Machine learning-based membrane bag sand dike slope whole cycle safety management and control method and system

Abstract

The application discloses a machine learning-based membrane bag sand dike slope whole-cycle safety management and control method and system, which comprises the following steps: in the design stage, based on an initial design scheme, a classification model is used to judge a failure mode, a corresponding regression prediction model is selected to predict a limit filling height H max and a safety factor F s , the design scheme is compared with a safety setting threshold [F s ], and it is judged whether the design scheme is feasible; if the design scheme is not feasible, an optimization model is used to generate an optimal scheme; in the construction stage, displacement monitoring data is output to new ground soil parameters through an inversion model, the original design parameters are replaced, and a scheme is updated; after the monitoring data passes through a classification model, a time sequence model is used to predict a safety factor F sp , a risk level is judged in combination with an [Fs] early warning threshold, and if the safety factor exceeds a safety range, an early warning is given and an optimal adjustment scheme is generated through the optimization model. The application constructs a whole-cycle closed loop of design prediction, construction updating, real-time evaluation and optimization adjustment, combines monitoring data with a safety factor, and realizes dynamic whole-cycle stability evaluation and adjustment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of engineering intelligent management and control, and particularly relates to a membrane bag sand dike slope full-cycle safety management and control method and system based on machine learning. BACKGROUND

[0002] The membrane bag sand dike slope has become the core enclosing structure of reclamation, offshore dike protection and waterway regulation projects due to its high degree of mechanization, fast construction speed, strong overall stability and strong adaptability. However, the soft soil foundation has the characteristics of low strength and large deformation, and the dike slope project has the pain points of design relying on experience analogy, construction monitoring lagging behind and safety warning not being timely, which leads to frequent cases of engineering instability. Under the condition of soft foundation, there may be "overall failure" or "partial failure" modes, such as dike slope subsidence or combined sliding failure caused by insufficient foundation bearing capacity. In the design aspect, the conventional design often uses equal-thickness sand bags and uniform-strength materials, which may not fully strengthen the bottom and may affect the bearing capacity, and the quality of the membrane bag dike slope also depends on the type, ratio and filling process of the filling material.

[0003] The traditional stability analysis and control method of the membrane bag sand dike slope mainly relies on the safety factor prediction in the design stage and the real-time monitoring on site in the construction stage. However, as a flexible structure, the membrane bag sand dike slope has continuous coordinated deformation and interaction between its own structure and the foundation soil during the construction stage and the normal use stage after the construction is completed. With the progress of the drainage consolidation process, the cofferdam stiffness and the foundation soil strength change with time, which leads to the dynamic adjustment of the overall stability of the dike slope. In addition, the flexible characteristics of the dike slope and the properties of the soft foundation soil together lead to large dike slope foundation deformation and a long process, and it is difficult to accurately judge the stability of the dike slope only by deformation monitoring. Therefore, the real-time monitoring data in the construction process cannot be effectively linked with the safety factor, which makes it difficult for engineering personnel to accurately judge the stability state of the dike slope, and only when the monitoring data appears obvious abnormalities can corresponding measures be taken, without making early predictions and taking corresponding measures in advance according to the trend of the monitoring data.

[0004] In view of the above problems, it is urgent to build an intelligent stability control method and system that integrates multi-source information and is suitable for the whole cycle process of the membrane bag sand dike slope. Machine learning technology provides strong technical support for solving this engineering problem due to its significant advantages in multi-modal data fusion, dynamic prediction and adaptive optimization. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a membrane bag sand dike slope full-cycle safety management and control method and system based on machine learning, which builds a closed loop of design prediction-construction update-real-time evaluation-optimization adjustment, covers the whole process from design to construction, can make early warning and timely response adjustment, and continuously optimizes the scheme.

[0006] The application is realized by the following technical solutions:

[0007] The membrane bag sand dike slope full-cycle safety management method based on machine learning comprises the following steps:

[0008] S1, in the design stage:

[0009] S1-1, based on the initial design scheme, the failure mode is judged by the classification model, and the design scheme comprises the membrane bag sand dike slope design parameter and the dike slope section geometric parameter;

[0010] S1-2, the corresponding regression prediction model is selected according to the failure mode obtained in the step S1-1, and the limit filling height H max And the safety factor F s Are predicted by the corresponding regression prediction model;

[0011] S1-3, by comparing the predicted safety factor F s With the safety factor setting threshold [F s ], it is judged whether the design scheme is feasible, if feasible, the design scheme is the optimal scheme, if not feasible, the optimal scheme is generated by the optimization model, and the optimal scheme comprises the membrane bag sand dike slope design parameter and the dike slope section geometric parameter;

[0012] S2, in the construction stage:

[0013] S2-1, the new foundation soil body parameter is obtained by the inversion model through the construction monitoring data, the original foundation soil body parameter is replaced by the new foundation soil body parameter, and the optimal scheme obtained in the step S1 is updated;

[0014] S2-2, the new optimal scheme and the new construction monitoring data are judged by the classification model according to the failure mode;

[0015] S2-3, according to the failure mode, the corresponding time sequence model is selected, and the safety factor F sp Is predicted by the corresponding time sequence model;

[0016] S2-4, by comparing the predicted safety factor F sp With the safety factor setting threshold [F s ], the safety risk level is judged, if the safety risk level is within the set safety range, no warning is needed and continuous monitoring is needed, if the safety risk level exceeds the set safety range, warning is carried out and the optimal scheme is generated by the optimization model, and the optimal scheme comprises the membrane bag sand dike slope design parameter and the dike slope section geometric parameter;

[0017] S2-5, during construction, the safety evaluation and scheme optimization are continuously carried out according to the method of S2-2 to S2-4 at regular or irregular time until the construction is completed.

[0018] Further, the membrane bag sand embankment slope design parameters include the tensile stiffness J of the geomembrane bag, the undrained shear strength S of the soft soil u , the shear strength τ of the contact surface between the geomembrane bag and the foundation, the filling thickness t of the membrane bag sand, the internal friction angle φ of the filling sand, the unit weight γ of the foundation soil c and the unit weight γ of the membrane bag filling sand s ; the embankment slope section geometric parameters include the embankment slope width W, the embankment slope height H and the embankment slope gradient k; the foundation soil parameters include the undrained shear strength S u of the soft soil; and the construction monitoring data includes displacement data.

[0019] Further, the classification model adopts a random forest classifier, taking into account the noise tolerance and classification accuracy of engineering data, and the output failure modes include overall failure and side failure;

[0020] The training method of the classification model is:

[0021] Collecting data, the data sources include actual engineering data (instability cases, including failure mode labels), simulation data and indoor scale physical model test data, the indoor scale physical model test data includes data under remolded silt working conditions;

[0022] The data is randomly divided into training set and test set according to 7:3;

[0023] The hyperparameters are optimized by 5-fold cross-validation, the number of decision trees is 100, and the maximum depth of each tree is 8.

[0024] Further, the regression prediction model adopts an XGBoost regression model structure, and the training method of the regression prediction model is:

[0025] The data sources of the training set include simulation data and indoor scale physical model test data, the indoor scale physical model test data includes limit bearing capacity and safety factor data derived from embankment load-settlement curve;

[0026] The test set data includes actual engineering data;

[0027] The hyperparameters are optimized by grid search, the learning rate is 0.1, the tree depth is 6, and the iteration number is 200.

[0028] Further, the optimization model comprises an NSGA-III optimization algorithm, the classification model and the regression prediction model, and the working process is: under the condition of clear optimization target and constraint condition, a candidate scheme is generated by the NSGA-III optimization algorithm;

[0029] The candidate scheme is judged by the classification model according to the failure mode, the corresponding regression prediction model is selected according to the failure mode, and the safety factor F s is obtained by the corresponding regression prediction model.

[0030] According to the safety factor F s of each candidate scheme, a feasible scheme is screened out.

[0031] Then, the optimal scheme is selected from the feasible scheme by comprehensively considering the economic and safety factors.

[0032] Further, the method for screening out the feasible scheme according to the safety factor F s of each candidate scheme is:

[0033] If it is the design stage, the safety factor F s of the candidate scheme is compared with the safety factor set threshold [F s ], if F sp s ≥ [F s ], the candidate scheme is a feasible scheme, otherwise, it is an infeasible scheme.

[0034] If it is the construction stage, the safety factor F sp of the candidate scheme is compared with the safety factor set threshold [F s ], and the safety risk level is judged, if the safety risk level is within the set safety range, it is a feasible scheme, otherwise, it is an infeasible scheme.

[0035] The constraint condition corresponding to the design stage is W∈30-120m, J∈140-8000kN / m, S u ≥5kPa, and the constraint condition corresponding to the construction stage is that the adjustment range is ≤20% of the original parameter, so as to avoid large changes in the construction scheme.

[0036] Further, the training method of the NSGA-III optimization algorithm is: 2000 groups of feasible schemes generated by simulation numerical simulation are used as initial population, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and the iteration number is set to 50 times.

[0037] Further, the inversion model adopts a BP neural network based on Bayesian optimization (BO-BP), which takes into account the inversion accuracy and calculation efficiency, the input data is monitoring displacement data, and the output data is soft soil undrained shear strength S u .

[0038] The training method of the inversion model is:

[0039] The data sources of the training set include actual engineering data (construction measured displacement and soil test data) and simulation data, which include input parameters-displacement responses simulated under the condition of taking values of 2.5-15 kPa, corresponding to different displacement outputs. u The data simulated under the input parameter-displacement response condition taking values of 2.5-15 kPa correspond to different displacement outputs.

[0040] The optimal initial weights and structure of the BP neural network are searched through Bayesian optimization, including the number of hidden layers and the number of neurons in each layer.

[0041] Further, the time series model adopts an XGBoost-LSTM coupled model, which includes an XGBoost module, an LSTM module and a feature fusion module. The XGBoost module is used for static nonlinear feature extraction of the static data in the optimal scheme, the LSTM module is used for time series feature extraction of the construction monitoring data, and the feature fusion module is used for splicing the static nonlinear feature vector output by the XGBoost module and the time series feature vector output by the LSTM module, and completing feature fusion through a fully connected network to obtain a joint feature representation. Through "feature-level fusion", the advantages are complementary, which not only solves the time series continuity problem of monitoring data, but also adapts to the nonlinear prediction demand after the dynamic update of engineering parameters, and the accuracy is improved compared with single LSTM.

[0042] The method for training the time series model is as follows:

[0043] Data sources: including construction monitoring time series data of actual projects, each data containing a timestamp, horizontal displacement, vertical displacement and cumulative settlement;

[0044] Data preprocessing: sliding window method (such as window size = 24) is used to construct time series samples, missing values are filled by linear interpolation, and trend noise is eliminated by difference method;

[0045] Training and verification: the data are divided into training set and test set according to 8:2, and the hyperparameters are optimized, including the dimension of hidden layer, the number of iterations and the batch size.

[0046] The membrane bag sand embankment slope full-cycle safety management and control system based on machine learning is used to realize the above-mentioned membrane bag sand embankment slope full-cycle safety management and control method based on machine learning, which comprises:

[0047] A data collection module is used to collect actual engineering data, simulation data and indoor scale physical model test data for model training and verification and scheme evaluation and optimization;

[0048] A failure mode discrimination module is used to discriminate the failure mode of the design scheme or the optimized scheme through a classification model.

[0049] A safety factor prediction module in the design stage is configured to predict the limit fill height Hmax and the safety factor F of the design scheme by a regression prediction model s ;

[0050] A safety factor prediction module in the construction stage is configured to predict the safety factor F of the optimized scheme by a time series model sp ;

[0051] A foundation soil inversion module is configured to obtain the foundation soil parameters by an inversion model according to the construction monitoring data

[0052] A scheme optimization module is configured to optimize the design scheme or the optimized scheme by an optimization model when the safety factor of the scheme (the design scheme or the optimized scheme) exceeds a set value, to optimize the initial design scheme in the design stage, and to continuously optimize the scheme (the optimized scheme in the design stage and the construction stage) after optimization in the construction stage.

[0053] The present application has the following advantages over the prior art:

[0054] The present application addresses the problem that the traditional method only relies on the static safety factor in the design stage and the displacement monitoring data in the construction stage, and cannot cope with the working conditions that the embankment slope and the soft soil foundation continuously coordinate deformation, the cofferdam stiffness and the foundation strength dynamically change with drainage consolidation, resulting in distorted and delayed stability evaluation in the construction stage.

[0055] In the design stage, the present application discriminates between the overall failure and the lateral failure mode by a random forest classification model, and predicts the limit fill height H max and the safety factor F s by an XGBoost regression model to lock the risk point in advance; in the construction stage, the present application dynamically updates the core parameters such as the undrained shear strength S u of the foundation soft soil based on the monitoring data by a BO-BP inversion model, and then continuously predicts the dynamic safety factor by an XGBoost-LSTM time series model to realize the transition from static prediction to full-cycle dynamic tracking.

[0056] The application aims at the problems of traditional monitoring hysteresis, untimely early warning and lack of optimized scheme for rapid response, which leads to the problems of instability caused by insufficient foundation bearing capacity and excessive settlement in Tianjin Port, Zhoushan Shenjiamen Port and other projects, an early warning mechanism is constructed based on an XGBoost-LSTM coupled model, the safety factor is dynamically predicted, when the predicted safety level reaches a high risk level, early warning is triggered, and the scheme is adjusted and optimized through an optimization model, which can identify safety hazards in advance and respond in time, while the cost of construction can be taken into account, such as reducing the amount of geotechnical materials to reduce costs, realizing the rapid response of early warning-optimization-execution.

[0057] The application realizes deep fusion of multi-modal data by integrating data from multiple sources for prediction optimization and model training, the data sources include actual design parameters (W, J, su, etc.), physical test (scale physical model test) data, numerical simulation data and construction monitoring data, etc., which can improve the accuracy of model judgment and prediction, and is closer to the actual engineering, and can overcome the defects of insufficient pertinence of the design scheme caused by the quantization influence of core parameters such as embankment slope bottom width, membrane bag stiffness and foundation strength which are not considered in traditional design, and the problem of insufficient reliability of construction safety prediction caused by the ineffective linkage of displacement monitoring data and safety factor in the conventional construction stage, and ensures the reliability of safety control. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 A workflow diagram for the design stage of the embodiment of the application is provided.

[0059] Figure 2 A workflow diagram for the construction stage of the embodiment of the application is provided.

[0060] Figure 3 A working principle diagram of the optimization model in the embodiment of the application is provided.

[0061] Figure 4 A working principle diagram of the time sequence model in the embodiment of the application is provided.

[0062] Figure 5 A structure block diagram of the safety control system in the embodiment of the application is provided. DETAILED DESCRIPTION

[0063] The membrane bag sand embankment slope full-cycle safety control method based on machine learning comprises the following steps:

[0064] S1, as shown in the design stage: Figure 1

[0065] S1-1, based on the initial design scheme, the failure mode is judged by the classification model, the design scheme includes the membrane bag sand embankment slope design parameters and the embankment slope section geometric parameters;

[0066] ​S1-2, the corresponding regression prediction model is selected according to the failure mode obtained in the step S1-1, and the safety factor F is predicted by the corresponding regression prediction model max and the safety factor F s ;

[0067] S1-3, by comparing the predicted safety factor F s with the safety factor threshold [F s ], it is judged whether the design scheme is feasible, if feasible (F s ≥ [F s ], the design scheme is the optimal scheme, if not feasible (F s <[F s ], the optimal scheme is generated by the optimization model, and the optimal scheme includes the membrane bag sand embankment slope design parameter and the embankment slope section geometric parameter;

[0068] S2, as shown in Figure 2 , in the construction stage:

[0069] S2-1, the new foundation soil parameters are obtained by the inversion model through the construction monitoring data, the original foundation soil parameters are replaced by the new foundation soil parameters, and the optimal scheme obtained in the step S1 is updated;

[0070] S2-2, the new optimal scheme and the new construction monitoring data are judged by the classification model to determine the failure mode;

[0071] S2-3, the corresponding time sequence model is selected according to the failure mode, and the safety factor F is predicted by the corresponding time sequence model sp ;

[0072] S2-4, by comparing the predicted safety factor F sp with the safety factor threshold [F s ], the safety risk level is judged, if the safety risk level is within the set safety range, no warning is needed and continuous monitoring is needed, if the safety risk level exceeds the set safety range (reaches the category of high risk), warning is needed and the optimal scheme is generated by the optimization model, and the optimal scheme includes the membrane bag sand embankment slope design parameter and the embankment slope section geometric parameter;

[0073] S2-5, in the construction process, the safety evaluation and scheme optimization are continuously carried out in the construction process according to the method of steps S2-2 to S2-4, until the construction is completed. Timely means that risk assessment and scheme optimization are carried out according to the set time interval, and at random means that sampling is carried out randomly or evaluation and optimization are carried out when data is abnormal.

[0074] In the embodiment, the membrane bag sand embankment slope design parameter includes the tensile stiffness J of the geomembrane bag, the undrained shear strength S u, the shear strength of the contact surface between the geomembrane bag and the foundation τ, the filling thickness of the bag sand t, the internal friction angle of the filling sand φ, and the specific gravity of the foundation soil γ c , and the specific gravity of the filling sand of the bag γ s ; the embankment slope section geometric parameters include the embankment slope width W, the embankment slope height H, and the embankment slope gradient k; the foundation soil parameters include the undrained shear strength of soft soil S u ; the construction monitoring data include displacement data such as horizontal displacement, vertical displacement, and settlement value.

[0075] The core function of the classification model is to accurately distinguish the failure mode according to the input parameters, and to provide the basis for subsequent regression prediction. It needs to adapt to the high dimensionality and nonlinearity of engineering data, and has strong anti-interference ability. As one of the implementation manners, the classification model adopts a random forest classifier, which takes into account the noise tolerance and classification accuracy of engineering data. The output failure modes include overall failure and side failure. The overall failure is mainly caused by insufficient foundation bearing capacity, mainly manifested as overall subsidence of the embankment slope, and the failure surface in the foundation is basically symmetrical and intersects below the center of the embankment slope; the side failure is mainly manifested as local failure of the foundation at the toe position of the embankment slope on both sides.

[0076] For example, the training method of the classification model is specifically as follows:

[0077] (1) Collect data, the data sources include actual engineering data, simulation data, and indoor scale physical model test data, the indoor scale physical model test data includes data under remolded silt conditions.

[0078] The actual engineering data of this embodiment has 12 groups of instability case data, containing failure mode labels; the simulation data is obtained by numerical simulation and simulation calculation (finite element method), containing 2400 groups of conditions, containing parameters such as embankment bottom width W, geomembrane tensile stiffness J, geomembrane sand filling thickness t, undrained shear strength of soft soil S u , internal friction angle of filling sand φ, embankment slope gradient k, and shear strength τ of the contact surface between the geomembrane bag and the foundation; the indoor scale physical model test data includes 30 groups of data under remolded silt conditions.

[0079] These data also need to be processed to select core features. Specifically, Z-score standardization can be used to eliminate the dimension effect, and redundant features can be removed by mutual information method.

[0080] (2) Randomly divide the data into a training set and a test set according to a ratio of 7:3.

[0081] (3) Optimize the hyperparameters by 5-fold cross-validation, the number of decision trees is 100, and the maximum depth of each tree is 8. Tests prove that the obtained model can accurately distinguish the two failure modes.

[0082] The role of the regression prediction model is mainly to predict the limit filling height H max and the safety factor F s , which meets the engineering level precision requirement. It should be noted here that although only the safety factor F s is used in subsequent operations, H max is not used, from the perspective of engineering personnel, they mainly focus on the overall stability of the embankment slope and the safety factor F s . But knowing the safety factor is based on the known embankment limit filling height. Therefore, predicting the limit filling height H max is a prerequisite for predicting F s , and it is also important for engineering personnel to understand the limit filling height of the embankment slope to determine whether it meets the engineering requirements.

[0083] Different failure modes correspond to different regression prediction models. When the classification model determines whether the scheme is overall failure or lateral failure, the corresponding regression prediction model is used to predict the predicted limit filling height H max and the safety factor F s according to the failure type. Different regression prediction models have the same model structure, but the types of training data are different, and the corresponding parameters may be different. As one of the implementation manners, in this embodiment, the regression prediction model adopts the XGBoost regression model structure, which can effectively fit the complex relationship between the parameters and the safety indicators, adapt the nonlinear mapping relationship between the parameters and the target values, and support feature importance analysis.

[0084] The method for training the regression prediction model is:

[0085] (1) The data sources of the training set include simulation data and indoor scale physical model test data, and the indoor scale physical model test data includes limit bearing capacity and safety factor data derived from the embankment load-settlement curve. In this embodiment, the simulation data is calculated by numerical simulation, including but not limited to finite element method, discrete element method, etc. The data specifically includes the corresponding relationship data of W, J, S u and other parameters and H max , F s , and the limit bearing capacity and safety factor data derived from the embankment load-settlement curve.

[0086] (2) The test set data includes actual engineering data.

[0087] In this embodiment, the actual engineering case data includes the actual data of Tianjin Port and Zhoushan Shenjiamen Port cofferdam. The data of the membrane bag sand cofferdam of the submarine immersed tunnel of Zhoushan Shenjiamen Port is as follows: the width of the embankment slope bottom W = 23.3 m, the height H = 7 m, the slope k = 1 / 2, the stiffness of the geomembrane bag J = 150 kN / m, the filling sand specific gravity is 20 kN / m 3 , the soil strength S u = 25 kPa, and the soil buoyant specific gravity is 9 kN / m 3 .

[0088] The data of the Tianjin Port reclamation membrane bag sand cofferdam is as follows: the embankment slope width W = 60 m, the height H = 6 m, the slope k = 1 / 2, the stiffness of the geomembrane bag J = 140 kN / m, the filling sand specific gravity is 18 kN / m 3 , the soil strength S u = 5 kPa, and the soil buoyant specific gravity is 6.8 kN / m 3 .

[0089] Here, the parameters used for the verification case are less than the 7 input parameters of the numerical simulation and simulation calculation described above, because according to the results of the simulation calculation, some parameters have less effect on the limit filling height and stability of the embankment slope and can be ignored.

[0090] (3) The grid search is used to optimize the hyperparameters, the learning rate is 0.1, the tree depth is 6, and the iteration number is 200. The verification proves that the regression prediction model can realize that the predicted values of H max and F s meet the engineering design and construction control requirements.

[0091] The function of the optimization model is to optimize the existing scheme, meet Fs≥[Fs] in the design stage, be lower than the high-risk early warning in the construction stage, generate an optimal parameter scheme, and need to adapt to multiple constraint conditions such as material consumption, construction difficulty, and safety threshold.

[0092] As one of the implementation manners, in this embodiment, the optimization model is a composite model which includes the NSGA-III optimization algorithm, the classification model and the regression prediction model, as shown in FIG. 8, and the working process is as follows: Figure 3

[0093] a. Under the condition of clear optimization target and constraint condition, the candidate scheme is generated by the NSGA-III optimization algorithm.

[0094] For example, the constraint condition corresponding to the design stage is W∈30-120m, J∈140-8000kN / m, S u ≥ 5 kPa, and the constraint condition corresponding to the construction stage is that the adjustment range is less than 20% of the original parameters to avoid large changes in the construction scheme.

[0095] ​The NSGA-Ⅲ optimization algorithm can simultaneously optimize three major objectives: maximizing the safety factor, minimizing the amount of geotechnical materials used, and minimizing the construction period.

[0096] b. The candidate solution determines the failure mode using the classification model, selects the corresponding regression prediction model based on the failure mode, and obtains the safety factor F through the corresponding regression prediction model. s .

[0097] c. Select feasible solutions based on the safety coefficient Fs of each candidate solution.

[0098] The specific method is as follows:

[0099] During the design phase, the safety factor F of the candidate solutions is... s With the safety factor set threshold [F] s If Fs ≥ [Fs], then the candidate solution is a feasible solution; otherwise, it is an infeasible solution.

[0100] During the construction phase, the safety factor F of the candidate scheme will be... sp With the safety factor set threshold [F] s In comparison, the safety risk level is determined. If the safety risk level is within the set safety range, it is a feasible solution; otherwise, it is an infeasible solution.

[0101] d. Taking into account economic and safety factors, select the optimal solution from the feasible options, such as the most economical and safest option.

[0102] In this embodiment, the training method of the NSGA-Ⅲ optimization algorithm is as follows: using 2000 feasible schemes generated by simulation numerical modeling as the initial population, setting the crossover probability to 0.8, the mutation probability to 0.1, and the number of iterations to 50. Experiments show that the F-value of the optimized scheme is... s The compliance rate has been effectively improved, and the amount of geosynthetic materials used has been effectively reduced.

[0103] The core function of the inversion model is to invert key foundation parameters such as the undrained shear strength of soft soil based on construction monitoring displacement data, meeting the input requirements for dynamic prediction during construction. As one implementation method, in this embodiment, the inversion model employs a Bayesian optimized BP neural network (BO-BP), balancing inversion accuracy and computational efficiency. The input data is the monitored displacement data, and the output data is the undrained shear strength S of the soft soil. u .

[0104] For example, the method for training the inversion model is as follows:

[0105] (1) The data sources for the training set include actual engineering data and simulation data.

[0106] In this embodiment, the actual engineering data includes 500 groups of construction measured displacement and soil test data, and the simulation simulation data is 1000 groups of data obtained by simulating the relationship between the "input parameters-displacement response", S u The value is 2.5-15kPa, corresponding to different displacement outputs.

[0107] More specifically, the simulation method can be: using numerical simulation software, including but not limited to finite element method, discrete element method, for example, using commercial finite element software abaqus, according to the geometric size, material parameters of the membrane bag sand embankment slope and foundation, establish a finite element model, simulate the load applied in the actual construction process until the embankment slope instability and failure, so as to obtain the limit filling height of the embankment slope; Then under the condition of equal geometric size and material parameters, by setting different embankment filling height, the safety factor corresponding to different filling height is calculated by using strength reduction method; In this way, the corresponding relationship between the embankment geometric size and material parameters and the limit filling height H max , filling height H, safety factor F s .

[0108] (2) The optimal initial weight and structure of the BP neural network are searched through Bayesian optimization, including the number of hidden layers and the number of neurons in each layer. Finally, the number of hidden layers is 2, the number of nodes is 32, the inversion parameter accuracy is high, and the foundation parameters can be dynamically updated.

[0109] The main function of the time series model is to dynamically predict the safety factor by combining the updated parameters and the monitoring data, and to realize risk early warning. Different destruction modes correspond to different time series models. When the classification model judges whether the scheme is overall destruction or side destruction, the safety factor is predicted by the corresponding time series model according to the destruction type. Different regression time series models have the same model structure, but the types of training data are different, and the corresponding parameters may be different.

[0110] As one of the implementation modes, in this embodiment, the time series model adopts an XGBoost-LSTM coupled model, as shown in Figure 4 , which includes an XGBoost module, an LSTM module and a feature fusion module. The XGBoost module is used for static nonlinear feature extraction of static data in the optimal scheme, the LSTM module is used for time series feature extraction of construction monitoring data, and the feature fusion module is used for splicing the static nonlinear feature vector output by the XGBoost module and the time series feature vector output by the LSTM module. And through one to two layers of fully connected neural network to complete feature fusion, get joint feature representation, used for subsequent safety factor prediction.

[0111] The related formula of the feature fusion module can be:

[0112] ,

[0113] ,

[0114] .

[0115] The application adopts an XGBoost-LSTM coupled model to realize advantage complementation through "feature-level fusion", solve the time sequence continuity problem of monitoring data, and adapt to the nonlinear prediction demand after the dynamic update of engineering parameters, and the accuracy is improved compared with single LSTM.

[0116] The static data of the XGBoost module is derived from related data in the optimization scheme, such as W, H, J, γ c ′, γ s , etc.

[0117] The training of the time sequence model is to independently train the XGBoost module, so as to fully fit the nonlinear relationship between the static design parameters and the safety factor, and after the training is completed, it is used as a static feature extractor; the second stage is to freeze the XGBoost module parameters, utilize the static nonlinear features output by the XGBoost module and the time sequence features extracted by the LSTM during construction monitoring, and jointly model in the feature-level fusion module, train the LSTM module and the feature fusion module through back propagation end to end, and realize the prediction of the dynamic safety factor. Specifically, the training method comprises:

[0118] (1) Data source: including construction monitoring time sequence data of actual projects, each data containing time stamp, horizontal displacement, vertical displacement and cumulative settlement. In this embodiment, the construction monitoring time sequence data involves three actual projects, the sampling frequency is 30 min / time, covers the whole stage of embankment slope filling, and more than 100,000 time sequence data.

[0119] (2) Data preprocessing: sliding window method (window size = 24) is adopted to construct time sequence samples, missing values are filled by linear interpolation, and difference method is adopted to eliminate trend noise.

[0120] (3) Training and verification: the data is divided into training set and test set according to 8:2, the hyperparameters are optimized, including hidden layer dimension, iteration number and batch size. Finally, the time sequence model is obtained, the hidden layer dimension is 64, the iteration number is 100, and the batch size is 32, which can effectively predict F s , the safety risk early warning can be given in advance, and the warning value can be flexibly adjusted according to the actual engineering demand.

[0121] The membrane bag sand embankment slope whole cycle safety management and control system based on machine learning is used to realize the above-mentioned membrane bag sand embankment slope whole cycle safety management and control method based on machine learning, such as Figure 5As shown, it includes a data collection module, a failure mode discrimination module, a design stage safety factor prediction module, a construction stage safety factor prediction module, a foundation soil inversion module, and a scheme optimization module. These modules can cover the safety control of the membrane bag sand embankment slope from the design stage to the construction stage, and realize closed-loop management from scheme design to construction control.

[0122] The data collection module is used to collect actual engineering data, simulation data, and indoor scale physical model test data for model training and verification. Meanwhile, the data collection module is also used to collect data of the design scheme to be predicted and controlled for scheme evaluation and optimization.

[0123] The failure mode discrimination module is used to discriminate the failure mode of the design scheme or the optimized scheme through a classification model. In the design stage, the initial design scheme is discriminated for failure mode, and in the construction stage, the optimized scheme is discriminated for failure mode. The failure modes include overall failure and side failure, etc. In some embodiments, the classification model adopts a random forest classifier.

[0124] The design stage safety factor prediction module is used to predict the limit filling height H max and the safety factor F s of the design scheme through a regression prediction model. Different failure modes are input into different regression prediction models. In some embodiments, the regression prediction model adopts an XGBoost regression model structure.

[0125] The construction stage safety factor prediction module is used to predict the safety factor F sp of the optimized scheme through a time series model. Different failure modes are input into different time series models. In some embodiments, the time series model adopts an XGBoost-LSTM coupled model structure.

[0126] The foundation soil inversion module is used to obtain the foundation soil parameters through an inversion model according to the construction monitoring data. In some embodiments, the inversion model adopts a BP neural network optimized by Bayes (BO-BP) structure.

[0127] The scheme optimization module is used to optimize the design scheme or the optimized scheme through an optimization model when the safety factor of the scheme (design scheme or optimized scheme) exceeds the set value. In the design stage, the initial design scheme is optimized, and in the construction settlement, the optimized scheme (including the optimized scheme in the design stage and the construction stage) is continuously optimized. In some embodiments, the optimization model is a composite model, which includes an NSGA-III optimization algorithm, the classification model, and the regression prediction model.

[0128] The above detailed description is for the specific implementation of the present application, which is not intended to limit the patent scope of the present application. Any equivalent implementation or modification without departing from the present application shall be included in the patent scope of the present application.

Claims

1. A machine learning-based method for full-cycle safety management of membrane bag sand embankment slopes, characterized in that, Includes the following steps: S1. During the design phase: S1-1. Based on the initial design scheme, the failure mode is determined by a classification model. The design scheme includes the design parameters of the membrane bag sand embankment and the geometric parameters of the embankment cross section. The classification model uses a random forest classifier, and its output destruction modes include overall destruction and partial destruction. S1-2. Based on the failure mode obtained in step S1-1, select the corresponding regression prediction model, and use the corresponding regression prediction model to predict the limit filling height H. max and safety factor F s The regression prediction model adopts the XGBoost regression model structure. S1-3, By predicting the safety factor F s With the safety factor set threshold [F] s In comparison, it is determined whether the design scheme is feasible. If it is feasible, the design scheme is taken as the optimal scheme. If it is not feasible, the optimal scheme is generated through the optimization model. The optimal scheme includes the design parameters of the membrane bag sand embankment and the geometric parameters of the embankment section. The optimization model includes the NSGA-Ⅲ optimization algorithm, the classification model and the regression prediction model. S2. During the construction phase: S2-1. Construction monitoring data is used to obtain new foundation soil parameters through an inversion model. These new parameters replace the original foundation soil parameters, thus updating the optimal solution obtained in step S1. The inversion model employs a Bayesian-optimized BP neural network. The input data is the monitored displacement data, and the output data is the undrained shear strength S of the soft soil. u ; S2-2. The new optimal solution and new construction monitoring data are used to determine the failure mode through a classification model; S2-3. Select the corresponding time series model based on the failure mode, and predict the safety factor F using the corresponding time series model. sp The time series model adopts an XGBoost-LSTM coupled model, which includes an XGBoost module, an LSTM module, and a feature fusion module. S2-4, Based on the predicted safety factor F sp With the safety factor set threshold [F] s In comparison, the safety risk level is determined. If the safety risk level is within the set safety range, no warning is needed and continuous monitoring is performed. If the safety risk level exceeds the set safety range, an warning is issued and an optimal solution is generated through an optimization model. The optimal solution includes the design parameters of the membrane bag sand embankment and the geometric parameters of the embankment section. S2-5. During the construction process, following the methods in steps S2-2 to S2-4, conduct regular or irregular safety assessments and optimize the plan until the construction is completed.

2. The method for full-cycle safety management of membrane bag sand embankment slopes based on machine learning according to claim 1, characterized in that, The design parameters for the geomembrane bag sand embankment slope include the tensile stiffness J of the geomembrane bag and the undrained shear strength S of the soft soil. u The shear strength τ of the contact surface between the geomembrane bag and the foundation, the thickness t of the sand filling in the geomembrane bag, the internal friction angle φ of the filling sand, and the unit weight γ of the foundation soil. c 'and membrane bag filling sand unit weight γ s The geometric parameters of the embankment slope section include the embankment slope width W, the embankment slope height H, and the embankment slope k; the parameters of the foundation soil include the undrained shear strength S of soft soil. u The construction monitoring data includes displacement data.

3. The method for full-cycle safety management of membrane bag sand embankment slopes based on machine learning according to claim 1, characterized in that, The training method for the classification model is as follows: Data is collected, and the data sources include actual engineering data, simulation data, and indoor scaled physical model test data, including data under the condition of remolded sludge. The data was randomly divided into a training set and a test set in a 7:3 ratio. The hyperparameters were optimized using 5-fold cross-validation, with 100 decision trees and a maximum depth of 8 for each tree.

4. The method for full-cycle safety management of membrane bag sand embankment slopes based on machine learning according to claim 1, characterized in that, The training method for the regression prediction model is as follows: The training set data sources include simulation data and indoor scaled physical model test data, which includes ultimate bearing capacity and safety factor data derived from the embankment load-settlement curve. The test set data includes actual engineering data; The hyperparameters were optimized using a grid search with a learning rate of 0.1, a tree depth of 6, and 200 iterations.

5. The method for full-cycle safety management of membrane bag sand embankment slopes based on machine learning according to claim 2, characterized in that, The working process of the optimization model is as follows: Given a clear optimization objective and constraints, candidate solutions are generated using the NSGA-Ⅲ optimization algorithm. Candidate solutions determine the failure mode using the classification model, select the corresponding regression prediction model based on the failure mode, and obtain the safety factor F through the corresponding regression prediction model. s ; Based on the safety coefficient Fs of each candidate scheme, feasible schemes are selected; Taking into account both economic and security factors, the optimal solution is selected from the feasible options.

6. The method for full-cycle safety management of membrane bag sand embankment slopes based on machine learning according to claim 5, characterized in that, The safety factor F of each candidate scheme s The method for selecting feasible solutions is as follows: If it is in the design phase, the safety factor F of the candidate scheme will be... s With the safety factor set threshold [F] s In comparison, if Fs≥[Fs], then the candidate solution is a feasible solution; otherwise, it is an infeasible solution. If it is during the construction phase, the safety factor F of the candidate scheme will be... sp With the safety factor set threshold [F] s In comparison, the safety risk level is determined. If the safety risk level is within the set safety range, it is a feasible solution; otherwise, it is an infeasible solution. The constraints for the design phase are W∈30-120m, J∈140-8000kN / m, S u For parameters ≥5kPa, the corresponding constraint during the construction phase is that the adjustment range is ≤20% of the original parameters.

7. The method for full-cycle safety management of membrane bag sand embankment slopes based on machine learning according to claim 5, characterized in that, The training method of the NSGA-Ⅲ optimization algorithm is as follows: 2000 feasible schemes generated by simulation numerical simulation are used as the initial population, and the crossover probability is set to 0.8, the mutation probability is set to 0.1, and the number of iterations is set to 50.

8. The method for full-cycle safety management of membrane bag sand embankment slopes based on machine learning according to claim 1, characterized in that, The method for training the inversion model is as follows: The training set data sources include actual engineering data and simulation data, wherein the simulation data includes S u Simulated input parameter-displacement response data with values ​​ranging from 2.5 to 15 kPa; The optimal initial weights and structure of the BP neural network are searched using Bayesian optimization, including the number of hidden layers and the number of neurons per layer.

9. The method for full-cycle safety management of membrane bag sand embankment slopes based on machine learning according to claim 1, characterized in that, The XGBoost module is used to extract static nonlinear features from the static data in the optimal solution, the LSTM module is used to extract time-series features from the construction monitoring data, and the feature fusion module is used to concatenate the static nonlinear feature vector output by the XGBoost module and the time-series feature vector output by the LSTM module, and complete the feature fusion through a fully connected network to obtain a joint feature representation. The training method for the time series model is as follows: Data sources include construction monitoring time-series data from actual projects, with each data point containing a timestamp, horizontal displacement, vertical displacement, and cumulative settlement. Data preprocessing: The sliding window method was used to construct time series samples, missing values ​​were filled by linear interpolation, and trend noise was eliminated by the difference method; Training and validation: The data is divided into training and testing sets in an 8:2 ratio, and hyperparameters are optimized, including hidden layer dimension, number of iterations, and batch size.

10. A machine learning-based full-cycle safety management system for membrane bag sand embankment slopes, used to implement the machine learning-based full-cycle safety management method for membrane bag sand embankment slopes as described in any one of claims 1 to 9, characterized in that, include: The data collection module is used to collect actual engineering data, simulation data, and indoor scaled physical model test data for model training and verification, as well as scheme evaluation and optimization. The failure mode identification module is used to identify the failure mode of the design or optimization scheme through a classification model; The safety factor prediction module in the design phase is used to predict the ultimate fill height H of the design scheme through a regression prediction model. max With safety factor F s The construction phase safety factor prediction module is used to predict the safety factor F of the optimized scheme through a time series model. sp ; The foundation soil inversion module is used to obtain foundation soil parameters from construction monitoring data through an inversion model. The scheme optimization module is used to optimize the design scheme or optimization scheme through the optimization model when the safety factor of the scheme exceeds the set limit. It optimizes the initial design scheme in the design stage and continuously optimizes the optimized scheme in the construction stage.

Images