Using BPN Method for EstimatingCementTake ofGrouting

Chau-Ping, Yang

Department of Civil Engineering and EngineeringInformatics, Chung-HauUniversity,

30 Tung Shiang, Hsinchu, Taiwan, 30067, E-mail address:

Keywords: Dam foundation; Grouting; Cement take; BPN

Abstract

Using cement grout to improve bedrock has been quite common. However, the cost for cement grout is the most difficult one to estimate.This study adopted the Back-Propagation Neural network (BPN) to analyze the grouting construction data of the Li-Yu-Tan dam, in order to estimate the cement take needed.The samples analyzed included data from 3,532 grout sections.The data from the first half of the groutingconstruction were used to derive the parameters of the predictive schemes, and then the second half of the groutingconstruction’s data were used to test the accuracy of those schemes. The accuracy level estimated by BPN on gross cement take was 75.3%. It was higher than the original design level of 43.4%.

1. Introduction

The bedrock inherently has discontinuities such as faults, folds, beddings, joints, and fractures, whichare the major factors that affect the engineering properties of rock foundationssuch as permeability, shear strength, and deformation. When a dam is located on bedrock that has unknown discontinuities, the underlying foundation needs to be improved to raise its engineering properties and ensurea watertight reservoir. Using cement grout to improve bedrock has been quite common, and there are numerous examples of its application infoundation improvement [1,2,3,4,5,6]. However, since the dam foundation is below the surface of theground,the cost for cement grout is the most difficult one to estimate. The cost of cement grout mainly includes the operational part and the material part. Thecostof materialsis calculated based on the cement take. Then the cost of the groutingoperation is determined based on the material’scost. Therefore, it is necessary to study various methods to estimate the cement take of the grouting based on actual construction data. The methods commonly used are mean method and linear regression method [7, 8].

In general, the status of the discontinuities in the dam foundation is indirectly expressed by the determined from the Lugeon tests. This information can also be used to design the water to cement ratio and the injection pressure used in the grouting process. Eq.(1) is the definition of the.

== (1)

Where is the water take (), is the

standard injection pressure (981), is the injection time (), is the injection pressure used (), is the length of grout section ().Generally speaking, if a dam foundation has a high, it will have more discontinuities with high permeability and more cement take is needed for the grout improvement.

The is the best physical parameter to express the status of discontinuities in a dam foundation. Theoretically, it is quite difficult to define the relationship between cement take and the [7,8]. Additionally, when researchers estimate the cement take needed for a new dam foundation from past experiences, they still encounter the problems of different geological properties for the proposed dam site. For example, the cement take designed for the improvement of the foundationof theLi-Yu-Tan dam, Miao-Li County,Taiwan, was 50. However, the average reading of cement take from the construction recordswas 115[9]. This difference resulted in a doubling of the amount of gross cement take from what was required in the stage of original design.This experience illustrates the difficulty in cement take estimation.

This study adopted the Back-Propagation Neural network (BPN) to analyze the construction data from the grout-curtain improvement of the Li-Yu-Tan dam’s foundation, and indicate how to estimate the cement take needed.The dam is located at the upper stream of the Jing-San brook, a tributary of the Da-An river. The dam is a zone-type-earth-dam with a height of 96, a bottom width along the foundation of the river of about 500and a gross volume of 3,700,000. The major terrain includes gravelly terra rossa and some riverbank outcrops. There are no faults or obvious folds on either side of the river.The major discontinuitiesin the foundation of the dam site aredozens of developed shear zones. Most shear zones are distributed in the right side of the abutments of the dam with slips of more than 2~5[10].

2. Factorsaffecting cement take

Theoretically, there are many factors that affect the cement take needed for improving dam foundations. Moreover, since some factors may have combined effects, it is not possible to clearly define the role of each factor. Some factors that can be categorized or quantified are the strata, zone of dam foundation, depth of grout section, injection pressure, and the.

2.1 Strata

This category covers properties such as the rock layers, the nature of discontinuities, the rock strength, the mineral components, and the cementation. Shallow bedrock tends to have a high density of cracks or openings and is subjected to grout leakage and hole’s collapse.If a rock foundation has little strength, the grout hole will be less independent. The disadvantages of bedrock mentioned above increase the amount of cement take needed for grout improvement.

As shown in Fig.1, the stratain thedam sitevicinity are northeastwards and meet the river valley vertically at 28~34 degrees. All the strata are leaning towards the upper stream at 30~34 degrees. The strata of the Li-Yu-Tan dam’s foundation include clean sandstone (CS), mudstone (MS), and alternations with sandstone and shale (AL). The major formation of clean sandstone contains quartz sand, which has a tensile strength of about 1,050and a hydraulic conductivityof about . However, since quartz sand has a poor cementation quality, the seepage pathsare more likely to cause a loss of fine material. Mudstone contains different amounts of mud; therefore, its tensile strength ranges from 1,140 to 2,010, and the average hydraulic conductivity is. Alternations with sandstone and shale have intertwined clean sandstone and shale or mudstone and shale in small alternating thickness. The thickness of mud accumulation between layers can reach 30.On the surface layer, seepage paths can form that cause deterioration of the shale into fragments or even seams.

2.2 Zone of dam foundation

When the overburdenof ground is relieved, riverbanks will move inwards, and tensile fractures will occur in the banks. This phase results in more cracks on the upper half of the dam’sabutments and induces greater permeability. For this reason, the cement takes needed for the grout improvement in the right zone, left zone, and the valley are different. This research has divided the dam foundation into the riverbed, the left upper zone, the left lower zone, the right upper zone and the right lower zone, as shown in Fig.2and Fig.3, according to the tunnel locations for the grout-curtain construction. However, because the riverbed has been dug to the level of fresh bedrock with a permeability lower than 10, there are only a few in-place grout holes. Thus, the analytical extent of this research covers only the left upper zone, the left lower zone, the right upper zone, and the right lower zone.The shaded part in Fig.3 is the outcome of the grout-curtain in the Li-Yu-Tan dam’s foundation. For the shallower parts, grouting can be performedfrom the top, but, in the deeper areas, the grouting will have to be performedfrom tunnels.

2.3 Depth of grout section

In a rock layers deeper into the underground,the cracks are narrow and comparatively do not take in groutbecause of the greater tectonic stresses in lower elevation. When the tectonic stress is taken into consideration, the depth of the grout section is considered as one of the factors that affect cement take. As to the grout-curtain construction in the Li-Yu-Tan dam, the diameter of the grout holes was 3.8and the greatest vertical depth of a grout hole was limited to 50. Inside of each grout hole, there were several grout sections, and the grout process was conducted from the bottom to the top ofthegrout hole. If the depth of the grout section was smaller than 30, the grout section length was 5. When the depth of a grout section was greater than 30, the section length was 10.

2.4 Injection pressure

The injection pressure is the major technical factor affecting cement take. Theoretically, the injection pressure should be smaller than the tectonic stress corresponding to the depth of a grout section, which is obtained from the hydraulic fracturing test. Moreover, the injection pressure should be smaller than the tensile strength of the rocks [11,12]. In Taiwan, dam engineersconsider that the injection pressure is determined based on the principle of additional pressure increasing about30 per meter depth.The injection pressure adopted for the grout-curtain construction of theLi-Yu-Tan dam was 150 to 1200from top to the bottom of the grout hole [13].

2.5

The is the only physical parameter that the researcher could obtain to evaluate the multiple factors that affect cement take. This value shows the degree of permeability in the dam foundation. Basically, in grout improvement, a dam foundation that has a high requires more cement take.

3. Data analysis

In the Li-Yu-Tan dam’s grout-curtain construction, the grout holes were of the split-spacing type. Split-spacing means that the grout holes were arranged in the sequence of primary holes, secondary holes, tertiary holes, and quaternary holes. Supplementary holes may be added to enhance the locations with more discontinuities in the bedrock or near the holes that required more cement take. Basically, the arrangement of grout holes was based on the quality of bedrock. The grout holes were arranged at intervals of1to 3. When the grouting process of a specific hole lasts for 60 minutes, but the amount of cement take does not reach 70, the grouting for this section should be stopped. Finally, the drill inspection holes used for performing the Lugeon test to check the permeability of the dam foundation wereimproved. The process of grouting in each grout section was arranged in the following sequence: drilling, washing, water testing, and grouting. During water testing, theLugeon tests need to be performed to obtain .

Table 1lists the data analyzed for 469 grout holes and 3,532 grout sections. Each grout section had data such as zone, sequence, hole depth, length of grout section, rock nature,, injection pressure, and cement take. All of the data were collected from the inspection chart of the grout-curtain construction for the Li-Yu-Tan dam in 1993. Then, all the data were entered into an Excel application program for calculations before the BPN analysis began.

For the convenience of analysis, this study has adopted the symbol to represent the of a specific grout section. In addition, because the lengths of the grout sections analyzed were not the same (between 5 and 10), the cement take of a grout section was divided by its length to obtain the cement take per unit length ().There were three reasons to use cement take instead of cement mortar take to define . First, the voids in the cracks were filled by solid cement. Secondly, the major material expense in grout construction is the quantity of cement. Thirdly, many documents related to groutingrefer to cement take in place of cement mortar take [14,15].

4. BPN method

The BPN is a branch of artificial neural networks (ANN). The growing interest in ANN among researchers is due to its excellent performance in learning ability, fault tolerance, pattern recognition, and the modeling of nonlinear relationships especially involving a multitude of non-digital variables in place of conventional techniques. Generally, a complex domain is characterized by a number of interacting factors. Yet, such factors are often incomplete or unreliable. If ANN is used to analyze complex engineering systems, it can alleviate noise interference and raise the accuracy level of the analysis. ANN has been widely applied to research in the field of geotechnical engineering in recent years [16,17,18].

Huangand Wanstedt [19] applied BPN to the categorization of rocks and found that the categorizing ability of BPN was much better than statistical methods. Additionally, a conventional method for modeling the stress-strain behavior of soil is the constitutive law. However, it is characterized by the difficulties in obtaining correct parameters, conducting mathematical calculations, and the oversimplification of the hypothesis. In a quite different way of research thinking the constitutive law was replaced with BPN to simulate the stress-strain behavior of soils [20,21,22].

4.1 Mechanism of BPN

The typical architecture of BPN used in this study is shown in Fig.4. The input layer uses linear transfer functions to handle the input variables in the network. The number of processing elements in the input layer depends on the problem. In the hidden layer, it learns how each processing element in the input layer affects the others through association of the connection weights. In the output layer, an S-shaped sigmoid transfer function is used to handle output variables to make the domain to be [0, 1]. The number of processing elements in the output layer depends on the problem.BPN learns by modifying the connection weights of the elements in response to the errors between the actual output values and the target output values. This is carried out through the gradient descent on the sum of squared error for all the training patterns.

The learning algorithm of BPN requires the following steps:

  1. Use the connection weight to show the correlation between the input variable and each processing element. Meanwhile, biases and activity function value will come out. Then, convert the value to either the target output value in the hidden layer and to the target output valuein the output layer.
  2. As to the processing elements in the output layer, useand the actual output value to calculate the offset. The calculation of the processing elements in the hidden layer also adopts,andto calculate the offset.
  3. In the input layer and the hidden layer, use the learning rate, and to calculate the correction value of theconnection weight. In the hidden layer and the output layer, use the learning rate,andto calculate. Then, update thein each processing element to complete the learning of one cycle.
  4. Repeat the computation described above until convergence or approximately 3,000 learning cycles are reached.

The BPN software used in this research was PC-Neuron, written in language [23]. With the assistance of the original programmer, a new subprogram was written to return to the target output value from the original domain [0, 1]. Then, this value was converted to a data file that Excel software can treat.

4.2Architecture of BPN for estimating

cement take

The input variables which needed to be fed into the BPN programwere the zone of thedam foundation, the type of rock layers, the injection pressure, the depth of grout section, and . The output variable was theof each grout section. Among these variables, both and are measured digital data and the others are represented by the classification codes. The codes of these input variables are listed in Table 2.

The learning algorithm of BPN can be divided into the training phase and the testing phase. Thelearning samplesfor these phases were collected from the first half of the grout construction in the four zones. The samples were randomly categorized into the training set and testing set in the first phase of data processing. The initial learning rate, the initial inertial factor, and the initial connection weight were set to be 5.0, 0.5 and 0.3 respectively. After a number of different hidden layers were tried, one hidden layer was used in the BPN model employed here. In the preliminary task, a network with different elements ranging from 2 to 8 in the hidden layer was trained for the same number of 3,000 cycles. It was found that the value of the average sum squared error () would reach the minimum value of 0.11 when the number of elements was equal to 5.Eq. (2) is used to calculate:

(2)

Where is the actual output value of processing element j in example p, is the target output value of processing element j in example p, is the number of example, is the number of processing element in the output layer.

So, a 551 network was set up as shown in Fig.5. The learning process was performed with a Pentium 586 computer, which took about 110 min. of CPU time. Finally, BPN was applied to the training set and produced the connection weights and biases. Then, the architecture of BPN for estimating cement take was built (see Fig.5).The accuracy for the training and testing data setsare described by the degree of correlation between output target values and actual values.The scatter of the target output values versus the actual output values were assessed using regression analysis and its degree of correlation of 0.82 was an acceptable one