Foundation pit deformation prediction method, foundation pit deformation prediction model training method and equipment

Abstract

The invention relates to the technical field of foundation pit engineering construction monitoring, and particularly discloses a foundation pit deformation prediction method, a foundation pit deformation prediction model training method and equipment. And then the data are respectively input into the corresponding prediction models to obtain prediction results, and the prediction results are fused, so that the prediction accuracy and robustness are improved, and more accurate and stable prediction of the monitoring data is realized. Besides, the prediction model is pre-trained through numerical simulation of the random field model, physical and mechanical characteristics of a geologic body can be fully considered to improve the adaptive performance of the prediction model under variable and complex geological conditions, and through multi-model fusion and integration, the over-fitting phenomenon can be reduced, and the prediction accuracy is improved. And good performance of the machine learning model under new data is ensured through transfer learning.

Description

Technical Field

[0001] The present invention relates to the technical field of foundation pit engineering construction monitoring, and in particular to a foundation pit deformation prediction method, a foundation pit deformation prediction model method and equipment. Background Art

[0002] In the existing foundation pit deformation monitoring, short-term and accidental external influences may cause the construction monitoring values ​​to show nonlinear and non-stationary characteristics. In this case, if prediction is made directly based on historical monitoring values, it is often difficult to obtain good prediction results, and it is impossible to effectively capture irregular accidental fluctuations and sudden changes. In addition, the "static" regression model is generally used in the existing prediction models. This regression model can only learn the collected monitoring time series data, and the training set size is generally small, which cannot effectively reflect the long-term change trend of the target monitoring value. In addition, most regression models rely on historical monitoring data of similar projects to train a general model, which makes it difficult to accurately speculate on the future changes of a specific project or a specific monitoring point, and cannot capture long-term change trends, and the generalization ability is poor. In addition, the existing prediction methods are mostly optimization algorithms combined with a single model for prediction. The prediction effect of a single model has an optimization upper limit, and the generalization performance of a single model is limited, which is prone to overfitting, resulting in large errors in the prediction results and loss of reliability. Summary of the invention

[0003] In order to overcome the problems existing in existing foundation pit deformation monitoring, such as the inability to effectively handle sudden changes and the limitations of the models used, the present invention provides a foundation pit deformation prediction method based on physical information machine learning, a foundation pit deformation prediction model training method and equipment.

[0004] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0005] In a first aspect, the present invention provides a method for predicting foundation pit deformation, the method comprising:

[0006] Based on the target monitoring points of the target foundation pit, real-time monitoring data of the target monitoring points are obtained;

[0007] Inputting the real-time monitoring data into the foundation pit deformation prediction model to obtain deformation prediction data of the target monitoring point;

[0008] The foundation pit deformation prediction model includes a trend term prediction model and a random term prediction model.

[0009] Inputting the real-time monitoring data into the foundation pit deformation prediction model specifically includes:

[0010] Decomposing the real-time monitoring data into first trend item data and first random item data;

[0011] Inputting the first trend item data into corresponding trend item prediction models through different preset time windows, and performing output fusion based on the weight of each trend item prediction model to obtain trend item prediction data;

[0012] Inputting the first random item data into a random item prediction model to obtain random item prediction data;

[0013] The trend item prediction data and the random item prediction data are integrated to obtain deformation prediction data of the target monitoring point. According to a specific implementation, in the above prediction method, the real-time monitoring data is decomposed into first trend item data and first random item data, including:

[0014] Decomposing the real-time monitoring data using variational mode decomposition to generate a number of mode functions;

[0015] The first mode function is taken as the first trend item data;

[0016] Combine the remaining modal functions into the first random term data.

[0017] According to a specific implementation, in the above-mentioned prediction method, the trend item prediction model is composed of multiple LSTM models, each LSTM model corresponds to a preset time window, and the weight of each LSTM model is determined by the inverse error method.

[0018] According to a specific implementation, in the above-mentioned prediction method, the random item prediction model includes a random forest regression model, a gradient boosting regression model and a linear regression model, and the first random item data are respectively input into the random forest regression model, the gradient boosting regression model and the linear regression model, and the output fusion is performed based on the weight of each model to obtain the random item prediction data; wherein the weight of each model is determined by the inverse error method.

[0019] In a second aspect, the present invention provides a method for training a foundation pit deformation prediction model, wherein the foundation pit deformation prediction model includes a trend item prediction model and a random item prediction model, and the method includes:

[0020] Based on the target monitoring point of the target foundation pit, historical monitoring data and numerical simulation data of the target monitoring point are obtained; wherein the numerical simulation data is deformation simulation data of the target monitoring point generated based on a random field model of the target foundation pit;

[0021] The numerical simulation data is used as a data set and input into the corresponding LSTM model through different preset time windows for training, and a weight is assigned to each LSTM model based on the output result to obtain a pre-trained trend item prediction model;

[0022] Decomposing the historical monitoring data into second trend item data and second random item data;

[0023] Inputting the second trend item data into the pre-trained trend item prediction model for transfer learning to obtain a trained trend item prediction model;

[0024] Inputting the second random item data into three different base learners for training, and assigning a weight to each base learner based on the output result to obtain a trained random item prediction model;

[0025] The trained trend item prediction model and the trained random item prediction model are integrated to obtain a foundation pit deformation prediction model. According to a specific implementation, in the above training method, weights are assigned to each LSTM model based on the output results, including:

[0026] Calculate the mean square error of each LSTM model output result, the formula is:

[0027]

[0028] Among them, m represents the number of output results obtained by the LSTM model; Indicates the nth output result; y n Represents the numerical simulation data corresponding to the nth output result;

[0029] The mean square error of each LSTM model is calculated using the inverse error method, and a weight is assigned to each LSTM model. The formula is:

[0030]

[0031] Among them, ε i , ε k represents the mean square error of the i-th and k-th LSTM models, I represents the total number of LSTM models, and w i Represents the weight of the i-th LSTM model.

[0032] According to a specific implementation, in the above training method, weights are assigned to each base learner based on the output results, which is the same as assigning weights to each LSTM model based on the output results; the trained random item prediction models include a random forest regression model, a gradient boosting regression model and a linear regression model.

[0033] According to a specific implementation, in the above-mentioned training method, the random field model is a number of foundation pit physical models generated by random field simulation based on the rock and soil parameters of the target foundation pit, and the rock and soil parameters include elastic modulus, cohesion and friction angle.

[0034] According to a specific implementation, in the above training method, decomposing the historical monitoring data into second trend item data and second random item data includes:

[0035] Splitting the historical monitoring data using variational mode decomposition to generate a number of mode functions;

[0036] The first mode function is used as the second trend item data;

[0037] Combine the remaining modal functions into the second random term data.

[0038] In a third aspect, the present invention provides an electronic device, the device comprising a memory and a processor;

[0039] Among them, the memory is used to store computer programs; the processor is used to call and execute the computer program, so that the device executes any of the above-mentioned methods for predicting foundation pit deformation or executes any of the above-mentioned methods for training a foundation pit deformation prediction model.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] Based on the technical solution of the first aspect provided by the present invention, the present invention improves the prediction accuracy and robustness by decomposing the monitoring data of the target foundation pit into trend item data and random item data, and then inputting them into the corresponding prediction model to obtain the prediction results and then integrating them, thereby achieving a more accurate and stable prediction of the monitoring data;

[0042] Based on the technical solution of the second aspect provided by the present invention, the present invention pre-trains the prediction model through numerical simulation of the random field model, which can fully consider the physical and mechanical properties of the geological body to improve the adaptability of the prediction model under changeable and complex geological conditions. In addition, the present invention can reduce the overfitting phenomenon through multi-model fusion and integration, and ensure the good performance of the machine learning model under new data through transfer learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A schematic diagram of a process flow of a foundation pit deformation prediction method provided by an embodiment of the present invention;

[0044] Figure 2 The overall flow chart of foundation pit deformation prediction provided by the embodiment of the present invention;

[0045] Figure 3 A schematic diagram of the modal decomposition of monitoring data provided by an embodiment of the present invention;

[0046] Figure 4 Another schematic diagram of monitoring data modal decomposition provided by an embodiment of the present invention;

[0047] Figure 5 A schematic diagram for comparing trend item data and random item data provided by an embodiment of the present invention;

[0048] Figure 6 A schematic diagram of elastic modulus distribution provided by an embodiment of the present invention;

[0049] Figure 7 A schematic diagram of cohesion distribution provided by an embodiment of the present invention;

[0050] Figure 8 A schematic diagram of friction angle distribution provided by an embodiment of the present invention;

[0051] Fig. 9 A schematic diagram of numerical simulation of deformation provided by an embodiment of the present invention;

[0052] Fig.10 A schematic diagram of the prediction effect of a pre-trained trend item prediction model provided in an embodiment of the present invention;

[0053] Fig.11 A schematic diagram of the prediction effect of an unpre-trained LSTM model provided in an embodiment of the present invention;

[0054] Fig.12 A schematic diagram of a random item prediction result provided by an embodiment of the present invention;

[0055] Fig.13 A schematic diagram of the prediction effect provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0057] The terms "first", "second", etc. in the embodiments of the specification, claims, and drawings of the present invention are used only for the purpose of distinguishing descriptions, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying an order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, including a series of steps or units. The method, system, product, or device is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0058] In addition, any embodiment or design described as "exemplary" or "for example" in the embodiments of the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way for easy understanding.

[0059] The present invention mainly splits the non-smooth nonlinear foundation pit deformation monitoring data with large fluctuations into trend items and random items, wherein the trend item retains the main trend direction of the original data, and the random item extracts the sudden fluctuations in the original data. The present invention then constructs prediction models for the trend item and the random item respectively for prediction, and finally combines the two into the foundation pit deformation prediction value.

[0060] Among them, the trend item reflects the long-term change trend of foundation pit deformation. The change of trend items is mainly affected by long-term factors such as geological conditions and construction procedures, and has strong regularity and time dependence. Therefore, the trend items can be predicted by time series prediction models. The pure data-driven deformation prediction model ignores the influence of physical factors such as geological conditions on the target monitoring value, resulting in limitations in the prediction results. For example, for a specific foundation pit project, its geological conditions cannot be accurately determined even through detailed surveys, and the pure data-driven deformation prediction model completely ignores the influence of geological body parameter uncertainty on foundation pit deformation. Therefore, the present invention performs numerical simulation of the excavation process based on the physical model, and introduces the simulation results into the foundation pit deformation prediction model. Specifically, the present invention takes into account the uncertainty of the physical and mechanical parameters of the rock and soil body (including the elastic modulus, cohesion and friction angle of the soil body). First, a physical model of the distribution of different rock and soil body parameters is established, and then the simulation numerical value is carried out to derive the time series data such as the settlement of the foundation pit periphery and the deformation of the foundation pit support during the process of foundation pit excavation to equilibrium. By generating a foundation pit deformation time series data set that takes into account the uncertainty of geological body parameters in the above manner, and using the time series data set to pre-train the foundation pit deformation trend item prediction model (using the physical information machine learning model to build the prediction model), the prediction model can learn the change law and potential trend of the monitoring data under different geological conditions, and improve its prediction ability for real monitoring data. After completing the pre-training of the foundation pit deformation trend prediction model, the foundation pit engineering deformation monitoring data is finally imported for transfer learning, that is, the foundation pit deformation trend prediction model training is completed, and the accurate prediction of the deformation trend of the specific foundation pit engineering construction is achieved. Among them, the random term is the high-frequency fluctuation part in the data, representing a short-term disturbance or sudden change, which is usually caused by a short-term external influence (such as random disturbances such as weather changes). Considering that the change of the random term does not have a strong time dependence, it is usually local and random. Therefore, the present invention directly uses the random term obtained by decomposing the foundation pit monitoring data as a training set, and uses a simple integrated learning model to construct a foundation pit deformation random term prediction model.

[0061] The following is a further introduction and description of the various methods provided in the embodiments of the present invention.

[0062] Specifically, please refer to Figure 1 , which shows a schematic flow chart of a method for predicting foundation pit deformation provided by an embodiment of the present invention, the method comprising:

[0063] S1. Based on the target monitoring point of the target foundation pit, obtain the real-time monitoring data of the target monitoring point;

[0064] S2. Input the real-time monitoring data into a foundation pit deformation prediction model to obtain deformation prediction data of target monitoring points.

[0065] The foundation pit deformation prediction model includes a trend item prediction model and a random item prediction model. The real-time monitoring data is input into the foundation pit deformation prediction model, which specifically includes:

[0066] Decomposing the real-time monitoring data into first trend item data and first random item data;

[0067] Inputting the first trend item data into the corresponding trend item prediction model through different preset time windows, and performing output fusion based on the weight of each trend item prediction model to obtain trend item prediction data;

[0068] Inputting the first random item data into a random item prediction model to obtain random item prediction data;

[0069] The trend item prediction data and the random item prediction data are integrated to obtain the deformation prediction data of the target monitoring point.

[0070] The embodiment of the present invention improves the prediction accuracy and robustness by decomposing the monitoring data of the target foundation pit into trend item data and random item data, and then inputting them into corresponding prediction models to obtain prediction results and then integrating them, thereby achieving more accurate and stable prediction of the monitoring data.

[0071] Specifically, the training method of the foundation pit deformation prediction model specifically includes:

[0072] Based on the target monitoring point of the target foundation pit, historical monitoring data and numerical simulation data of the target monitoring point are obtained; wherein the numerical simulation data is deformation simulation data of the target monitoring point generated based on a random field model of the target foundation pit;

[0073] The numerical simulation data is used as a data set and input into the corresponding LSTM model through different preset time windows for training, and a weight is assigned to each LSTM model based on the output result to obtain a pre-trained trend item prediction model;

[0074] Decomposing the historical monitoring data into second trend item data and second random item data;

[0075] Inputting the second trend item data into the pre-trained trend item prediction model for transfer learning to obtain a trained trend item prediction model;

[0076] Inputting the second random item data into three different base learners for training, and assigning a weight to each base learner based on the output result to obtain a trained random item prediction model;

[0077] The trained trend item prediction model and the trained random item prediction model are integrated to obtain a foundation pit deformation prediction model.

[0078] It is understandable that in the embodiment of the present invention, the real-time monitoring data includes deformation values ​​such as the settlement data around the foundation pit, the displacement data of the foundation pit support, and the uplift data of the pit bottom. Then, the real-time monitoring data is predicted by the method provided in the embodiment of the present invention to obtain an accurate prediction value. The embodiment of the present invention pre-trains the prediction model through numerical simulation of the random field model, which can fully consider the physical and mechanical properties of the geological body to improve the adaptability of the prediction model under changeable and complex geological conditions. In addition, the present invention can reduce overfitting through multi-model fusion and integration, and ensure the good performance of the machine learning model under new data through transfer learning.

[0079] The present invention is further introduced and illustrated below in conjunction with specific implementation modes.

[0080] Generally, in the actual monitoring process, the above foundation pit deformation prediction method and foundation pit deformation prediction model training method are carried out simultaneously. Figure 2 , which shows a schematic diagram of the overall process of foundation pit deformation prediction provided by an embodiment of the present invention.

[0081] Suppose you want to predict the deformation value of a specific foundation pit during construction:

[0082] (1) Historical monitoring data preprocessing and variational mode decomposition

[0083] The deformation monitoring points of a specific foundation pit continuously collect monitoring data as historical monitoring data during the construction process. After missing value supplementation, normalization and other processing, the collected historical monitoring data are split into second trend item data and second random item data using the variational mode decomposition (VMD) method. The decomposed trend item reflects the long-term trend of foundation pit deformation during construction, while the random item contains information about short-term, random, and accidental fluctuations in foundation pit deformation. The above-mentioned decomposed second trend item data and second random item data will be used for the training of the trend item prediction model and the random item prediction model of the foundation pit deformation prediction model, respectively.

[0084] (2) Generation of numerical simulation data (physical information)

[0085] Considering the uncertainty of rock and soil parameters (elastic modulus, cohesion and friction angle), a large number of foundation pit physical model samples are generated. The samples reflect the physical characteristics of the foundation pit (such as excavation depth, support form, and surrounding building distribution), and the random field simulation of rock and soil parameters is used to generate samples with different distributions of rock and soil parameters. Numerical simulation of the foundation pit excavation process is carried out based on the physical model samples. By setting up History monitoring points, deformation simulation data of target monitoring points (such as deformation at the top of the support structure, settlement around the foundation pit, etc.) are collected as numerical simulation data. These data reflect the deformation trend that may occur during the foundation pit construction process under different distributions of rock and soil parameters.

[0086] (3) Pre-training of trend item prediction model (physical information machine learning model):

[0087] In order to predict the deformation trend of foundation pit, a physical information machine learning model is constructed based on the transfer learning strategy as the trend item prediction model of foundation pit deformation. That is, the physical model data (numerical simulation data) is used to pre-train the model, and then the experimental data (foundation pit monitoring data) is introduced for transfer learning to enhance the performance of the model on the new task (foundation pit deformation prediction).

[0088] Specifically, the numerical simulation data is first used as a data set and input into the corresponding LSTM model through different preset time windows for training. In order to simultaneously capture the long-term and short-term characteristics of the foundation pit deformation trend, LSTM models with different time window sizes (such as 24h, 48h, 72h, 96h, 120h) are used as base models, and data preparation is performed on the data set according to the window size, and time series training data sets with different time window sizes are created to train the LSTM model. For each LSTM model, the mean square error on its test set is calculated, and then the inverse error method is used to assign weights to the LSTM models of different window sizes, and the prediction results of the LSTM model are weighted fused to obtain the final foundation pit deformation prediction results. In this way, the pre-training of the trend item prediction model is completed, and the prediction model has a good prediction ability for the foundation pit construction deformation trend under different geological conditions.

[0089] (4) Transfer learning of trend item prediction model

[0090] The second trend item data decomposed in step (1) is used to perform transfer learning of the trend item prediction model, and the second trend item data is input into the pre-trained trend item prediction model for re-training to fine-tune the model parameters. Pre-training enables the physical information machine learning model to obtain knowledge of the deformation law and response characteristics of the foundation pit under different working conditions; while transfer learning enables the model to identify the characteristics and change trends of real-time monitoring data in a specific project. Therefore, the trend item prediction model after transfer learning using specific foundation pit monitoring data can predict the deformation trend of the foundation pit.

[0091] (5) Random item prediction model (ensemble learning model) training

[0092] The processing of random items is equally important in the prediction of foundation pit deformation. The embodiment of the present invention chooses to build a simple and efficient integrated model to train and predict random items. The integrated model consists of three different base learners, namely, a random forest regression model (RandomForestRegressor), a gradient boosting regression model (GradientBoostingRegressor) and a linear regression model (LinearRegression). Each of these three has unique advantages and characteristics. Combining the three to form an integrated learning model can improve the overall prediction performance. The second random item data decomposed in step (1) is used as a training data set, input into the three base learners for training, and then the weights are assigned to the three base learners by the inverse error method to obtain a trained random item prediction model. This method can not only reduce the accidental errors caused by the prediction of a single model, but also better reflect the overall trend of the data.

[0093] (6) Model prediction

[0094] Based on the target monitoring points of the target foundation pit, the real-time monitoring data is decomposed to obtain the first trend item data and the first random item data, which are input into the trend item prediction model and the random item prediction model respectively to make a comprehensive prediction of the subsequent deformation of the foundation pit. The specific steps are as follows:

[0095] Trend item prediction: input the first trend item data into the trend item prediction model, and output the trend item prediction value Tpred(t) of the foundation pit deformation at a subsequent moment, where t represents the time point.

[0096] Random item prediction: At the same time, the first random item data is input into the random item prediction model, and the random item prediction value Rpred(t) of the foundation pit deformation at a subsequent moment is output.

[0097] Final deformation prediction value: Finally, the total predicted value of foundation pit deformation Dpred(t) is calculated. Foundation pit deformation is a comprehensive result based on long-term trends and short-term fluctuations. Therefore, the prediction results of trend terms and random terms are combined to obtain:

[0098] Dpred(t)=Tpred(t)+Rpred(t).

[0099] Through the above steps, the deformation of the target foundation pit can be predicted.

[0100] The solution proposed in this invention not only improves the accuracy of prediction, but also significantly enhances the adaptability of the model under complex and uncertain conditions. The specific effects are mainly reflected in the following aspects:

[0101] (1) Improved prediction accuracy and reliability: By decomposing the monitoring data into trend items and random items, the present invention can apply targeted prediction models to predict the two items respectively. The trend item prediction model can capture the long-term change trend of foundation pit deformation, while the random item prediction model focuses on short-term random fluctuations. This decomposition method effectively improves the overall prediction accuracy, making the prediction results more consistent with the actual deformation process of the foundation pit.

[0102] (2) Strengthened consideration of the impact of geological conditions: The present invention generates numerical simulation data through a physical model, fully considering the uncertainty and complexity of rock and soil parameters. Compared with the traditional pure data-driven model, the pre-training combined with geological characteristics makes the trend item prediction model more robust in dealing with diverse geological conditions, effectively reducing the safety hazards caused by the uncertainty of geological conditions.

[0103] (3) Improved the generalization ability of the model: By implementing the strategies of transfer learning and ensemble learning, the present invention significantly enhances the generalization ability of the model. Transfer learning enables the model to quickly adapt to the monitoring data of a specific foundation pit, thereby improving the prediction performance of the trend item prediction model in a new environment. In addition, ensemble learning combines the strengths of different base learners, reduces the overfitting risk of the random item prediction model, and ensures that the model can still provide reliable predictions when faced with inputs of various postures.

[0104] (4) Strong adaptability. Due to the adoption of an innovative method based on the combination of physical modeling and machine learning, the present invention has strong adaptability and can be widely used in different types of foundation pit engineering projects. It can also be further extended to various geotechnical engineering deformation prediction tasks such as slopes, tunnels, and foundations, thereby improving the transformation potential of scientific and technological achievements.

[0105] In summary, the present invention significantly improves the prediction ability and adaptability of foundation pit deformation by innovatively combining the advantages of data-driven and physical models, and provides strong technical support for the safety management and construction optimization of foundation pit projects.

[0106] Furthermore, the technical solution of the present invention is introduced and described in detail below in conjunction with specific embodiments.

[0107] In order to illustrate the specific implementation steps of the embodiment of the present invention, a foundation pit project is selected to predict the deformation of the top of the foundation pit support. The foundation pit project arranges GNSS automatic displacement monitoring points at the top of the support. The monitoring data collection time is from March 15, 2024 to June 21, 2024. A total of 2206 monitoring data (hereinafter referred to as historical monitoring data) are obtained, and the monitoring frequency is 1 item / hour.

[0108] Data partitioning:

[0109] Training set and test set: For the historical monitoring data from March 15 to June 2, 2024, the first 80% of the data is used as the training set and the last 20% of the data is used as the test set.

[0110] Validation set: The time range is from June 3 to June 21, 2024, which is used to compare with the foundation pit deformation predicted by the method of the present invention to illustrate the prediction accuracy of the method.

[0111] Step (1): Monitoring data variational mode decomposition.

[0112] The historical monitoring data from March 15 to June 2, 2024 are decomposed into the second trend item data and the second random item data through variational mode decomposition (VMD). The decomposition process is as follows:

[0113] ①Data preprocessing

[0114] Due to signal blocking and other reasons, some historical monitoring data are missing, so it is necessary to supplement the missing values ​​of historical monitoring data. In order to improve the training efficiency, the data needs to be normalized. The preprocessed historical monitoring data is recorded as x(t).

[0115] ②VMD decomposition parameter setting

[0116] The VMD function in the vmdpy library in Python is used to decompose the historical monitoring data. The key input parameters include:

[0117] 1. data: the input signal or time series data to be decomposed, i.e. GNSS automated monitoring data. Here, GNSS automated monitoring data needs to be preprocessed by missing value supplementation and normalization before entering the VMD function.

[0118] 2.alpha: Penalty coefficient used to control the sparsity of the mode. Larger values ​​will result in smoother modes, while smaller values ​​may result in more complex modes.

[0119] 3.tau: time smoothing factor, which controls the smoothness of the mode in time.

[0120] 4.K: The number of extracted modes, which indicates how many different frequency components you want to extract from the input signal.

[0121] 5.tol: Convergence tolerance, used to control the convergence conditions of the algorithm. Smaller values ​​may lead to more accurate results, but may also increase the calculation time.

[0122] In specific applications, the above parameters need to be adjusted and the optimal values ​​of each parameter need to be determined through experiments.

[0123] ③VMD variational mode decomposition

[0124] The VMD goal is to decompose the historical monitoring data x(t) into several modal functions u k (t) (where k = 1, 2, ..., K). For example, when K = 3, the decomposition result is as follows Figure 3 As shown:

[0125] x(t)=u1(t)+u2(t)+u3(t).

[0126] The embodiment of the present invention uses the first modal function u1(t) as the second trend item data, and combines all modal functions u2(t) and u3(t) other than u1(t) into the second random item data. After comparative tests, for the historical monitoring data of this embodiment, when the number of extracted modes = 4 (K = 4) and other key parameters take the default values, the best decomposition effect can be obtained. The historical monitoring data from March 15 to June 2, 2024 is decomposed, and the decomposition results are as follows: Figure 5 As shown, the second trend item data and the second random item data are compared as follows: Figure 6 shown.

[0127] Step (2): Numerical simulation data generation (physical information).

[0128] The variability of rock and soil parameters (elastic modulus, cohesion and friction angle, etc.) has an important impact on the stability and deformation characteristics of the support structure, but it is impossible to directly estimate the rock and soil parameters accurately, and naturally it is impossible to know the accurate displacement trend of the target monitoring point. Therefore, the embodiment of the present invention obtains the possible displacement trend of the target monitoring point by numerical simulation, taking into account the variability of rock and soil parameters in advance, for pre-training of the trend item prediction model. The specific steps are as follows:

[0129] ①Physical model building

[0130] Firstly, a numerical model of the foundation pit is constructed according to the geometric characteristics of the foundation pit and the statistical characteristics of rock and soil parameters.

[0131] ② Generation of foundation pit physical model samples considering parameter variability

[0132] The spatial variability of the physical and mechanical parameters of the rock and soil (elastic modulus, cohesion and friction angle) was simulated by random field method, and 500 sets of rock and soil parameter distribution samples were generated. The 500 sets of parameter distribution samples were assigned to the foundation pit numerical model, and 500 sets of foundation pit physical model samples were obtained.

[0133] The geological stratification and rock and soil parameters are shown in Table 1.

[0134] Table 1 is an example table of geological stratification and rock and soil parameters

[0135]

[0136] The elastic modulus, cohesion and friction angle distribution of a typical sample are shown as follows: Figures 6 to 8 shown.

[0137] ③ Numerical simulation

[0138] FLAC3D was used to simulate the soil excavation and support structure construction process of 500 samples. A History displacement collection point was set at a distance of 2m from the top edge of the slope to generate 500 sets of numerical simulation data of the deformation of the foundation pit support top under different distributions of rock and soil parameters. Fig. 9 shown.

[0139] Step (3): Pre-training of trend item prediction model.

[0140] The trend of foundation pit engineering monitoring data is affected by multiple time scales such as construction procedures, design changes, temperature changes, humidity changes, rainfall cycles, and changes in the surrounding environment. Using a fixed time window may not be able to fully capture the impact of these complex factors on the monitoring data. In order to effectively capture features at different time scales, the embodiment of the present invention constructs multiple LSTM models, each model corresponds to a specific time window, and finally the outputs of all models are fused. The steps for constructing and training the foundation pit deformation trend item prediction model are as follows:

[0141] ①Time window setting:

[0142] Each LSTM model corresponds to a different preset time window. In this embodiment, five preset time windows are set to 24 hours, 48 ​​hours, 72 hours, 96 hours and 120 hours, corresponding to 5 LSTM models. That is, the data of the first 24 hours are used to predict the data of the next hour (LSTM-24), the data of the first 48 hours are used to predict the data of the next hour (LSTM-48), the data of the first 72 hours are used to predict the data of the next hour (LSTM-72), the data of the first 96 hours are used to predict the data of the next hour (LSTM-96), and the data of the first 120 hours are used to predict the data of the next hour (LSTM-120).

[0143] ②Preparation of training data set:

[0144] The numerical simulation data obtained in step (2) are used to prepare training data sets according to different window sizes, and a total of 5 training data sets are obtained (corresponding to LSTM models with different window sizes).

[0145] ③LSTM model training:

[0146] The training data sets are input into the LSTM model for training.

[0147] ④Model output fusion:

[0148] According to the performance of each LSTM model on the test set, the output results of multiple LSTM models are fused through the inverse error method. Fusion prediction can integrate the information provided by LSTM models in different time windows, so combined prediction can improve prediction accuracy more effectively than single methods. The inverse error method is a method for determining weighted coefficients in combined prediction. In the inverse error method, the base model with smaller error is given a larger weight. The weight calculation formula of the inverse error method is as follows:

[0149]

[0150] Where k∈{24,48,72,96,120},k≠i; i∈{24,48,72,96,120}.

[0151] w i is the fusion weight of the i-th LSTM model, ε i is the root mean square error (RMSE) of the i-th LSTM model on the test set. The calculation formula of the root mean square error (RMSE) is as follows:

[0152]

[0153] Where m represents the number of predicted values; Indicates the nth output result; y n Represents the numerical simulation data corresponding to the nth output result.

[0154] For example, the weight calculation formulas for LSTM-24 and LSTM-120 are as follows:

[0155] w 24 =(ε 48 +ε 72 +ε 96 +ε 120 ) / 2(ε 24 +ε 48 +ε 72 +ε 96 +ε 120 ),

[0156] w 120 =(ε 24 +ε 48 +ε 72 +ε 96 ) / 2(ε 24 +ε 48 +ε 72 +ε 96 +ε 120 ).

[0157] The inverse error method can effectively weight the prediction results of each model, so that the model can highlight the better performance in the synthesized output, thereby improving the overall prediction accuracy. Experiments show that the LSTM multi-sliding window based on the inverse error method fusion performs well in capturing the complex change trend of foundation pit engineering monitoring data, which helps to improve prediction accuracy and reliability and provide more accurate support for engineering safety monitoring.

[0158] According to the output results of the LSTM model, the errors and weights of the LSTM models with different window sizes are shown in Table 2.

[0159] Table 2 shows the error and weight examples of LSTM models with different window sizes.

[0160] Model Window size <![CDATA[Root mean square error ε i > <![CDATA[Weight w i > LSTM-24 24 hours 0.023 0.380208 LSTM-48 48 hours 0.018 0.40625 LSTM-72 72 hours 0.015 0.421875 LSTM-96 96 hours 0.019 0.401042 LSTM-120 120 hours 0.021 0.390625

[0161] The calculation formula of the deformation trend item prediction value T outputted in the end is as follows:

[0162] T(t)=∑ i w i T i (t) = w 24 T 24 +w 48 T 48 +w 24 T 24 +w 24 T 24 +w 24 T 24 ,

[0163] T i (t) is the predicted value of the deformation trend item of the LSTM model at time t, i∈{24,48,72,96,120}.

[0164] Step (4): Transfer learning and prediction of trend item prediction model.

[0165] The second trend item data (March 15, 2024 to June 2, 2024) obtained by VMD decomposition in step (1) is used as the target data for transfer learning, and is input into the pre-trained trend item prediction model for transfer learning.

[0166] After completing the transfer learning, the trend item prediction model is used to predict the subsequent foundation pit deformation monitoring data trend (June 3 to June 21, 2024) to obtain the trend item prediction data. Please refer to Fig.10 , which is a schematic diagram of the prediction effect of the pre-trained trend item prediction model provided in an embodiment of the present invention. Please refer to Fig.11, which is a schematic diagram of the prediction effect of the LSTM model without pre-training provided in the embodiment of the present invention. It can be foreseen that the pre-trained model can make the prediction value constrained by the source data (deformation trend of the foundation pit numerical simulation), while the directly trained LSTM model can only capture the characteristics of the limited field monitoring data itself, verifying the reliability of the physical information machine learning model.

[0167] Step (5): Random item prediction model training and prediction.

[0168] For random items in foundation pit deformation monitoring, the embodiment of the present invention constructs an integrated model for random item prediction, which includes three different regression models: Random Forest Regressor, Gradient Boosting Regressor and Linear Regression. By combining the outputs of multiple regression base models, the generalization ability of the model is enhanced and the risk of overfitting is reduced.

[0169] The specific process is as follows:

[0170] ①Preparation of training data set

[0171] The second random item data from March 15 to June 2, 2024 obtained by decomposition in step (1) is used to prepare a training data set according to a window size of 24 hours.

[0172] ②Learning model training

[0173] The three base learners (random forest regression, gradient boosting regression and linear regression) are trained using the above training data sets.

[0174] ③Model integration output

[0175] The three trained base learners are used to predict the subsequent foundation pit deformation monitoring data fluctuations (June 3 to June 21, 2024). The inverse error method (see step (3) for details) is also used to fuse the prediction results of the three learning models as the final random item prediction model output R(t).

[0176] R(t)=∑ i v i R i (t)=v1R1(t)+v2R2(t)+v2R2(t),

[0177] R i is the predicted value of the deformation random term of the i-th base learner at time t, v i is the weight of the i-th base learner, i∈{1, 2, 3}.

[0178] The root mean square errors of the three base learners on the test set and the corresponding weights calculated are shown in Table 3.

[0179] Table 3 is an example table of the root mean square error and weights of the three base learners.

[0180] Model <![CDATA[Root mean square error ε i > <![CDATA[Weight v i > Random Forest 0.017 0.339623 Gradient Boosting 0.016 0.349057 Linear Regression 0.020 0.311321

[0181] The random term prediction model is used to predict the subsequent random fluctuations of foundation pit deformation monitoring data (from June 3 to June 21, 2024). The performance of the random term prediction model on the validation set is as follows: Fig.12 shown.

[0182] Step (6): Deformation prediction model verification.

[0183] The pit deformation is the comprehensive result of long-term trend and short-term fluctuation. Therefore, the trend item prediction data T(t) obtained in step (4) and the random item prediction data R(t) obtained in step (5) are fused to obtain the final pit deformation prediction value.

[0184] D(t)=T(t)+R(t).

[0185] In summary, through the technical solution provided by the embodiment of the present invention, a trend item prediction model for foundation pit deformation prediction is constructed based on the physical information machine learning model, and a random item prediction model for foundation pit deformation prediction is constructed based on the integrated learning model. Compared with the general LSTM prediction model, the prediction accuracy and reliability can be improved, and the problem of prediction value lag in the LSTM model can be solved. For comparison, please refer to Fig.13 The present invention provides a new idea for deformation monitoring and prediction of foundation pit engineering.

[0186] On the other hand, an embodiment of the present invention further provides an electronic device, which includes a processor, a network interface and a memory, wherein the processor, the network interface and the memory are interconnected, wherein the memory is used to store a computer program, and the computer program includes program instructions, and the processor is configured to call the program instructions to execute the above-mentioned slope support reliability assessment method.

[0187] In the embodiments of the present invention, the processor may be an integrated circuit chip having the ability to process signals. The processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0188] The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present invention can be directly embodied as a hardware decoding processor for execution, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The processor reads the information in the storage medium and completes the steps of the above method in combination with its hardware.

[0189] The storage medium may be a memory, which may be, for example, a volatile memory or a nonvolatile memory, or may include both volatile and nonvolatile memory.

[0190] Among them, the non-volatile memory can be a read-only memory (ROM), a programmable ROM (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory.

[0191] The volatile memory may be a random access memory (RAM) which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), sync link DRAM (SLDRAM), and direct rambus RAM (DRRAM).

[0192] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0193] It should be understood that the system disclosed in the present invention can be implemented in other ways. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the communication connection between modules can be through some interfaces, indirect coupling or communication connection of servers or units, which can be electrical or other forms.

[0194] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, each module may exist physically separately, or two or more modules may be integrated into one processing unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional unit.

[0195] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0196] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting foundation pit deformation, characterized in that: The method comprises: Based on the target monitoring points of the target foundation pit, real-time monitoring data of the target monitoring points are obtained; Inputting the real-time monitoring data into the foundation pit deformation prediction model to obtain deformation prediction data of the target monitoring point; The foundation pit deformation prediction model includes a trend term prediction model and a random term prediction model. Inputting the real-time monitoring data into the foundation pit deformation prediction model specifically includes: Decomposing the real-time monitoring data into first trend item data and first random item data; Inputting the first trend item data into corresponding trend item prediction models through different preset time windows, and performing output fusion based on the weight of each trend item prediction model to obtain trend item prediction data; Inputting the first random item data into a random item prediction model to obtain random item prediction data; The trend item prediction data and the random item prediction data are integrated to obtain the deformation prediction data of the target monitoring point.

2. A foundation pit deformation prediction method according to claim 1, characterized in that: Decomposing the real-time monitoring data into first trend item data and first random item data includes: Decomposing the real-time monitoring data using variational mode decomposition to generate a number of mode functions; The first mode function is taken as the first trend item data; Combine the remaining modal functions into the first random term data.

3. A foundation pit deformation prediction method according to claim 1, characterized in that: The trend item prediction model is composed of multiple LSTM models, each LSTM model corresponds to a preset time window, and the weight of each LSTM model is determined by the inverse error method.

4. A foundation pit deformation prediction method according to claim 1, characterized in that: The random item prediction model includes a random forest regression model, a gradient boosting regression model and a linear regression model. The first random item data are respectively input into the random forest regression model, the gradient boosting regression model and the linear regression model, and the output fusion is performed based on the weight of each model to obtain the random item prediction data; wherein the weight of each model is determined by the inverse error method.

5. A foundation pit deformation prediction model training method, characterized in that: The foundation pit deformation prediction model includes a trend term prediction model and a random term prediction model, and the method includes: Based on the target monitoring point of the target foundation pit, historical monitoring data and numerical simulation data of the target monitoring point are obtained; wherein the numerical simulation data is deformation simulation data of the target monitoring point generated based on a random field model of the target foundation pit; The numerical simulation data is used as a data set and input into the corresponding LSTM model through different preset time windows for training, and a weight is assigned to each LSTM model based on the output result to obtain a pre-trained trend item prediction model; Decomposing the historical monitoring data into second trend item data and second random item data; Inputting the second trend item data into the pre-trained trend item prediction model for transfer learning to obtain a trained trend item prediction model; Inputting the second random item data into three different base learners for training, and assigning a weight to each base learner based on the output result to obtain a trained random item prediction model; The trained trend item prediction model and the trained random item prediction model are integrated to obtain a foundation pit deformation prediction model.

6. A foundation pit deformation prediction model training method according to claim 5, characterized in that: Assign weights to each LSTM model based on the output results, including: Calculate the mean square error of each LSTM model output result, the formula is: Among them, m represents the number of output results obtained by the LSTM model; Indicates the nth output result; y n Indicates the monitoring simulation data corresponding to the nth output result; The mean square error of each LSTM model is calculated using the inverse error method, and a weight is assigned to each LSTM model. The formula is: Among them, ε i , ε k represents the mean square error of the i-th and k-th LSTM models, I represents the total number of LSTM models, and w i Represents the weight of the i-th LSTM model.

7. A foundation pit deformation prediction model training method according to claim 6, characterized in that: Assigning weights to each base learner based on the output results is the same as assigning weights to each LSTM model based on the output results; the trained random item prediction models include a random forest regression model, a gradient boosting regression model, and a linear regression model.

8. A foundation pit deformation prediction model training method according to claim 5, characterized in that: The random field models are several foundation pit physical models generated by random field simulation based on the rock and soil parameters of the target foundation pit, and the rock and soil parameters include elastic modulus, cohesion and friction angle.

9. A foundation pit deformation prediction model training method according to claim 5, characterized in that: Decomposing the historical monitoring data into second trend item data and second random item data includes: Splitting the historical monitoring data using variational mode decomposition to generate a number of mode functions; The first mode function is used as the second trend item data; Combine the remaining modal functions into the second random term data.

10. An electronic device, characterized in that: The device comprises a memory and a processor; In which, the memory is used to store computer programs; the processor is used to call and execute the computer program so that the device executes a foundation pit deformation prediction method described in any one of claims 1 to 4 or executes a foundation pit deformation prediction model training method described in any one of claims 5 to 9.

Images