A meta-learning strategy-based early warning method for bearing capacity attenuation of soft soil pile foundation

Abstract

The present invention relates to a soft soil pile foundation bearing capacity attenuation early warning method based on a meta-learning strategy, the steps are as follows: S1. Building a cyclic loading test device; S2. Making a data set; S3. Building and training a base model; S4. Introducing a meta-learning strategy into the base model obtained in S3, inputting a pre-training data set for pre-training, and obtaining a pre-trained model; S5. Inputting a target data set into the pre-training model for training, and obtaining a model prediction result. The method constructs an intelligent early warning model with rapid adaptability, so that it can make full use of existing engineering experience. When facing new engineering scenarios, only a small amount of on-site monitoring data is needed to quickly adjust and realize accurate prediction of the pile foundation bearing capacity attenuation trend.

Description

Technical Field

[0001] The present invention relates to the technical field of foundation engineering and structural health monitoring, and in particular to a soft soil pile foundation bearing capacity attenuation early warning method based on a meta-learning strategy. Background Art

[0002] In the field of geotechnical engineering, monitoring and early warning of pile foundation bearing capacity attenuation has always been an important research topic. In current engineering practice, traditional methods such as settlement observation and stress-strain monitoring are mainly used to evaluate changes in pile foundation bearing capacity. These methods can record the stress state and deformation characteristics of the pile foundation in real time, providing an important reference for engineering safety. However, this type of observation method is essentially passive monitoring, which can only reflect the state of pile foundation bearing capacity attenuation that has occurred, and cannot effectively predict the bearing capacity attenuation trend that may occur in the future, and it is even more difficult to achieve early warning.

[0003] Finite element analysis can simulate the pile-soil interaction mechanism and calculate the evolution of stress and displacement fields by establishing numerical models, so as to conduct predictive research on the attenuation law of pile foundation bearing capacity. However, finite element analysis has obvious limitations: first, the establishment of accurate geological models and material constitutive relationships requires a large amount of field survey and indoor test data; second, the complex nonlinear calculation process consumes considerable time and computing resources; most importantly, due to the differences in geological conditions and engineering environments, the finite element model established for a specific project is difficult to be directly transplanted to other projects, which seriously restricts the promotion and application of this method in engineering practice.

[0004] Therefore, developing a fast, accurate and universal prediction method for pile foundation bearing capacity attenuation is of great significance to improving the safety and economy of pile foundation engineering. This requires the establishment of a more efficient and practical prediction and early warning system based on traditional monitoring methods and numerical analysis methods combined with new sensing technologies and intelligent algorithms. Summary of the invention

[0005] In view of the shortcomings of the background technology, the technical problem to be solved by the present invention is to provide a soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy. The method can build an intelligent early warning model with rapid adaptability, so that it can make full use of existing engineering experience. When facing new engineering scenarios, only a small amount of on-site monitoring data is needed to quickly adjust and realize accurate prediction of the pile foundation bearing capacity attenuation trend.

[0006] The present invention is completed by adopting the following technical solution: a soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy, the steps are as follows:

[0007] S1. Build a cyclic loading test device.

[0008] A cyclic loading test is performed to obtain the horizontal bearing capacity signal of the pile end of the model pile under the cyclic horizontal load, wherein in the cyclic loading test, the prototype pile is equivalent to a scaled model pile based on the second similarity theory and a cyclic loading test is performed;

[0009] S2. Prepare a data set, wherein the data set includes a training data set of a base model, a target data set and a pre-training data set;

[0010] S3. Build and train the base model, and obtain the base model by training the bidirectional echo state network;

[0011] S4. The base model obtained in S3 introduces a meta-learning strategy, inputs a pre-training data set for pre-training, and obtains a pre-trained model;

[0012] S5. The target data set is input into the pre-trained model for training to obtain the model prediction results.

[0013] Furthermore, the steps for building and training the base model in S3 are as follows:

[0014] S31. Prepare a training data set for the base model;

[0015] S32. Inputting the training data set into the bidirectional echo state network for training to obtain a trained base model;

[0016] S33. Use the Adam optimizer to optimize the parameters of the trained model to obtain the optimized base model.

[0017] Furthermore, in S4, the meta-learning strategy is the Reptile meta-learning algorithm, which optimizes the pre-trained model and updates the pre-trained model parameters by performing inner-loop training and outer-loop training in Reptile meta-learning.

[0018] Furthermore, the steps for pre-training the pre-training dataset in S4 are as follows:

[0019] S41. Copy two sets of identical models φ and φ';

[0020] S42. Randomly sample a set of pre-training sets in S2, so that the model φ' uses a two-way echo state network to reconstruct the sequence on the pre-training set, completing a meta-learning inner loop training;

[0021] S43. Based on the training result of φ', update the parameters of model φ according to the formula φ=φ+β*(φ'-φ), and complete a meta-learning outer loop, where β is the meta-learning rate.

[0022] S44. Repeat steps S42 and S43, select a group of pre-training sets for pre-training in turn, and after completing the set number of inner and outer loops, the model completes pre-training.

[0023] Furthermore, in S31, the step of preparing the training data set of the base model includes resampling the loading displacement and horizontal bearing capacity signals into time series data with a sampling frequency of 10 Hz, recording a total of 8000 cycles of loading signals, taking the first 2000 cycles as the training data set of the base model, and adding the signal to the sliding window for data division preprocessing.

[0024] Further, in S32, the bidirectional echo state network includes an echo state pool and a fully connected read layer that recursively returns to a time series, the training data set in S31 is added to the echo state pool for training, the input signal is projected into a sparsely connected high-dimensional space, and a recursive relationship of the space is established on the time axis, the high-dimensional echo state pool is reduced to a target output signal through the fully connected read layer, and a trained base model is obtained;

[0025] Furthermore, the data set acquisition step in S3 includes performing multiple groups of tests under different working conditions through the cyclic loading test in S1 to obtain time series data of the horizontal bearing capacity attenuation of the pile foundation under different working conditions, and making them into multiple groups of data sets, any one of which is the target data set, and the remaining data sets are pre-training data sets.

[0026] Furthermore, the data set acquisition step in S3 includes performing multiple groups of tests under different working conditions through the cyclic loading test in S1 to obtain time series data of the horizontal bearing capacity attenuation of the pile foundation under different working conditions, and making them into multiple groups of pre-training data sets, and the monitoring data in the actual working condition is the target data set.

[0027] Furthermore, the cyclic loading test device includes a model box, a model pile, a loading system and a monitoring system. The loading system includes a horizontal cyclic loading system, a data collector and a first computer. The first computer is used to set the loading frequency and the number of cycles, control the loading mode, and record the load and displacement data changes at the loading position of the pile body. The monitoring system includes a laser displacement sensor, a dynamic demodulator and a second computer. The second computer records the wavelength changes of the pile body by connecting to the dynamic demodulator.

[0028] Furthermore, a target pile foundation monitoring system is established in actual working conditions, and the monitoring system acquires monitoring data to obtain a target data set of the horizontal bearing capacity of the pile foundation in actual working conditions.

[0029] This invention solves the hysteresis problem of traditional early warning methods, while avoiding the tedious modeling process and computing resource consumption in finite element analysis, and constructs an intelligent early warning model with rapid adaptability. Through the meta-learning strategy, the model can extract common features and laws from different working condition data obtained from historical engineering cases or laboratories, and establish a deep understanding of the mechanism of pile foundation bearing capacity attenuation. When facing new engineering scenarios, only a small amount of on-site monitoring data is needed to quickly adjust and accurately predict the trend of pile foundation bearing capacity attenuation, realize rapid migration and accurate early warning, and provide more efficient and practical decision support tools for engineering practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a general flow chart of the early warning method for the bearing capacity decay of soft soil pile foundation based on meta-learning strategy;

[0031] Figure 2 It is a flow chart of the meta-learning method in the early warning method of the bearing capacity attenuation of soft soil pile foundation based on the meta-learning strategy;

[0032] Figure 3 It is a schematic diagram of boundary effect judgment in cyclic loading test in the soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy;

[0033] Figure 4 It is a schematic diagram of the structure of a cyclic loading test device in a soft soil pile foundation bearing capacity attenuation early warning method based on a meta-learning strategy;

[0034] Figure 5 It is a schematic diagram of the structure of the bidirectional echo state network in S3 of a soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy;

[0035] Figure 6 Schematic diagram of the Reptile meta-learning strategy in S4 of a soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy;

[0036] Figure 7 A schematic diagram showing the comparison of the few-sample signal reconstruction effect using the meta-learning method in a soft soil pile foundation bearing capacity attenuation early warning method based on the meta-learning strategy;

[0037] Figure 8 Schematic diagram of the comparison of the few-sample signal reconstruction effect without meta-learning method in a soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy. DETAILED DESCRIPTION

[0038] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0039] Reference Figure 1-7 As shown, the present invention provides a soft soil pile foundation bearing capacity attenuation early warning method based on a meta-learning strategy, the steps are as follows:

[0040] S1. Build a cyclic loading test device, conduct a cyclic loading test, and obtain the horizontal bearing capacity signal of the pile end of the model pile under cyclic horizontal load. In the cyclic loading test, the prototype pile is equivalent to a scaled model pile based on the second similarity theory and a cyclic loading test is performed.

[0041] S2. Create a data set, which includes a training data set of a base model, a target data set, and a pre-training data set.

[0042] Specifically, the acquisition of the target data set and the pre-training data set includes two methods. The first method includes performing multiple groups of tests under different working conditions through the cyclic loading test in S1 to obtain monitoring data signals, and adding the signals to the sliding window for data division preprocessing to obtain the time series data of the horizontal bearing capacity attenuation of the pile foundation under different working conditions, and making them into multiple groups of data sets, any one of which is the target data set, and the remaining data sets are pre-training data sets. The second method includes performing multiple groups of tests under different working conditions through the cyclic loading test in S1 to obtain monitoring data signals, and adding the signals to the sliding window for data division preprocessing to obtain the time series data of the horizontal bearing capacity attenuation of the pile foundation under different working conditions, all of which are made into multiple groups of pre-training data sets, and the monitoring data in the actual working condition is the target data set.

[0043] S3. Build and train the base model, train the bidirectional echo state network through the training data set of the base model to reconstruct the time series signal of the horizontal bearing capacity of the single pile, and obtain the base model. The hysteresis characteristics of the pile foundation bearing capacity attenuation can be better predicted through bidirectional feature extraction.

[0044] S4. Introduce a meta-learning strategy into the base model obtained in S3, input a pre-training data set for pre-training, and obtain a pre-trained model. The meta-learning strategy is the Reptile meta-learning algorithm, which optimizes the pre-trained model and updates the pre-trained model parameters by performing inner loop training and outer loop training in Reptile meta-learning. The introduction of the Reptile meta-learning method enables the model to learn prior knowledge from related tasks, obtain strong generalization capabilities, and handle prediction tasks under different working conditions.

[0045] S5. The target data set is input into the pre-trained model for training to obtain the model prediction results. The target data set is preferably composed of monitoring data in actual working conditions. Specifically, a target pile foundation monitoring system is established in the actual working conditions, and the monitoring data is obtained by the monitoring system to obtain the target data set of the horizontal bearing capacity of the pile foundation in the actual working conditions.

[0046] Through the meta-learning strategy, the base model is pre-trained based on multiple sets of pre-training data sets, and the parameters of the pre-training model are updated. The model can extract common features and laws from historical engineering cases or laboratories, accumulate relevant experience on the pre-training set, and establish a deep understanding of the mechanism of pile foundation bearing capacity attenuation, so as to achieve rapid migration and accurate early warning under new engineering conditions, and provide more efficient and practical decision support tools for engineering practice. This method is aimed at the research status and actual engineering needs of soft soil pile foundation bearing capacity attenuation early warning, and makes full use of existing engineering experience or test data to build an intelligent early warning model with rapid adaptability. When facing new engineering scenarios, only a small amount of on-site monitoring data is needed to quickly adjust and realize accurate prediction of the pile foundation bearing capacity attenuation trend. It solves the hysteresis problem of traditional early warning methods, while avoiding the cumbersome modeling process and computing resource consumption in finite element analysis.

[0047] Specifically, in the above-mentioned soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy, the steps of building and training the base model in S3 are as follows:

[0048] S31. Training dataset for the base model.

[0049] The steps for making the training data set of the base model include: resampling the loading displacement and horizontal bearing capacity signals into time series data with a sampling frequency of 10 Hz, recording a total of 8000 cycles of loading signals, taking the first 2000 cycles as the training data set of the base model, and adding the signals to the sliding window for data partition preprocessing.

[0050] S32. Input the training data set of the base model in S31 into the bidirectional echo state network for training, reconstruct the time series signal of the horizontal bearing capacity of the single pile, and obtain the trained base model.

[0051] For details, please refer to the attached Figure 5 As shown, the bidirectional echo state network includes an echo state pool and a fully connected read layer (attached Figure 5 The lower middle layer is the echo state pool, and the upper layer is the fully connected read layer). The training data set in S31 is input into the echo state pool for training, the input signal is projected into a sparsely connected high-dimensional space, and a recursive relationship of the space is established on the time axis. The high-dimensional echo state pool is then reduced to the target output signal through the fully connected read layer. The echo state pool is fixed after initialization. The fully connected read layer optimizes its high-dimensional feature extraction capability to obtain the trained base model. Among them, the echo state pool is derived as shown in the following formula:

[0052] (2)

[0053] in Rt for t The echo state pool at a point in time, λ represents the leaked parameter during recursive derivation, Ψ ɑ represents the activation function that introduces nonlinear parsing capabilities into the model. X t and b R are the input signal and bias of the echo pool, respectively. In the bidirectional echo state network, in addition to the forward recursive derivation, a reverse recursive mode is also introduced. This method uses the input signal at a future time point to assist the derivation at the current time point. The reverse recursive derivation is as follows:

[0054] (3)

[0055] S33. Use the Adam optimizer to optimize the parameters of the trained model to obtain the optimized base model.

[0056] Reference Figure 6 As shown, the above-mentioned soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy, the steps of inner loop and outer loop pre-training of the pre-training data set in S4 are as follows:

[0057] S41. Copy two sets of identical models φ and φ';

[0058] S42. Randomly sample a set of pre-training sets in S3, and make the model φ' perform an inner loop training on the pre-training set to optimize the pre-training set;

[0059] S43. According to the training result of φ', update the parameters of the model φ according to the formula φ=φ+β*(φ'-φ) (where β is the meta-learning rate), complete an outer loop, update the parameters of the model based on the pre-training set, and accumulate relevant experience;

[0060] S44. Repeat steps S42 and S43, select a set of pre-training sets for pre-training in turn, and after completing the set number of inner and outer loops, the model completes pre-training. As shown below:

[0061]

[0062] Reference Figure 7-Figure 8 As shown, attached Figure 7 To illustrate the effect of few-shot signal reconstruction using meta-learning methods, Figure 8 This is the reconstruction effect of the few-sample signal without using the meta-learning method. It can be seen that after the introduction of the meta-learning method, only a shorter cycle test is required for the single pile to more accurately predict the attenuation of the horizontal bearing capacity of the single pile.

[0063] For details, please refer to the attached Figure 3-4 As shown, in the S1 step of the above-mentioned soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy, the cyclic loading test device includes a model box 1, a model pile 2, a loading system and a monitoring system. Among them, the loading system includes a horizontal cyclic loading system 3, a collector 4 and a first computer 5, the horizontal cyclic loading system 3 is connected to the model pile 2, and the horizontal cyclic loading system is connected to the first computer 5 through a wire. The loading frequency, number of cycles and control loading mode of the horizontal cyclic loading system are set by the first computer 5, the horizontal cyclic loading system 3 loads the model pile according to the instructions of the first computer, and the first computer 5 records the load and displacement data changes at the loading position of the pile body. The monitoring system includes a laser displacement sensor 6, a dynamic demodulator 7 and a second computer 8. The second computer 8 is connected to the dynamic demodulator 7, the laser displacement sensor 6 is connected to the dynamic demodulator 7, the laser displacement sensor 6 records the wavelength changes of the pile body, and the second computer obtains the monitoring data of the laser displacement sensor 6 through the dynamic demodulator 7.

[0064] Specifically, firstly, the model pile in the cyclic loading test is equivalent to the prototype pile as a scaled model pile based on the second similarity theory. In order to make the model pile and the prototype pile similar in geometry and deformation, similarity processing is performed according to the second similarity theory. The geometric similarity processing is performed using the following formula:

[0065] (1)

[0066] Where L is the pile length (m), the subscripts p and m represent the geometric similarity constants of the prototype and the model respectively; in addition, the similarity conversion relationship of the pile part is shown in Table 1. The prototype pile parameters (taking a certain pile as an example) are similarly converted to obtain the basic parameters of the model pile.

[0067] Table 1 Model conversion relationship

[0068]

[0069] Secondly, after completing the model pile design, a cyclic loading test is carried out on the model pile. Figure 4 As shown in the figure, the effective volume of the model box used in this test is 900mm×700mm×700mm (length×width×height). When the distance between the buried position of the model pile and the inner wall of the model box is greater than 2.82 times the pile diameter, and the distance between the pile end and the bottom of the model box is greater than 4 times the pile diameter, the influence of the boundary effect on the test results is negligible. The cyclic loading test is to conduct cyclic loading tests under different working conditions through the loading system, and the monitoring system monitors the horizontal bearing capacity signal of the pile end of the model pile under cyclic horizontal load, and the computer obtains the monitoring data as the basis of the data set.

[0070] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A soft soil pile foundation bearing capacity attenuation early warning method based on meta-learning strategy, characterized by: The steps are as follows, S1. Build a cyclic loading test device. A cyclic loading test is performed to obtain the horizontal bearing capacity signal of the pile end of the model pile under the cyclic horizontal load, wherein in the cyclic loading test, the prototype pile is equivalent to a scaled model pile based on the second similarity theory and a cyclic loading test is performed; S2. Prepare a data set, wherein the data set includes a training data set of a base model, a target data set and a pre-training data set; S3. Build and train a base model, and obtain the base model by training a bidirectional echo state network; S4. The base model obtained in S3 introduces a meta-learning strategy, inputs a pre-training data set for pre-training, and obtains a pre-trained model. The meta-learning strategy is the Reptile meta-learning algorithm. Inner-loop training and outer-loop training are performed in Reptile meta-learning to optimize the pre-trained model and update the pre-trained model parameters. The steps for pre-training the pre-training dataset are as follows: S41. Copy two sets of the same model φ and φ' ; S42. Randomly sample a set of pre-training sets in S2 so that the model φ' The sequence is reconstructed using a bidirectional echo state network on the pre-training set to complete a meta-learning inner loop training; S43.According to φ' The training results are as follows: φ=φ+β * (φ'-φ) Update the model φ The parameters of , complete a meta-learning outer loop, where β is the meta-learning rate; S44. Repeat steps S42 and S43, select a group of pre-training sets for pre-training in turn, and after completing the set number of inner and outer loops, the model completes pre-training; S5. The target data set is input into the pre-trained model for training to obtain the model prediction results.

2. According to the method of early warning of soft soil pile foundation bearing capacity attenuation based on meta-learning strategy in claim 1, it is characterized in that: the steps of building and training the base model in S3 are as follows: S31. Training data set of base model; S32. Inputting the training data set into the bidirectional echo state network for training to obtain a trained base model; S33. Use the Adam optimizer to optimize the parameters of the trained model to obtain the optimized base model.

3. The method for early warning of soft soil pile foundation bearing capacity attenuation based on meta-learning strategy according to claim 2 is characterized by: In the S31, the step of preparing the training data set of the base model includes resampling the loading displacement and horizontal bearing capacity signals into time series data with a sampling frequency of 10 Hz, recording a total of 8000 cycles of loading signals, taking the first 2000 cycles as the training data set of the base model, and adding the signal to the sliding window for data division preprocessing.

4. The method for early warning of soft soil pile foundation bearing capacity attenuation based on meta-learning strategy according to claim 2 or 3 is characterized in that: In S32, the bidirectional echo state network includes an echo state pool that recursively returns to a time series and a fully connected reading layer. The training data set in S31 is added to the echo state pool for training, the input signal is projected into a sparsely connected high-dimensional space, and a recursive relationship of the space is established on the time axis. The high-dimensional echo state pool is reduced to the target output signal through the fully connected reading layer to obtain a trained base model.

5. The method for early warning of soft soil pile foundation bearing capacity attenuation based on meta-learning strategy according to claim 1, 2 or 3, characterized in that: The steps of obtaining the pre-training data set and the target data set include performing multiple groups of tests under different working conditions through the cyclic loading test in S1, obtaining the horizontal bearing capacity attenuation time series data of the pile foundation under different working conditions, and making them into multiple groups of data sets, wherein any one group of data sets is the target data set, and the remaining data sets are pre-training data sets.

6. The method for early warning of soft soil pile foundation bearing capacity attenuation based on meta-learning strategy according to claim 1, 2 or 3, characterized in that: The steps of obtaining the pre-training data set and the target data set include: performing multiple groups of tests under different working conditions through the cyclic loading test in S1, obtaining the horizontal bearing capacity attenuation time series data of the pile foundation under different working conditions, and making them into multiple groups of pre-training data sets, and the monitoring data in the actual working condition is the target data set.

7. The method for early warning of soft soil pile foundation bearing capacity attenuation based on meta-learning strategy according to claim 1 is characterized by: The cyclic loading test device includes a model box, a model pile, a loading system and a monitoring system. The loading system includes a horizontal cyclic loading system, a collector and a first computer. The first computer is used to set the loading frequency and the number of cycles, control the loading mode and record the load and displacement data changes at the loading position of the pile body. The monitoring system includes a laser displacement sensor, a dynamic demodulator and a second computer. The second computer records the wavelength changes of the pile body by connecting to the dynamic demodulator.

8. The method for early warning of soft soil pile foundation bearing capacity attenuation based on meta-learning strategy according to claim 6 is characterized in that: A target pile foundation monitoring system is established in the working condition, and the monitoring data is acquired by the monitoring system to obtain the target data set of the horizontal bearing capacity of the pile foundation in the actual working condition.

Images