Evaluation of soft soil foundation reinforcement effect of prefabricated building based on BP neural network

Abstract

The use of traditional reinforcement methods in construction sites often causes problems such as pore water pressure, which can not effectively form a solid foundation. Aiming at this problem, the evaluation model of soft soil foundation reinforcement effect of prefabricated buildings is established based on BP neural network, combined with the geological characteristics of soft soil and the elements of foundation reinforcement; The L-M algorithm is used to optimize the slow convergence problem of BP neural network, and finally its evaluation effect is verified through practical application. The results show that the strengthening effect of 1550 kN $\cdot$ m/m ${}^{2}$ is better than that of 2000 kN $\cdot$ m/m ${}^{2}$ with the more times of tamping for marine and river facies, and there is a positive correlation between the times of strengthening and the effect. At the same time, similar qualitative conditions also show that the greater the burial depth, the worse the reinforcement effect. When the overlying soil layer is soft, the shallow buried soil layer can be reinforced by laying a cushion to improve the overall reinforcement effect. The laws reflected in the final model output data are the same as those reflected in the construction, and the accuracy of the proposed model is up to 87%, indicating that the model has superior performance in the reinforcement effect evaluation.

Keywords

BP neural network prefabricated soft soil foundation reinforcement L-M algorithm pore water pressure

1. Introduction

As a typical labor-intensive enterprise, the green and sustainable development of the construction industry has become an important research topic in related fields in recent years [1]. As an important construction technology that changes the operation mode of the construction industry, prefabricated buildings completely break the traditional process of field production and construction in the traditional construction industry [2]. With the technical support of precise component design, it is assembled and constructed into batches in a factory assembly line, and is conveyed to the engineering spot for on-site assembling [3]. This method is more conducive to shortening the actual construction period of the project site, while improving the construction efficiency and reducing environmental pollution. But the disadvantage is that due to the separation of the manufacturing process and the construction site, it is difficult to carry out timely and effective on-site response in the face of different climatic and geological environments, resulting in insufficient engineering safety or practicability [4]. Therefore, it is necessary to evaluate the adaptability to different geological environments. This study takes the soft soil environment as the main perspective, and uses the BP neural network to establish a soft soil foundation reinforcement effect evaluation model for prefabricated buildings. By effectively evaluating the reinforcement effect, it offers a foundation for the progress of the construction of prefabricated buildings in the field construction process.

2. Related wsork

Research on prefabricated buildings has gradually enriched in recent years. Chen [5] established an evaluation model for the continuous scientific progress of prefabricated components from the point of view of ecological supply cycle. The model identifies and evaluates key supply nodes through a hierarchical clustering method. This means can achieve the sustainable management of the ecological supply chain cycle of prefabricated buildings. Yudina et al. [6] proposed a prefabricated building pile foundation construction method for multi-purpose buildings. The method takes affordable housing in Russia as the main target, so it needs to adapt to the extreme local climate and meet the multi-functional requirements. The research team conducted a synthetical assessment of the feasibility, and safety of the means. This evaluation results showed that the means can effectively meet the construction standards, and the construction efficiency is much higher than that of traditional multi-functional buildings. Ma et al. [7] studied the cooling roof formed by polymer hybrid materials, and simulated the economic potential and environmental potential of cooling roof in prefabricated buildings by taking a single-storey prefabricated building as an example, through electricity consumption in different climates In contrast, the study concluded that compared with traditional tile roof buildings, cooling roof buildings consume less energy, have a simple payback period, and can take into account economic benefits as well as environmental benefits. The extensive application in architecture has laid the theory and foundation.

On the other hand, the application of BP neural network equivalent Deep Learning (DL) algorithm in architecture has gradually deepened. Ramadoss and Prabath [8] introduced the experimental and numerical research of high-performance fiber concrete. Based on a large amount of collected data, a trained BP neural network was used to build a prediction model. The experimental results show that when the fiber volume fraction is 1.5%, the compressive strength and flexural strength of steel fiber reinforced concrete cylinder are greatly improved. Yang et al. [9] proposed a model based on BP neural network to predict the residual strength of carbon fiber reinforced composites after low-speed impact. 12 groups of simulation experiments were carried out to verify the residual strength prediction performance of the model. The results show that the residual strength predicted by the model is consistent with the finite element results, and the error is less than 5%. Haddad and Haddad [10] summarized and collected more than 440 data points from the literature, and used the Artificial Neural Network (ANN) to build a model to predict the adhesive strength of fiber reinforced polymer concrete. The results show that the model has high fitting ability and accurate prediction ability in training and test data, with lower mean square error, and its prediction accuracy is far higher than the empirical model. Fu et al. [11] proposed a building energy consumption prediction method based on deep neural networks transfer reinforcement learning (DNN-TRL) to solve the problem of low accuracy of traditional building energy consumption prediction methods. This method introduces a stack denoising automatic encoder to extract the deep characteristics of building energy consumption, and shares the hidden layer structure to transmit the public information between different building energy consumption problems. The results show that the model has significant prediction accuracy. Wang et al. [12] proposed a model based on Genetic Algorithm (GA) and BP neural network to predict the geometric characteristics of laser induction hybrid welding joints, which provides a basis for the energy transfer control and real-time monitoring of welding quality of the welding rivet hybrid welding method. Saleem and Gutierrez [13] proposed a non destructive testing method based on ANN, which estimates the concrete cracks around the reinforcement through ultrasonic pulse velocity testing. The model is used to predict the crack width and analyze the sensitivity of various factors affecting the bond deterioration. The results show that the model has reached a high accuracy level in prediction. Dey et al. [14] proposed a prediction model based on ANN to predict the reliable service life of reinforced concrete structures. The influence of corrosion rate, thickness of protective layer and diameter of reinforcement on the prediction accuracy of the model is compared. The results show that the model has superior accuracy and reliability in predicting the service life of reinforced concrete structures. Li et al. [15] proposed a prediction model based on ANN to predict the compressive strength of cylindrical concrete. The experimental data of 58 cylindrical concrete specimens under concentric loading were collected from the literature for training and testing the neural network, and the model was compared with other models. It is found that the model can logically capture the basic shape of concrete.

In the previous research on the reinforcement effect of soft soil foundation, the dynamic response and influencing factors of soft soil foundation under impact load have not been deeply studied, and the predictive evaluation method of reinforcement effect is not perfect, and the evaluation of reinforcement effect is relatively less. In view of these problems, the research will summarize the dynamic response of soft soil under impact load and its influencing factors, and analyze the factors affecting the reinforcement effect combined with engineering practice; The improved BP neural network model is used to establish the evaluation model according to the engineering data, and it is used for quantitative analysis to provide theoretical guidance for the actual construction of soft soil foundation reinforcement.

3. Evaluation of the reinforcement effect of soft soil foundation for prefabricated buildings based on BP neural network

Soft soil foundation has high natural water content, large void ratio, large compressibility and small permeability coefficient, and its structure is stable, and its strength is not easy to recover after being disturbed. When using traditional methods such as rammers to reinforce soft soil foundations, the pore water pressure in the foundation soil layer increases after tamping [16]. If the increase of pore pressure is too small, the soil layer cannot be fully consolidated to realize the needed reinforcement effect; if the increase of pore pressure is too large and the pressure is not evanished in time, the soil layer will easily become rubber soil, and the intensity of the surface layer will be reduced instead [17]. Therefore, the key to strengthening soft soil foundations is to control the variation range of pore water pressure and its dissipation rate.

The most widely used and more mature one in geotechnical engineering is the Back Propagation neural network (BP), a multi-layer forward artificial neural network with one-way propagation. The neurons are completely linked, but are not linked in the same layer [18]. Compared with the traditional numerical method of geotechnical engineering, BP neural network can determine the reasonable network structure, and use a large number of test results to train the network to obtain the model, avoiding the labor of looking for empirical expressions; All experimental data are applied to the model, so the model has good fault tolerance; It also has self-learning function and can gradually improve the model through continuous learning. But at the same time, BP is easy to fall into local optimization, and there is no theoretical guidance on the number of hidden layer and hidden layer nodes. In addition, the algorithm requires that all instance input vectors have the same dimension. If new examples are added, the training effect of the original dataset will be affected. Therefore, the L-M algorithm with fast convergence speed is adopted to improve it [19, 20, 21]. BP is an iterative algorithm whose data is transmitted in the forward direction and back propagation of errors constitute its learning process. The structure of BP is shown in Fig. 1.

Figure 1.

Basic structure of BP.

In Fig. 1, the first layer is the output layer neuron, the second layer is the hidden layer neuron, and the third layer is the input layer neuron. Input the vector from the input node $X_{k}=\left({x_{1k},x_{2k},x_{3k},\ldots,x_{nk}}\right)$ , among which $k=1,2,3,\ldots,n$ , after the processing of the hidden layer, the network calculation value, vector $Y_{k}=\left({y_{1k},y_{2k},y_{3k},\ldots,y_{nk}}\right)$ , among which $k=1,2,3,\ldots,n$ , and $k$ the expected output obtained $\hat{Y}_{k}=\left({\hat{y}_{1k},\hat{y}_{2k},\hat{y}_{3k},\ldots,\hat{y}_{mk}}\right)$ by the $k=1,2,3,\ldots,m$ output node $k$ are compared. If the error between the calculation and the expected output data is within the allowable range, end the process; otherwise, it will enter the back-propagation error handling, $Y_{k}$ and $\hat{Y}_{k}$ the error between and will be back-propagated from the output layer to the input layer. Revise. Establish a BP-network structure, and decide that levels number is greater than or equal to the number of nodes in each level, and the number of nodes in the first $l$ level is $n^{\left(l\right)}=n$ , $n^{\left(L\right)}=m$ . Frame the matrix formed by the weights of connection values between different levels, the first $l$ level connecting the first $l+1$ level can be $W^{\left(z\right)}=\left[{W_{ij}^{\left(z\right)}}\right]_{n}^{\left(z\right)}% \times_{n}^{z+1}$ , $l=1,2,\ldots,L-1$ , and initialize the element value of each connection weight matrix. Error within allowable range $\delta$ and learning behavior ratio $\eta$ , and initialize iterations number $t=1$ . The output data of different nodes at the input level is calculated as shown in Eq. (1).

$\displaystyle O_{zk}^{l}=f\left({x_{zk}}\right)$ (1)

In Eq. (1) $z=1,2,\ldots,n$ . The auxiliary words of different nodes in different levels are calculated and the corresponding input and output are obtained, as shown in Eq. (2).

$\displaystyle I_{ji}^{\left(l\right)}=\sum\limits_{i=1}^{n^{\left({l-1}\right)% }}{W_{ij}^{\left({L-1}\right)}O_{JK}^{l-1}}$ (2)

In Eq. (2) $l=2,3,\ldots,L$ , $j=1,2,\ldots,n^{\left(l\right)}$ . The error calculation process between different output nodes in the output hierarchy is as follows:

$\displaystyle\left\{{\begin{array}[]{l}y_{jk}=O_{jk}^{\left(L\right)}\\ E_{jk}=(-y_{jk}+\hat{y}_{jk})^{2}/2\\ \end{array}}\right.$ (3)

In Eq. (3) $j=1,2,\ldots,n$ . If it corresponds to $N$ any instance $K$ value of each learning instance $E_{jk}\leqslant\varepsilon,j=1,2,\ldots,m$ , then the learning operation is washed-up, but the reverse propagation is carried out to revise each connection weight matrix. Revise the matrix formed by connecting the generated weights from the hidden level of level L-1 to the output level (level L), as shown in Eq. (4).

$\displaystyle\left\{{\begin{array}[]{l}\delta_{jk}^{\left(L\right)}=-\left({-Y% _{jk}+Y_{jk}^{\ast}}\right)f^{\prime}\left({i_{jk}^{\left(L\right)}}\right)\\ W_{ij}^{\left({L-1}\right)}\left({1+t}\right)=W_{ij}^{\left({L-1}\right)}\left% (T\right)+\eta\delta_{jk}^{\left({L+1}\right)}O_{ik}^{\left(L\right)}\\ \end{array}}\right.$ (4)

In Eq. (4) $j=1,2,\ldots,m$ , $i=1,2,\ldots,n^{\left({L-1}\right)}$ . The weight matrix of the connections generated by different hidden levels is continuously modified inversely, as shown in Eq. (5).

$\displaystyle\left\{{\begin{array}[]{l}\delta_{jk}^{\left(l\right)}=f^{\prime}% \left({I_{jk}^{\left(l\right)}\sum\limits_{q=1}^{n^{\left({i+1}\right)}}{% \delta_{jk}^{\left(l\right)}}W_{jq}^{\left(l\right)}}\right)\\ W_{ij}^{\left({L-1}\right)}\left(t\right)+\left({-\eta\delta_{jk}^{\left(l% \right)}O_{ik}^{\left({l-1}\right)}}\right)=\left({t+1}\right)\ast W_{ij}^{% \left({L-1}\right)}\\ \end{array}}\right.$ (5)

In Eq. (5) $j=1,2,3,4,\ldots,n^{\left(l\right)}$ , $i=1,2,3,4,\ldots n^{\left({l-1}\right)}$ , $k-1=k\left({\bmod N}\right)$ , $t-1=t$ . The essence of the traditional BP-network is a plain static algorithm of steepest descent, which is corrected on the basis of $k$ the negative gradient of the time $w\left(k\right)$ . It does not need to be considered that previous accumulate information in the process, so it is easy to oscillate during the learning process, resulting in slow convergence. The flow of the BP algorithm is shown in Fig. 2.

Figure 2.

Flow of BP algorithm.

BP adopts the gradient descent method in the learning process, which is likely to be caught in the minimum data in local interval and cannot obtain the global optimum, and there is no theoretical guidance for the number of hidden level and its nodes [22]. The main reason for the slow convergence of BP algorithm is the improper selection of learning rate. If it is too small, the convergence will be too slow. If it is too large, the correction will be excessively oscillating and divergent. The adaptive learning rate method can cut down the learning time and adjust the step size as needed. As shown in Eq. (6).

$\displaystyle\left\{{\begin{array}[]{l}W_{k+1}=W_{k}+2^{\lambda}\alpha\left({k% -1}\right)D_{k}\\ \lambda=sign\left[{D_{k}D_{k-1}}\right]\\ \end{array}}\right.$ (6)

The gradient direction is identical in two successive iterations, it means the sliding speed is slow, the step width may be doubled, which can improve learning efficiency. The learning rate can be set as a function of the number of learning times, the learning rate will decrease as the number of learning times increases to avoid oscillation [23]. Levenberg-Marquardt is an improved algorithm of Gauss-Newton algorithm with more efficient optimization effect. The weight adjustment rate of the G-N algorithm is shown in Eq. (7).

$\displaystyle\Delta Z=J^{T}\left(Z\right)J\left(Z\right)^{-1}J^{T}\left(Z% \right)e\left(Z\right)$ (7)

The weight adjustment rate of the LM algorithm is shown in Eq. (8).

$\displaystyle\Delta Z=\left[{J^{T}\left(Z\right)J\left(Z\right)+\mu I}\right]^% {-1}J^{T}\left(Z\right)e\left(Z\right)$ (8)

In Eq. (8), it $\mu$ represents the adaptive factor. When the result of a certain step in the calculation process increases the sum of squares of the error, it will be multiplied by $\mu$ a number $\omega$ , otherwise, it will be $\mu$ divided by a number $\omega$ . When $\mu$ the value is larger, the L-M algorithm and G-N algorithm are not much different. $Z$ represents the weights and threshold matrices of the network in a BP neural network $W$ . Set the data origin $X^{\left(0\right)}$ , its precision is $\varepsilon_{0}$ , the value of $k$ is 0, $t^{k}=$ 10 ${}^{4}$ , $f_{i}\left({x^{k}}\right)$ obtain the vector $f\left({X^{k}}\right)=\left[{f_{1}\left({X^{k}}\right),\ldots,f_{m}\left({X^{k% }}\right)}\right]$ , $i=1,2,3,4,\ldots,m$ , $j=1,2,3,4,\ldots,n$ . Then, the Jacobi matrix is obtained as shown in Eq. (9). $J_{ij}\left({X^{k}}\right)=\frac{\partial f_{i}\left({X^{k}}\right)}{\partial X% _{j}}$ .

$\displaystyle J\left({X^{b}}\right)=\left[{J_{ij}\left({X^{k}}\right)}\right]$ (9)

The system of linear equations is shown in Eq. (10).

$\displaystyle\left[{J\left({X^{k}}\right)\times J\left({X^{k}}\right)^{T}+t_{k% }I}\right]d^{k}=f\left({X^{k}}\right)\times\left[{-J\left({X^{k}}\right)^{T}}\right]$ (10)

Obtain the search direction $d^{k}$ , when solving a linear equation system, if $J\left({X^{k}}\right)\times J\left({X^{k}}\right)^{T}+t_{k}I$ the rank of the matrix is not n, the equation system will not be solved, and the negative gradient direction is taken $d^{k}$ , as shown in Eq. (11).

$\displaystyle d^{k}=-\nabla F\left({X^{k}}\right)/2=-J\left({X^{k}}\right)^{T}% f\left({X^{k}}\right)$ (11)

A straight line search is performed, as shown in Eq. (12).

$\displaystyle X^{\left({k+1}\right)}-X^{k}=\lambda_{k}d^{k}$ (12)

In Eq. (13) $\lambda_{k}$ should satisfy the conditions shown in Eq. (13).

$\displaystyle F\left({\lambda_{k}d^{k}+x^{k}}\right)=\min\lambda F\left({% \lambda_{k}d^{k}+x^{k}}\right)$ (13)

If so $\left\|{-X^{k}+X^{\left({k+1}\right)}}\right\|<\varepsilon$ , the solution can be obtained $X_{o}$ , and the calculation is terminated at this time, otherwise continue. If $F\left({X^{k}}\right)>F\left({X^{\left({k+1}\right)}}\right)$ , then let $t_{k}\ast 2=t_{k}$ , $k-1=k$ , re-find $f_{i}\left({x^{k}}\right)$ ; otherwise $t_{k}=2t_{k}$ , re-solve the system of linear equations. The LM algorithm has a fast convergence speed, but it is still a local optimization algorithm and is prone to oscillation. In practical applications, several operations are required to achieve satisfactory results.

This study will analyze the pore water pressure of soft clay under impact load and its influencing factors, select soft clay from a foundation pit on a river bank, grind and sieve it to prepare a disturbed experimental soil sample. The basic properties of the soil sample are, size $\phi=$ 3.82 cm, $h=$ 7 cm, liquid index $W_{L}=$ 35.7%, plasticity index $W_{P}=$ 16.2%, plasticity index $I_{P}=$ 21.3, and water content 37.0%. In order to improve the drainage conditions of the samples, high-water-permeability paper drainage cores were set at the central axis of some samples and connected with the filter paper above and below the samples. The sample is placed under a certain ambient pressure to complete the main consolidation, and N times an axial impact load is applied to it without drainage. After the pore water pressure reaches a certain level, to ensure that the sample is consolidated and stabilized under drainage conditions, the last impact load is carried out, and then the CU shear test is carried out. The relationship under different $N$ impact energies is $\varepsilon_{a}$ shown in Fig. 3.

Figure 3.

Relationship between $\varepsilon_{a}$ and $N$ under different impact energies.

In case of equal confining pressure, strain force in axial direction and pore water pressure rise as the grows of the impact number $n$ . When $N<$ 10, the pore water pressure is obviously affected by the impact force, and the axial strain tends to converge. When $N>$ 10, the variation range of pore water pressure tends to be gentle and gradually approaches the limit value, $\varepsilon_{a}$ roughly maintaining the same rate of change, and the threshold value appears. Under the condition that the impact energy remains the same, although the next impact energy increases, the axial strain and pore pressure increase decrease compared with the previous ones, indicating that the sample resists external impact after impact reconsolidation. The load capacity is improved. When $N=$ 20, the incremental change of pore water pressure tends to be gentle. At this time, the impact energy is increased to carry out the impact, and the growth rate of pore pressure will increase and enter a stable state for the second time. But the axial strain becomes larger and larger as the pore pressure increases. Coordinating the relationship between pore pressure and impact energy, the growth rate of pore pressure rises as the grows of impact energy, and when large impact force occurs, the pore pressure can quickly stabilize.

The number of impacts can reflect the influence of the change of impact load on the properties of the soft soil samples under the subsequent impact load, and can be $n_{q}$ quantitatively expressed by the quasi-over-consolidation ratio of the samples that are re-consolidated after impact, as shown in Eq. (14).

$\displaystyle u=f\left({\sigma^{\prime}_{3c},E,N,n_{q}}\right)$ (14)

In Eq. (14), $u$ represents the accumulated pore water pressure after impact load under undrained conditions; $\sigma^{\prime}_{3c}$ is the surrounding pressure of the soft soil sample; $E$ is the impact energy; $N$ is the number of impacts. The relationship curve of pore pressure increment, tamping amount and ramming number is shown in Fig. 4.

When tamping is performed on a certain point, the more tamping times, the greater the pressure value generated by voids, it increases at a faster rate at the initial stage, and then gradually decreases with the growth of what tamping times, its augmenter rises gradually.

After analyzing the relationship between the pore water pressure of soft soil under impact load and its influencing factors, the experimental samples were selected. Due to the engineering properties of soft soil are different, the study selects the sedimentary soft soils of fluvial facies and marine facies as the training set, and randomly selects some of them as the test set. The regional geomorphological conditions selected for the study are relatively simple. According to the characteristics of geological engineering, the soil layer is separated into 5 engineering geological and layering of top and bottom distinction. The mechanical properties of different soil levels are different is Table 1.

Table 1

Physical and mechanical property indexes of soil layer

Layer	Surface characteristics of the project area	Natural water content W (%)	Natural gravity $r$ (kN/m ${}^{3}$ )	Natural aperture ratio	Compression index $C_{C}$	Plasticity index $I_{P}$
1	Plain fill	30.2	18.5	0.852	0.245	15.2
2	Silty clay	31.4	19.2	0.864	0.276	16.4
3	Muddy silty clay	38.7	18.0	1.086	0.308	14.1
4	Silty clay mixed with fine sand	39.8	17.6	1.154	/	13.3
5	Fine sand mixed with silty clay	32.3	18.4	0.930	/	12.4
6	Silty fine sand	29.5	18.5	0.842	/	/
7	Silty fine sand mixed with silty clay	28.9	19.1	0.832	/	/

Figure 4.

Relationship curve between pore water pressure increment, tamping settlement and tamping blow count.

The data is normalized to speed up the convergence and improve the accuracy, as shown in Eq. (15).

$\displaystyle\left\{{\begin{array}[]{l}\bar{x}_{i}=\frac{x_{i}-\left({x_{\max}% +x_{\min}}\right)/2}{x_{\max}-x_{\min}}\\ \bar{y}_{i}=\frac{y_{i}-\left({y_{\max}+y_{\min}}\right)/2}{y_{\max}-y_{\min}}% \\ \end{array}}\right.$ (15)

In Eq. (15), $\bar{x}_{i}$ and $\bar{y}_{i}$ represent the sample results obtained after normalization; $x_{i}$ and $y_{i}$ represent the data in the sample; $x_{\max}$ and $y_{\max}$ are the maximum values in the sample; $x_{\min}$ and $y_{\min}$ are the minimum values in the sample. In the normalized data input model, there is no theoretical guidance for the determination of hidden layers’ number and units’ number in this layer. According to past experience and research data, the network model is determined to be 3 layers. The more neurons in the hidden level, the fewer iterations it takes for the model to converge to the required error, but at the same time, the longer the single training time. The fewer neurons in the BP network, especially in the hidden layer, the more iterations the model needs to converge to the required error, but the time for a single training is shorter. Therefore, the higher the number of neurons, the better the function approximation of the network, the slower the operation because it requires more memory. Usually, when neurons number in the input level is more than the hidden level, the network error decreases rapidly. Under the condition that the total number of neurons remains unchanged, the more neurons in the hidden layer, the longer the time required for model training. The number of neurons should be appropriate, if too much or too little, it will lead to the reduction of the generalization ability. If the mean square error is not appropriate, it will also influence the function approximation ability and generalization ability of the network [24].

4. Evaluation and analysis of soft soil foundation reinforcement effect of prefabricated buildings based on BP neural network

To ensure that the soil conditions of samples are the same, the burial depth is 6 m, the liquid index is 1.18 IL, the void ratio is 1.12 e, and the upper layer thickness is 3.5 m. Using different construction methods, input the sample data into the model for calculation, and the reinforcement effect is shown in Table 2.

Table 2
Reinforcement effect

Number	Vertical drainage	construction technology			Estimated effect
	1/Well diameter ratio	Single point compaction parameters kN $\cdot$ m/m ${}^{2}$	Tamping times	Full compaction parameters	Increase of specific penetration resistance %
1	0.0376	1550	2	1400	211.85
2	0.0376	1550	1	1400	206.74
3	0.0376	1550	4	1400	247.52
4	0.0376	1550	3	1400	218.32
5	0.0376	1550	5	1400	379.68
6	0.0376	2000	1	1400	171.25
7	0.0376	2000	2	1400	198.42
8	0.0376	2000	3	1400	232.02
9	0.0376	2000	4	1400	345.68
10	0.0376	1550	3	2000	223.17

For the soft soil samples of fluvial facies and marine facies, under the condition of the same area and the same ramming energy, the more ramming times, the better the reinforcement effect. The reinforcement effect of 1550 kN $\cdot$ m/m ${}^{2}$ is better than that of 2000 kN $\cdot$ m/m ${}^{2}$ when only ramming is performed once. When other conditions are kept the same, the higher the ramming energy, the better the reinforcement effect. A certain value is changed under the condition that other conditions remain unchanged. After normalization, it is input into the network model, and the least squares method is used to fit the reinforcement effect under different ramming energies, and the numerical value and the approximate value are obtained. The relationship between the composite curve and the reinforcement effect is shown in Fig. 5.

Figure 5.

Relationship curve between numerical value and fitting curve and reinforcement effect.

The buried depth of the reinforced soil layer is 5 m, the thickness is 3.7 m, the liquidity index is 1.26, the void ratio is 1.15, there is no permeable boundary, and the well diameter ratio is 26. It can be seen from Fig. 5 that the average tamping energy of fitting curve 1 is 1570 kJ/m ${}^{2}$ , and the average tamping energy of fitting curve 2 is 2000 kJ/m ${}^{2}$ . According to the analysis of the two fitting curves, the more times of tamping, the greater the specific penetration resistance and the multiple consolidation of the soft soil foundation, with the average tamping energy unchanged, have a better reinforcement effect. With the increase of tamping times, the tamping energy decreases gradually. When the tamping energy is high, the increase of specific penetration resistance after tamping is higher than that when the tamping energy is low, indicating that the soft soil structure has a certain strength. When the times of tamping are different, the increase of specific penetration resistance increases linearly with the increase of tamping energy, and the curves are basically parallel. This shows that the reinforcement effect of tamping energy is not affected by the times of tamping, and both of them individually affect the reinforcement effect, but they are not affected by each other. The increase of tamping energy has a great influence on the increase of penetration resistance. The relationship between void ratio and reinforcement effect is shown in Fig. 6.

Figure 6.

Relationship between void ratio and reinforcement effect.

Figure 7.

Relationship between compressibility of overlying soil layer and reinforcement effect.

Figure 8.

Relationship between various conditions and reinforcement effect.

The larger the void ratio, the smaller the increase in specific penetration resistance, and there is an eigenvalue, which is 1.4. When the void ratio is lower than 1.4, the increase or decrease of the void ratio has little effect on the reinforcement effect, and the increase of the specific penetration resistance diminishes as the grows of the void ratio, the diminish rate also diminishes. When the void ratio is higher than 1.4, the increase of the specific penetration resistance diminishes as the grows of the void proportion, the decreasing range is faster. In the case of single-sided drainage, the increase in specific penetration resistance is higher than that without drainage boundary. However, under different drainage conditions, the shapes of the curves are basically the same, indicating that both drainage conditions and void ratio have an influence on the increase of penetration resistance independently, but there is no interaction between the two. The relationship between the influencing factors and reinforcement effect of marine soft soil is similar to that of fluvial soft soil, but also different. The general trend is roughly the same as that of river facies soft soil, the greater the burial depth, the worse the reinforcement effect. With the improvement of the drainage boundary conditions, the reinforcement effect also increases. When the soil layer is buried deeper, the reinforcement effect does not change much. Correlation between the compressibility of overlying surface layer and the reinforcement effect is shown in Fig. 7.

In Fig. 7, the compressibility of the overburden varies according to the burial depth of the reinforced soil layer. When the burial depth is 3 m, with the increase of the compressibility of the overlying soil layer, the increase of specific penetration resistance decreases from 2.7 to about 0.4. When the burial depth is 6 m, with the increase of the compressibility of the overlying soil layer, the specific penetration resistance increases from about 0 to about 3.0. It shows that the soil layer with higher compressibility is more favorable for the spread of ramming to the deep, but the soil layer with shallow burial depth reduces the reinforcement effect because it absorbs most of the energy. Therefore, when the compressibility of the overlying soil layer is large, the soil layer with shallow burial depth can be reinforced by laying cushion. The relationship between soil layer thickness, vertical drainage facilities, ramming times, full ramming energy and reinforcement effect is shown in Fig. 8.

Figure 9.

Model accuracy comparison.

In Fig. 8a, the increase of specific penetration resistance decreases with the increase of the thickness of the reinforced soil layer. The increase of sample 3 decreased the most, from 2.7 to 0.6. The thinner the soil layer, the worse the reinforcement effect. In Fig. 8b, with the improvement of vertical drainage facilities, the specific penetration resistance increases. When the burial depth of soil sample is 2 m, the specific penetration resistance increases from 1.0 to 11.0 with the improvement of shaft drainage facilities. It can be seen from the analysis that the deeper the burial depth of soil sample is, the smaller the increase of specific penetration resistance with the improvement of shaft drainage facilities. In Fig. 8c, the increase of specific penetration resistance increases with the increase of tamping times. When the burial depth of soil sample is 3 m, the specific penetration resistance increases from 0.4 to 7.8. It can be seen from the analysis that the deeper the burial depth of soil sample is, the smaller the increase of specific penetration resistance with the increase of tamping times. In Fig. 8d, the increase of specific penetration resistance increases with the increase of full tamping energy. When the burial depth of soil sample is 2 m, the specific penetration resistance increases from about 1.0 to about 1.25. It can be seen from the analysis that the deeper the burial depth of soil sample is, the smaller the increase of specific penetration resistance with the increase of full tamping energy is. The laws reflected in the data output from the model are the same as those reflected in the construction, indicating that the prediction results of the model are reliable and can be used for the preliminary estimation of the reinforcement effect. Compare the evaluation model proposed in the study with other evaluation models of reinforcement effect, as shown in Fig. 9.

It can be seen from Fig. 9 that the evaluation accuracy of the four models improves with the increase of the number of iterations, but the model proposed in the study has the highest accuracy in evaluating the effect of soft soil foundation reinforcement. When the number of iterations is 300, the accuracy rate reaches 87%, which is significantly improved compared with other models. The experimental results show that the evaluation effect of the proposed soft soil foundation reinforcement model is remarkable, the laws reflected in the model output data are consistent with those reflected in the construction, the accuracy of the model evaluation is extremely high, and the evaluation performance of the model is the best.

5. Conclusion

In order to improve the reinforcement effect of prefabricated buildings on soft soil foundation, BP neural network is applied to the evaluation of reinforcement effect of soft soil foundation. The evaluation model of soft soil foundation reinforcement effect of prefabricated building based on BP neural network is established. In the process of establishing the model, in order to solve the problem of slow convergence speed of BP neural network, L-M algorithm is used to optimize it and improve the operation performance of the model. Finally, the evaluation model of prefabricated building soft soil foundation reinforcement based on BP neural network is applied to the actual evaluation scene to test its evaluation effect. The results show that when the area is equal and the tamping energy is the same, the soft soil samples of river facies and marine facies show the characteristics that the more times of tamping, the better the reinforcement effect. When only one tamping, the reinforcement effect of 1550 kN $\cdot$ m/m ${}^{2}$ is better than that of 2000 kN $\cdot$ m/m ${}^{2}$ . At the same time, soft soil samples also show the law that the greater the burial depth, the worse the reinforcement effect. In addition, it can be seen from the analysis that the increase of specific penetration resistance decreases with the increase of the thickness of the reinforced soil layer, and the reinforcement effect shows a downward trend when the soil layer is very thin. The laws reflected in the data output from the model are the same as those reflected in the construction, indicating that the prediction results of the model are reliable. The accuracy rate of the four evaluation models is compared, and the accuracy rate of the proposed model is as high as 87%, which shows that it can achieve accurate evaluation, stability evaluation and specific evaluation in the reinforcement effect evaluation.

References

Guri

Brzev

Lluka

. Performance of prefabricated large panel reinforced concrete buildings in the november 2019 Albania earthquake. J Earthq Eng. 2021; 26(3): 1-27.

Wang

Shao

Wang

Liu

. Study on the effect of pile foundation reinforcement of embankment on slope of soft soil. Sustainability. 2022; 14(21): 14359.

Zhang

. Resilience-cost tradeoff supply chain planning for the prefabricated construction project. J Civil Eng Manag. 2021; 27(1): 45-59.

Ahn

Han

Al-Hussein

. Improvement of transportation cost estimation for prefabricated construction using geo-fence-based large-scale GPS data feature extraction and support vector regression. Adv Eng Inform. 2020; 43: 1010121-101012.15..

Wei

Men

Zhu

Fan

. The reinforcement effects of deep soft soil foundation with high degree of saturation under dynamic compaction. Adv Mech Eng. 2018; 10(8): 1687814018782633.

Zheng

Wang

. Application of visualization modeling technology in the determination of reinforcement range of deep soft soil foundation. Environ Earth Sci. 2022; 81(7): 215.

Zhang

Chen

Tang

. Analysis of the impact of a novel cool roof on cooling performance for a low-rise prefabricated building in China. Build Serv Eng Res T. 2021; 42(1): 26-44.

Ramadoss

Prabath

NVN

. Engineering properties and prediction of strength of high performance fibre reinforced concrete using artificial neural networks. Electron J Struct Eng. 2021; 12(1): 76-90.

Yang

Lee

. Artificial neural network (ANN)-based residual strength prediction of carbon fibre reinforced composites (CFRCs) after impact. Appl Compos Mater. 2021; 28(3): 809-833.

10.

Haddad

. Predicting fiber-reinforced polymer-concrete bond strength using artificial neural networks: A comparative analysis study. Struct Concrete. 2021; 22(1): 38-49.

11.

Liu

Gao

Chen

. A building energy consumption prediction method based on integration of a deep neural network and transfer reinforcement learning. Int J Pattern Recogn. 2020; 34(10): 2052005.1-2052005.18.

12.

Wang

Zhang

Liu

. Prediction and fitting of weld morphology of Al alloy-CFRP welding-rivet hybrid bonding joint based on GA-BP neural network. J Manuf Processe. 2021; 63: 109-120.

13.

Saleem

Gutierrez

. Using artificial neural network and non-destructive test for crack detection in concrete surrounding the embedded steel reinforcement. Struct Concrete. 2021; 22(5): 2849-2867.

14.

Dey

Miyani

Sil

. Application of artificial neural network (ANN) for estimating reliable service life of reinforced concrete (RC) structure bookkeeping factors responsible for deterioration mechanism. Soft Comput. 2020; 24(3): 2109-2123.

15.

Rayhani

MHT

El Naggar

. Numerical modeling of seismic response of rigid foundation on soft soil. Int J Geomech. 8(6): 336-346.

16.

Roy

Pandey

Saha

. Shake table study on seismic soil-pile foundation-structure interaction in soft clay. Structures. 2021; 29(9): 1229-1241.

17.

Guo

Jiang

Wang

Zhou

. Design of floor plan and components of prefabricated building with energy-cost effect. Appl Sci-Basel. 2022; 12(8): 3740.

18.

Sun

Zhang

Huang

. Prediction of phthalates concentration in household dust based on back propagation neural network. Indoor Built Environ. 2022; 31(1): 230-244.

19.

Pan

Yan

Lan

Badal

Qin

. A sparse spike deconvolution algorithm based on a recurrent neural network and the iterative shrinkage-thresholding algorithm. Energies. 2020; 13(12): 3074.

20.

Sharifi

Moghbeli

. Shear capacity assessment of steel fiber reinforced concrete beams using artificial neural network. Innov Infrastruct So. 2021; 6(2): 1-19.

21.

Xing

Zhang

Liu

Yao

. Comprehensive monitoring of talus slope deformation and displacement back analysis of mechanical parameters based on back-propagation neural network. Landslides. 2021; 18(5): 1889-1907.

22.

Wen

Sun

Ding

Song

Zhang

. Localization of low velocity impacts on CFRP laminates based on FBG sensors and BP neural networks. Mech Adv Mater Struc. 2021; 29(26): 5478-5487.

23.

Zhao

. Application of construction waste in the reinforcement of soft soil foundation in coastal cities. Environ Technol Inno. 2021; 21: 101195.

24.

Raza

Shah

SAR

ul Haq

Arshad

Raza

Farhan

, et al. Prediction of axial load-carrying capacity of GFRP-reinforced concrete columns through artificial neural networks. Structures. 2020; 28: 1557-1571.

Evaluation of soft soil foundation reinforcement effect of prefabricated building based on BP neural network

Abstract

Keywords

1. Introduction

2. Related wsork

3. Evaluation of the reinforcement effect of soft soil foundation for prefabricated buildings based on BP neural network

Table 2 Reinforcement effect

References

Table 2
Reinforcement effect