for comparison. Note that these theoretical results may also be found available in previous report [20], which are calculated from Eqs. (4) (8) as discussed in Section 3. Values shown in Table 1 are the coefcient of correlation for different combination of input parameters. It can be found that in neural networks model M1, the groundwater drawdown depth has the highest averaged coefcient value R 0:7598 for all line sections. In neural networks model M2, the averaged coefcient value does not improve very signicantly, some sections even have a lower coefcient value than that of the single parameter case. In this combination model, the soil void ratio and total unit weight exhibit less important R 0:2606 in average; but the groundwater drawdown depth seems still the dominating factor. Finally in neural networks model M3, which uses three parameters as the input and to the best result, where four sections have the coefcient value over 0.9, and R 0:8755 in average for all sections. Therefore, as long as three input parameters are arranged properly, the common effect of each parameter may become signicant to achieve a better output, and this model will be chosen for further analysis. Fig. 2 shows the approaching tendency between consolidation settlement according to the Terzaghi theory and

Table 1 Coefcients of correlation for different neural networks models and the combination of input parameters NN Model Section M1 e R1 R2 R2 R3 R3 R4 R4 R5 R5 R6 R6 R7 R7 R8 R8 R9 R9 R10 R10R11 R11R12 R12R13 R13R14 R14R15 R15R16 Average 0.2827 0.0465 0.3609 0.3776 0.2699 0.4830 0.2948 0.5275 0.2746 0.7045 0.5264 0.3469 0.3975 0.2970 0.2406 0.3596 h 0.8743 0.6249 0.6708 0.6721 0.7228 0.7141 0.7968 0.7016 0.7301 0.6293 0.8710 0.8222 0.9663 0.8009 0.8009 0.7598 M2 M3 e&h&g h&g 2 0.7347 0.4877 2 0.6943 0.5381 0.7288 0.4639 0.5671 2 0.6131 0.7522 0.6301 0.7177 0.8291 2 0.7977 0.6428 2 0.8142 0.6674 e&g 2 0.1505 0.0151 2 0.3608 2 0.2123 0.2829 0.3131 0.1661 2 0.4235 0.2873 0.7045 0.2146 0.3459 2 0.3405 0.0158 2 0.0755 0.2606 0.9225 0.9131 0.8275 0.8512 0.8008 0.8042 0.8857 0.9693 0.8876 0.9348 0.8474 0.8425 0.8842 0.8821 0.8799 0.8755

g
2 0.3307 0.1681 2 0.1079 2 0.9776 0.4240 0.4754 2 0.8452 2 0.2685 0.4519 2 0.5844 0.3638 0.3428 0.0937 0.3357 0.0732 0.3895

e&h 0.6591 0.6075 0.6186 0.6711 0.8012 0.6637 0.5808 0.8138 0.8889 0.9306 0.7233 0.8436 0.9701 0.6419 0.8730 0.7525

The averaged values are calculated from absolute value of correlation coefcients.

T. Kerh et al. / Advances in Engineering Software 34 (2003) 559568

563

Fig. 2. Approaching tendency between theoretical results and ANN estimations.

estimating results based on the M3 neural networks algorithm from every available boring test data for three groundwater drawdown depths. This scattering diagram shows the value of correlation coefcient is 0.8791, which implies that the estimated result is in reasonable agreement with the theoretical result. Next, in each groundwater drawdown depth, the comparison of ANN estimations and theoretical calculations along the Red line sections is shown in Fig. 3 for (a) h 1 m; (b) h 2 m; and (c) h 3 m; respectively. Note that all the 15 Red line sections are distinguished from train stations, and the results in each section are calculated by averaging the obtained results from boring test information in the neighborhood of the line section. These comparison plots with statistical t-test results t 22:093; t 1:634 and t 1:215 for the above three groundwater drawdown depths, respectively, exhibit that the estimations are in good agreement with the theoretical calculations, as their test values are within the acceptance interval 22:145 , t , 2:145 under degrees of freedom n 14 and signicance level a 5%. These obtained results conrm that a satisfactory estimation of consolidation settlement is achievable from the learning ANN model, which may provide a condence for us to estimate the results on the way down to excavation surface of the Red line sections. Since the excavation depths along the main Red line will around 17 20 m according to the construction plan, in this study, the GWL is assumed to be dropped down till 20 m below ground surface for evaluating the amount of settlements. Thus, the highest groundwater drawdown depth at each section is by deducting the initial GWL from the 20 m excavated depths. However, for observing

the variations of settlement due to groundwater drawdown, two cases of 5 and 10 m drawdown depths are applied to make comparison with the case of highest drawdown depth. Now by taking the effective soil ratios and total unit weights input to the M3 neural networks model as trained previously, the estimating consolidation settlements are obtained accordingly. Exhibited in Fig. 4 is the bar chart result for three groundwater drawdown depths. It can be seen that in the case of 5 m drawdown depths, there are about 10 cm settlement for most of sections, but more settlements are found in sections 10, 14 and 15, with the settlements 19.9, 20.0 and 19.4 cm, respectively. When the groundwater drawdown depth becomes 10 m, all the settlements increase larger and the averaged value is increased to about 293%, which is almost three times of the previous case. In this case, sections 11, 12 and 15 have lager settlements owing to thicker soft clay layers existed in these sections. For the highest drawdown depth, the averaged settlement is about 66 cm for all sections. The most settling regions are in sections 6, 11, 12, and 15, which nearly reach up to 1 m of settlements. From the tendency of GWL, it can be seen that the difference of settlement in section 11 is very small between the cases of 10 m and the deepest groundwater drawdown depths, owing to the existence of a deeper initial GWL. But in sections 6, 14 and 15, as they have a higher initial GWL, which may reect that the more layers of soil are affected by the groundwater drawdown depth, so that the stress in soil is increased which results in a larger consolidation settlement. To view more clearly overall in the related actual geometrical locations, now by taking the results of

564

T. Kerh et al. / Advances in Engineering Software 34 (2003) 559568

Fig. 3. Comparison of theoretical calculations and ANN estimations along the Red line sections.

settlements to become a three-dimensional plot as shown in Figs. 5 7 for three groundwater drawdown depths. It can be observed that the consolidation settlements at each section have a tendency of high

variation, and the potential hazardous regions may be identied from the case of highest groundwater drawdown depth. The Red line sections 6, 11, 12 and 15 have relatively larger settlements due to a higher initial GWL

Fig. 4. Estimations of consolidation settlement at Red line sections for three different groundwater drawdown depths.

T. Kerh et al. / Advances in Engineering Software 34 (2003) 559568

565

Fig. 5. Three-dimensional plot of estimated consolidation settlement for 5 m groundwater drawdown depth.

Fig. 6. Three-dimensional plot of estimated consolidation settlement for 10 m groundwater drawdown depth.

Fig. 7. Three-dimensional plot of estimated consolidation settlement for the deepest groundwater drawdown depth.

566

T. Kerh et al. / Advances in Engineering Software 34 (2003) 559568

Fig. 8. Averaging soil proles from boring tests around the Red line sections, (a) section 6, (b) section 11, and (c) section 12.

T. Kerh et al. / Advances in Engineering Software 34 (2003) 559568

567

or a thicker clay layer existing underground, so these sections must be carefully evaluated during the construction. As the Red line is situated in the recent alluvial sediments, and the distributions of various soil textures may include mainly sandy silt, low compressibility clay and their mixtures. To view more clearly in these potential hazardous sections, the averaging soil proles from on-site boring tests in the vicinity of each section are shown in Fig. 8 [21]. Basically, the experimental results show that loose to medium tight sands and clays are distributed in these regions. But most importantly, it can be found that the low compressibility clay layers either occupy a large portion of the soil prole or present in the top of soil prole in these sections. Note that in section 6, the rock is in a very deep area, which is not even found at 89 m depths. In addition, the changing of soil layer is very signicant in section 15 from several boring holes information around this site, which is difcult to present a typical soil prole for this section, but this uneven soil distribution may also be one of the reasons for a larger consolidation settlement.

a thicker low compressibility clay layer in soil prole, those might be the major reasons to cause the severe settlements in these sections. Actually, to determine the depth of groundwater drawdown might be a tedious task, which relied upon various on-site geological conditions such as soil permeability coefcient, radius of water well, underground water pressure, tunneling radius and excavation area. However, the information might not be found available from boring tests for all cases. Therefore, although the present neural networks model neither used many training data nor taken other parameters as the input for improving the estimations, the obtained results have proved that the neural networks was an applicable tool in this type of problem. Overall in this study case, the distribution of surface settlements along the Red line sections could be seen clearly from the post-plots. In practice, the information might have provided a useful reference for evaluating the potential hazardous regions due to construction works.

5. Summary and conclusion The evaluation of potential hazardous regions is a crucial issue, which must be taken into account in any civil engineering project. In the present study, the method of back-propagation neural networks has been applied to develop a model for estimating the consolidation settlements caused by transient or long-term groundwater drawdown along the main Red line sections of KMRT. The available on-site boring test data including soil void ratio, groundwater drawdown depth and total unit weight of soil were taken as the input parameters. Three neural networks models with different combinations of these inputs have been compared from correlation coefcients to obtain a relatively better estimating model for further analysis. It could be found that the groundwater drawdown depth was the dominating factor to affect the consolidation settlement. The estimated results by neural networks and the calculated results by Terzaghi theory for the cases of 1, 2 and 3 m groundwater drawdown depths have been checked for the performance with statistical t-tests, to enhance the reliability of neural networks model. Then, from the estimated results of 5, 10 m and the deepest groundwater drawdown depths, it could be observed that the consolidation settlements at each section exhibited a high variation tendency, and the potential hazardous regions might be identied along the Red line sections. The results have shown that there were relatively larger settlements occurred in sections 6, 11, 12 and 15. From the boring test information around these regions, it might be found that there exists either a higher initial GWL or

Acknowledgements The authors wish to thank Bureau of Kaohsiung Rapid Transit for providing information and boring test data of the Red line system.

References
[1] Kerh T, Chu D. Neural networks approach and microtremor measurements in estimating peak ground acceleration due to strong motion. Adv Engng Software 2002;33:73342. [2] CTCI Corporation., Boring test report on major sections of red line, KMRT. Bureau of Kaohsiung Mass Rapid Transit, vol. 1.; 1991. pp. 1202. [3] Chang KR, Lin ML. Analysis of consolidation settlement caused by withdrawing of multi-well water at Meiliao area. Master Thesis of Civil Engineering, National Taiwan University, Taiwan; 1999. [4] Duh FL, Lue CJ. A study on the problem of consolidation settlement due to withdrawing of water. Master Thesis of Civil Engineering, Chung Fua University, Taiwan; 1992. [5] Adeli H. Neural networks in civil engineering: 19892000. ComputAided Civil Infrastruct Engng 2001;16:126 42. [6] Arslan A, Ince R. Neural network-based design of edge-supported reinforced concrete slabs. Struct Engng Rev 1996;8:329 35. [7] Bodri L. Prediction of extreme precipitation using a neural network: application to summer ood occurrence in Moravia. Adv Engng Software 2000;31:311 21. [8] Flood I, Kartam N. Neural networks in civil engineering ii: systems and applications. J Comput Civil Engng 1994;8:14962. [9] Flood I, Kartam N. Neural networks in civil engineering i: principles and understanding. J Comput Civil Engng 1994;8(2):13148. [10] Liong SY, Lim WH, Paudyal GN. River stage forecasting in Bangladesh: neural network approach. J Comput Civil Engng 2000; 14(1):18. [11] Pandey PC, Barai SV. Multilayer perceptron in damage detection of bridge structures. Comput Struct 1995;54(4):597608.

568

T. Kerh et al. / Advances in Engineering Software 34 (2003) 559568 editor. Advances in engineering computational technology. Edinburgh: Civil-Comp Press; 1998. p. 173 87. Terzaghi K. Theoretical soil mechanics. New York: Wiley; 1943. Das BM. Advanced soil mechanics. USA: Taylor and Francis; 1998. p. 399453. Das BM. Principles of soil dynamics. USA: Pws Kent Publishing Company; 1992. p. 301455. Huang YS, Tsai KJ, Wang HY. Analysis of potential geological disaster at the red line sections of Kaohsiung mass rapid transit system. Master Thesis of Civil Engineering, National Pingtung University of Science and Technology, Taiwan; 2001 Kerh T, Chen LY, Lee JH. Analysis of soil distribution and period measurement along the main lines of Kaohsiung mass rapid transit. J Civil Engng Technol 1996;4:81100.

[12] Tsai CP, Lee TL. Back-propagation neural network in tidal-level forecasting. J Waterway Port, Coastal, Ocean Engng 1999;125(4): 195202. [13] Zeng P. Articial neural network computing in structural engineering. In: Topping BHV, editor. Developments in neural networks and evolutionary computing for Civil and Structural Engineering. Edinburgh: CivilComp Press; 1995. p. 37 50. [14] Hirose Y, Yamashita K, Hijiya S. Back-propagation algorithm which varies the number of hidden units. Neural Networks 1991;4: 616. [15] Villiers J, Barnard E. Back-propagation neural nets with one and two hidden layers. IEEE Trans Neural Networks 1993;4(1):13641. [16] Waszczyszyn Z. Some new results in application of back propagation neural networks in structural and civil engineering. In: Topping BHV,

[17] [18] [19] [20]

[21]

Вам также может понравиться