Method for correcting multi-scale single pile p-y curve of soft soil layer based on recurrent neural network

Abstract

The present application relates to a kind of multi-scale soft soil layer single pile P-Y curve correction method based on recurrent neural network, can accurately predict the deformation of single pile pile head after horizontal load, especially in the field of pile body horizontal load and deformation calculation.The present application obtains single pile micro P-Y curve by numerical simulation, obtains single pile mechanical P-Y curve by empirical formula, constructs multi-scale P-Y curve sequence data combined with pile and stratum environmental variable, compared with prior art, actual environmental influence is considered, and the model multi-factor analysis ability is improved.Then based on LSTM recurrent neural network, the P-Y LSTM model is constructed, the model comprehensively considers the influence of multi-scale mechanical derivation result and multi-modal experience data, greatly improves the accuracy of P-Y curve prediction, can accurately calculate the accurate value of load and pile deformation of single pile pile head in soft soil layer after horizontal load.

Description

Technical Field

[0001] This invention relates to a multi-scale soft soil stratum single pile PY curve correction method based on recurrent neural networks, which can accurately predict the deformation of the single pile head after being subjected to horizontal load, and is particularly related to the field of pile horizontal load and deformation calculation. Background Technology

[0002] The existing standard uses a single pile PY curve that has errors in actual application and field measurement data. The acquisition of the single pile PY curve mainly adopts mechanical calculation method and simulation method based on finite difference. The correction of the PY curve is also an improvement proposed for these two methods. At present, the methods and shortcomings of the PY curve prediction are mainly: (1) After the single pile head is subjected to horizontal load in soft soil layer, the deformation of the pile body is related to the surrounding soil layer conditions, the elastic modulus of the pile body and the load size. According to the mechanical empirical method of single pile horizontal load and displacement given in the "Port Engineering Pile Foundation Code", it is difficult to accurately obtain the deformation characteristics of the pile body after the pile head is subjected to horizontal load. In particular, the horizontal displacement of a single pile exceeds 8y 50 (1) The results calculated by the empirical mechanical method are increasingly different from the measured values ​​on site. (2) The PY curve obtained by the finite difference method at the microscopic pile scale also has a large error with the deformation of the pile under horizontal load in actual engineering. (3) The correction technology for the PY curve is limited to single-scale correction, or introduces multiple parameters to correct the microscopic data, or modifies the macroscopic mechanical calculation formula, without considering the combination of macroscopic and microscopic conditions. (4) The correction technology for the PY curve fails to incorporate actual empirical data and lacks targeted correction for specific environments. Therefore, accurately predicting the actual deformation curve of the pile under horizontal load has become an urgent need in this field. Summary of the Invention

[0003] To address the critical issue of accurately predicting the true deformation curve of a pile under horizontal load, this invention uses recurrent neural networks to correct existing PY curves and construct a corrected PY curve model that accurately reflects the relationship between the horizontal load at the pile top and deformation. Specifically, this invention employs a multi-scale soft soil stratum single-pile PY curve correction method based on recurrent neural networks. This method considers not only the microscale but also the mechanical properties, and compared to single-scale PY curve calculation methods, it significantly improves the accuracy and speed of the calculation. The specific steps include:

[0004] S1. Based on the working conditions of a single pile and the surrounding soft soil layer, a model is established using finite difference numerical simulation software and parameters are assigned. By applying a horizontal load to the pile head and recording the horizontal displacement value of the pile head under the horizontal load, the microscopic PY curve of the single pile is obtained.

[0005] S2. Based on the empirical formula of the single pile PY curve in the "Port Engineering Pile Foundation Code", and after modification, obtain the PY curve of single pile mechanics.

[0006] The empirical formula for correcting the PY curve of a single pile is as follows:

[0007]

[0008] In the formula: p is the horizontal load on the pile head, N; p u The ultimate soil resistance around the pile is N; y 50 The horizontal displacement of the pile side, in meters, is the displacement of the soil around the pile when it reaches half of the ultimate soil resistance.

[0009] Furthermore, only for pile side deformation exceeding 8y 50 Partial revisions.

[0010] S3. Based on the single pile microscopic PY curve determined in step S1 and the single pile mechanical PY curve determined in step S2, multi-scale data fusion is performed, specifically including:

[0011] S31, taking the same displacement starting point as the sampling starting point, continuously sample the microscopic PY curve of a single pile and the mechanical PY curve of a single pile n times to obtain a 2D single pile multi-scale sequence of (n, 2), and sample the corresponding points of the measured PY curve to obtain a 1D single pile measured PY sequence of (n).

[0012] S32, normalizes the soil strength, soil moisture content, and pile section stiffness of a single pile;

[0013] S33, the normalized soil strength, normalized soil moisture content and normalized pile section stiffness are expressed as a data sequence of (1, 3), which is copied and expanded n times in the first dimension to obtain a single pile environmental variable sequence of (n, 3);

[0014] S34, the single-pile multi-scale sequence of the form (n, 2) and the single-pile environmental variable sequence of the form (n, 3) are spliced ​​in the second dimension to obtain the single-pile fusion sequence of (n, 5);

[0015] S35, the sampling starting point moves backward along the displacement axis, and steps S31-S34 are repeated to obtain m single pile fusion sequences, described as (m, n, 5), which respectively represent (number of single pile fusion sequences, number of single pile fusion sequence data points, number of data point features), and m single pile measured PY sequences, described as (m, n), which respectively represent (number of single pile measured PY sequences, number of measured sequence data points).

[0016] S4 uses a single-pile fusion sequence of the form (m, n, 5) as input and a single-pile measured PY sequence of the form (m, n) as output to train a recurrent neural network, thus constructing a P-YLSTM model based on a recurrent neural network.

[0017] In step S4, preferably, the P-YLSTM model includes a PY data encoding structure and a PY data decoding structure. The PY data encoding structure consists of five consecutive LSTM layers with n units. The PY data encoding structure and the PY data decoding structure are connected by a dimensionality reduction layer. The PY data decoding structure consists of an MLP module composed of three consecutive fully connected layers. The PY data decoding structure is followed by a ReLU activation function layer for output mapping.

[0018] Furthermore, in step S4, the P-YLSTM model is built using the Tensorflow 2.0 framework, with 80 training rounds and a training batch size of 4. MSE is used as the loss function during training, and Adam is selected as the optimization algorithm.

[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: It obtains the microscopic PY curve of a single pile through numerical simulation and the mechanical PY curve of a single pile through empirical formulas, constructing a multi-scale PY curve sequence data that combines pile and soil environmental variables. Compared with existing technologies, this considers the actual environmental impact and improves the model's multi-factor analysis capability. Then, a P-YLSTM model is constructed based on an LSTM recurrent neural network. This model comprehensively considers the multi-scale mechanical derivation results and the influence of multi-modal empirical data, greatly improving the accuracy of PY curve prediction. It can accurately calculate the load and pile deformation values ​​of a single pile head subjected to horizontal loads in soft soil strata. Attached Figure Description

[0020] Figure 1 This is a flowchart of the algorithm for the multi-scale soft soil single-pile PY curve correction method of the present invention;

[0021] Figure 2 This is a comparison chart of the PY curve of a single pile based on parameter inversion and the field measured data;

[0022] Figure 3 This is a structural diagram of the P-YLSTM model of the present invention. Detailed Implementation

[0023] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0024] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0025] like Figure 1 The diagram shown is a flowchart of the algorithm of this invention. The multi-scale soft soil single-pile PY curve correction method based on recurrent neural networks includes the following steps:

[0026] S1. Obtain microscopic PY curve data of a single pile. Based on the actual dimensions of the single pile, a numerical simulation model is established using the finite difference method (FLAC), and soil and pile parameters are assigned. The model dimensions are 10m × 10m × 10m, with a single pile diameter of 0.3m, a pile length of 10m, and a pile depth of 6m. Cambridge simulation is applied to the soil around the pile, and a linear elastic constitutive model is applied to the pile. The boundary conditions and load conditions of the model are controlled, and horizontal loads are applied to obtain the microscopic PY curve of the single pile (the pile head is loaded using a cyclically increasing horizontal load method, with each horizontal load corresponding to a specific pile displacement). Figure 2 As shown in curve (1), combined with... Figure 2 The measured PY curve is shown in the middle curve (4). Some data points are selected from the two curves to obtain Table 1. It can be seen from Table 1 that the error between the pile head horizontal load obtained based on microscopic observation and the measured value is more than 10.8%.

[0027] Table 1 Error of Horizontal Load Calculation in Microscopic Numerical Simulation of Single Pile

[0028] Pile displacement y / m Microscopic horizontal load value p / N Measured value p / N error 0.00172 51722.8 45127.7 14.5％ 0.00403 70876.2 63410.24 11.7％ 0.00639 72376.2 65306.944 10.8％ 0.0091 79576.2 71065.152 11.9％

[0029] S2. Obtain the PY curve data for single pile mechanics. Based on the recommended PY formula for single piles in the "Port Engineering Pile Foundation Code", then calculate the PY curve for piles with side deformation exceeding 8y. 50 The PY curve of a portion is corrected to obtain the corrected PY curve of single pile mechanics; specifically, formula (1) is used for calculation and correction to obtain the PY curve of the pile body; such as Figure 2 As shown in the curve (2), some data points are selected on the curve to obtain Table 2. The displacement of the data points is the same as that in Table 1. From Table 2, it can be seen that the error between the horizontal load of the pile head calculated based on formula (1) and the measured value is more than 7.2%.

[0030]

[0031] In the formula: p is the horizontal load on the pile head, N; p u The ultimate soil resistance around the pile is N; y 50 The horizontal displacement of the pile side, in meters, is the displacement of the soil around the pile when it reaches half of the ultimate soil resistance.

[0032] Table 2 Error of Horizontal Load in Mechanical Calculation of Single Pile

[0033] Pile displacement y / m Mechanical horizontal load value p / N Measured value p / N error 0.00172 36786.2796 45127.7 -18.5％ 0.00403 52873.38 63410.24 19.9％ 0.00639 60126.918 65306.944 -7.9％ 0.0091 66012.794 71065.152 -7.2％

[0034] S3. Perform multi-scale data fusion based on the single pile microscopic PY curve determined in step S1 and the single pile mechanical PY curve determined in step S2. Specifically, this includes the following steps:

[0035] S31, select 2×10 -4 m is used as the sampling interval. Starting from the same displacement point, the PY curve of the single pile microstructure and the PY curve of the single pile mechanics are sampled continuously 100 times to obtain 200 data points. The 200 data points are matched to obtain a 2D sequence of (100, 2) called the single pile multi-scale sequence. The corresponding points of the measured PY curve are sampled to obtain the (100) single pile measured PY sequence.

[0036] S32, the soil strength, soil moisture content and pile section stiffness of a single pile are normalized. The normalization method is to scale the data to the range of [0,1], as shown in formula (2).

[0037] Dout=(D-Dmin) / (Dmax-Dmin) (2)

[0038] Where: D—soil strength, soil moisture content, or pile cross-sectional stiffness;

[0039] Dout—Normalized soil strength, normalized soil moisture content, and normalized pile section stiffness;

[0040] Dmax—corresponds to the maximum value of the characteristic. If the soil strength is 600kPa or more, take 600kPa. If the water content is 1, take 300MPa. If the pile section stiffness is 300MPa or more, take 300MPa.

[0041] Dmin—corresponds to the minimum value of the characteristic, the soil strength is 200kPa or less, the water content is 0, and the pile section stiffness is 50MPa.

[0042] S33, the normalized soil strength, normalized soil moisture content and normalized pile section stiffness can be expressed as a data sequence of (1, 3), which is copied and extended 100 times in the first dimension to obtain a single pile environmental variable sequence of (100, 3).

[0043] S34, the single-pile multi-scale sequence of the form (100, 2) and the single-pile environmental variable sequence of the form (100, 3) are spliced ​​in the second dimension to obtain the single-pile fusion sequence of (100, 5);

[0044] S35, the sampling starting point is moved backward by one sampling point along the displacement axis, and steps S31-S34 are repeated to obtain 1000 single pile fusion sequences, described as (1000, 100, 5), which respectively represent (number of single pile fusion sequences, number of single pile fusion sequence data points, number of data point features). 1000 single pile measured PY sequences are obtained, described as (1000, 100), which respectively represent (number of single pile measured PY sequences, number of measured sequence data points).

[0045] In other words, a single-pile fusion sequence is a two-dimensional sequence data, represented as (n, b), where n = 100 indicates that there are 100 continuous displacement discontinuities in the sequence. These continuous displacement discontinuities are defined as the displacement intervals on the horizontal axis of the PY curve, which are 2 × 10⁻⁶. -4 The continuous sampling of m, b=5 indicates that each point contains 5 characteristic values, which are the ordinate values ​​corresponding to the continuous displacement discontinuity of the PY curve of the single pile microscopic model, the ordinate values ​​corresponding to the continuous displacement discontinuity of the PY curve of the single pile mechanics model, the normalized soil strength, the normalized soil moisture content, and the normalized pile section stiffness.

[0046] S4. Construct a P-YLSTM model based on LSTM (long-short-term-memory recurrent neural network). Use a single-stack fusion sequence of the form (1000, 100, 5) as input and a measured PY curve of the form (1000, 100) as output for recurrent neural network training. The P-YLSTM model includes a PY data encoding structure and a PY data decoding structure. The PY data encoding structure consists of five consecutive LSTM layers with 100 units each. The output of the five neural network layers... The dimensions are (100, 5), (100, 20), (100, 40), (100, 80), and (100, 160), respectively. The PY data encoding structure and the PY data decoding structure are connected by a dimensionality reduction layer. The PY data decoding structure consists of an MLP module composed of three consecutive fully connected layers. The output dimensions of the three fully connected layers are (500), (200), and (100). After the PY data decoding structure, a ReLU activation function layer is connected for output mapping, and the output dimension is (100). The P-YLSTM model structure is as follows. Figure 3 As shown; the P-YLSTM model is built using the Tensorflow 2.0 framework, with 80 training rounds and a training batch size of 4. MSE is used as the loss function during training, as shown in equation (3). Adam is selected as the optimization algorithm.

[0047]

[0048] In the formula: Ij represents the predicted value of the j-th data in the model output sequence, Kj represents the true value of the j-th data in the measured data sequence, and n = 100 is the sequence data size.

[0049] The corrected PY curve trained using the P-YLSTM model is shown below. Figure 2 (3) As shown, some data points were selected on the curve to obtain Table 3. The displacement of the data points is the same as that in Table 1-2. As shown in Table 3, the error of the PY curve was reduced to less than 4% after correction, indicating that the P-YLSTM model training was effective and the error was reduced.

[0050] Table 3. Error in Calculating Horizontal Loads Using the Modified P-YLSTM Model for Single Piles

[0051] Pile displacement y / m Corrected horizontal load value p / N Measured value p / N error 0.00172 43127.7568 45127.7 -3.6％ 0.00403 62010.24 63410.24 -2.2％ 0.00639 67306.944 65306.944 -3.1％ 0.0091 71065.152 71065.152 -0.6％

[0052] S5. Conduct practical application based on the P-YLSTM model. The practical application first requires obtaining the microscopic PY curve and the mechanical PY curve of a single pile. The acquisition method is the same as in steps S1 and S2. Then, according to step S3, the entire single pile fusion sequence (5, 100, 5) to be corrected is output. Unlike S3, in S35, the sampling starting point is moved 100 sampling points backward along the displacement axis. The obtained single pile fusion sequence is used as the input of the P-YLSTM model established in step S4 to obtain the corrected PY sequence considering the influence of multi-scale mechanical derivation results and multi-modal empirical data. Its size is (5, 100). It is stitched together in the first dimension to obtain a corrected PY curve with a length of (500). Select some data points on this curve to obtain Table 4. It can be seen from Table 4 that the error of the corrected PY curve has decreased to less than 3%, indicating that the correction result based on the P-YLSTM model has improved the accuracy of PY curve prediction.

[0053] Table 4 shows the application of horizontal loads and errors in the P-YLSTM model.

[0054]

[0055] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for correcting the PY curve of a single pile in multi-scale soft soil strata based on recurrent neural networks, characterized in that, The method includes the following steps: S1. Based on the working conditions of a single pile and the surrounding soft soil layer, a model is established using finite difference numerical simulation software and parameters are assigned. By applying a horizontal load to the pile head and recording the horizontal displacement value of the pile head under the horizontal load, the microscopic PY curve of the single pile is obtained. S2. Based on the empirical formula for the PY curve of a single pile in the "Code for Design of Pile Foundations for Port Engineering", and for piles with side deformation exceeding 8y... 50 Partial corrections were made to obtain the PY curve of single pile mechanics; S3. Based on the single pile microscopic PY curve determined in step S1 and the single pile mechanical PY curve determined in step S2, multi-scale data fusion is performed, specifically including: S31, taking the same displacement starting point as the sampling starting point, continuously sample the microscopic PY curve of a single pile and the mechanical PY curve of a single pile n times to obtain a 2D single pile multi-scale sequence of (n, 2), and sample the corresponding points of the measured PY curve to obtain a 1D single pile measured PY sequence of (n). S32, normalizes the soil strength, soil moisture content, and pile section stiffness of a single pile; S33, the normalized soil strength, normalized soil moisture content and normalized pile section stiffness are expressed as a data sequence of (1, 3), which is copied and expanded n times in the first dimension to obtain a single pile environmental variable sequence of (n, 3); S34, the single-pile multi-scale sequence of the form (n, 2) and the single-pile environmental variable sequence of the form (n, 3) are spliced ​​in the second dimension to obtain the single-pile fusion sequence of (n, 5); S35, the sampling starting point moves backward along the displacement axis, and steps S31-S34 are repeated to obtain m single pile fusion sequences, described as (m, n, 5), which respectively represent (number of single pile fusion sequences, number of single pile fusion sequence data points, number of data point features), and m single pile measured PY sequences, described as (m, n), which respectively represent (number of single pile measured PY sequences, number of measured sequence data points). S4 uses a single-pile fusion sequence of the form (m, n, 5) as input and a single-pile measured PY sequence of the form (m, n) as output to train a recurrent neural network, thus constructing a P-YLSTM model based on a recurrent neural network.

2. The method for correcting the PY curve of a single pile according to claim 1, characterized in that, In step S2, the corrected empirical formula for the PY curve of a single pile is: In the formula: p is the horizontal load on the pile head, N; p u The ultimate soil resistance around the pile is N; y 50 The horizontal displacement of the pile side, in meters, is the displacement of the soil around the pile when it reaches half of the ultimate soil resistance.

3. The method for correcting the PY curve of a single pile according to claim 1, characterized in that, In step S4, the P-YLSTM model includes a PY data encoding structure and a PY data decoding structure. The PY data encoding structure consists of five consecutive LSTM layers with n units. The PY data encoding structure and the PY data decoding structure are connected by a dimensionality reduction layer. The PY data decoding structure consists of an MLP module composed of three consecutive fully connected layers. The PY data decoding structure is followed by a ReLU activation function layer for output mapping.

4. The method for correcting the PY curve of a single pile according to claim 3, characterized in that, The P-YLSTM model was built using the Tensorflow 2.0 framework, with 80 training epochs and a training batch size of 4. MSE was used as the loss function during training, and Adam was selected as the optimization algorithm.

Images