Water-rich tunnel stress mechanism analysis method based on physical-data dual drive

Abstract

The invention discloses a water-rich tunnel stress mechanism analysis method based on physical-data dual drive. A data driving model is introduced, and the model can capture various complex phenomena and change trends possibly occurring in the running process of the water-rich tunnel through analysis of actual observation data, so that the stress condition is predicted more accurately. By adopting the dual-drive method, the defects of a single model are overcome, stress analysis under complex working conditions can be coped, and a new method is provided for tunnel design and construction. According to the method, the theoretical basis of a physical model and the data capturing capacity of a data driving model are combined, and a more comprehensive and accurate analysis method is formed.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel stress analysis methods under water-rich conditions, and specifically relates to a water-rich tunnel stress mechanism analysis method based on physics-data dual drive. Background Art

[0002] Water-rich tunnels usually pass through geological bodies with abundant groundwater. The hydrogeological conditions of these geological bodies are complex, which have a significant impact on the stability and stress state of the tunnel structure.

[0003] There are currently two methods for analyzing the stress state of water-rich tunnels: 1. A physical model method based on the basic principles of solid mechanics, fluid mechanics, and rock mechanics can be used to further derive mathematical equations to describe the stress state of the tunnel under different working conditions. Although this model method can provide theoretical support for tunnel behavior, in practical applications, this method needs to make many assumptions in order to simplify calculations. These assumptions may not be consistent with the actual situation and cannot fully cover complex physical processes, thus affecting the accuracy of the analysis results. For example, the physical model may ignore the nonlinear characteristics of water-soil interaction, making the results inaccurate; 2. A data-driven neural network model method that uses machine learning technology to analyze the patterns and laws of tunnel stress behavior based on the monitored and collected data related to tunnel construction and operation (such as stress, strain, etc.), but the prediction effect of this method is limited by data quality and data volume, and the data-driven method is usually a "black box" model with an opaque internal mechanism and prediction process, making it difficult to explain and understand the working principle of the model and its prediction results. Summary of the invention

[0004] The present invention provides a method for analyzing the stress mechanism of a water-rich tunnel based on physical-data dual drive, and the specific technical scheme is as follows:

[0005] A method for analyzing the stress mechanism of a water-rich tunnel based on physical and data dual drive, the method mainly includes the following steps:

[0006] Step 1: Comparative analysis of data-driven neural network models

[0007] They are data collection, data aggregation, data preprocessing, data splitting, model training, model testing, and model evaluation;

[0008] During data collection, mechanical sensor data in the water-rich tunnel is collected. ω , consult the design drawings and exploration data to obtain tunnel structural data r, r0, r1, r2, r3, R and geological data k1, k2, k3, k4, P i ; where P ωis the water pressure, in MPa; r is the radial seepage radius of groundwater, in m; r0 is the inner radius of the secondary lining of the tunnel, in m; r1 is the inner radius of the initial support, in m; r2 is the inner radius of the reinforcement circle, in m; r3 is the outer radius of the reinforcement circle, in m; R is the outer radius of the study area, in m; k1 is the permeability coefficient of the secondary lining, in m / d; k2 is the permeability coefficient of the initial support, in m / d; k3 is the permeability coefficient of the reinforcement circle, in m / d; k4 is the permeability coefficient of the surrounding rock, in m / d; P i is the water pressure of the original seepage field, in MPa;

[0009] Step 2: Physical Modeling

[0010] The physical model is based on the following basic assumptions:

[0011] 1) In high head areas, the circular tunnel is surrounded by groundwater;

[0012] 2) The surrounding rock is a homogeneous, isotropic, continuous medium;

[0013] 3) Seepage potential is caused only by changes in water pressure;

[0014] 4) The water load is axisymmetrically distributed; the physical model is further derived into the tunnel force calculation formula, which is expressed as:

[0015]

[0016] In the formula, a and b are calculation parameters, which can be expressed as:

[0017]

[0018] According to formulas (1), (2), and (3), the water pressure behind the secondary lining, initial support, and reinforcement ring of the tunnel can be calculated under basic assumptions. However, this physical model cannot accurately simulate the complex geological conditions and physical relationships in actual situations. Therefore, it is necessary to add a deep neural network to learn monitoring data and further improve the model.

[0019] Step 3: Neural network model establishment

[0020] The input layer of the deep learning neural network model is the 11 parameters r, r0, r1, r2, r3, R, k1, k2, k3, k4, P in step 1. i There are two hidden layers, each consisting of 16 neurons. Each neuron performs a weighted summation of the output of the previous layer and performs a nonlinear transformation through the ReLU activation function to generate the output of this layer. The connection between the hidden layers forms the depth of the neural network. The last layer generates the network output, that is, the water pressure P at the radial seepage radius r of the groundwater. ω; ReLU stands for Rectified Linear Unit.

[0021] Step 4: Establishing a physical-data dual-driven model

[0022] The model is based on previously successful deep learning models and physical models, and innovatively integrates physical knowledge with machine learning methods. The core of this innovation lies in the construction of a custom loss function, which not only includes the mean square error of the data, but also creatively introduces physical formulas as part of the loss function. Specifically, the equation describing the relationship between the water pressure of the tunnel structure and the geological structure and tunnel structure is first derived, and then the physical equation is applied to evaluate the consistency between the model prediction value and the actual physical law. The specific loss function is as follows:

[0023]

[0024] Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample; y i1 is the calculated value of the i-th sample; λ is the weight of the data-driven loss value, which is adjusted according to different working conditions.

[0025] The preferred embodiment of the water-rich tunnel stress mechanism analysis method based on physical-data dual drive is that the analysis method is established based on a DNN deep neural network model, where DNN stands for Deep Neural Networks.

[0026] The model is used to explore the complex relationship between features and targets; and the model performance is represented by the model's loss function; the loss function is expressed as:

[0027]

[0028] Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample.

[0029] The method for analyzing the stress mechanism of a water-rich tunnel based on dual physics and data drive is preferably a method in which the mean square error between the predicted value and the true value is used to further reduce the impact of noise and disturbance on the model and improve the interpretability of the model, and the traditional physical model is incorporated into it to enhance the robustness and interpretability of the model.

[0030] The preferred solution of the stress mechanism analysis method of a water-rich tunnel based on physical-data dual drive is that a physical loss term is introduced into the loss function of the model to achieve this goal; the loss function after introduction is expressed as:

[0031]

[0032] Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample; y i1 is the calculated value of the i-th sample; λ is the weight of the data-driven loss value, which is adjusted according to different working conditions;

[0033] The preferred solution of the water-rich tunnel stress mechanism analysis method based on physics-data dual drive is that in step 3, the neural network training process includes the following four parts:

[0034] 1) Forward propagation: Starting from the input layer, the output of each layer of neurons is calculated in sequence until the result of the output layer is obtained;

[0035] 2) Back propagation: Use the chain rule to calculate the gradient of the loss function with respect to each weight and bias term. These gradients indicate how to adjust the weights to reduce the loss.

[0036] 3) Optimization algorithm: Adopt the Adam optimization algorithm and use the gradient calculated by back-propagation to update the weights and biases, and gradually iterate and optimize the model; the full name of Adam is AdaptiveMomentEstimation;

[0037] 4) Loss function: The mean square error (MSE) is used to measure the gap between the model prediction result and the true label. The full name of MSE is Mean Squared Error.

[0038] The preferred scheme of the water-rich tunnel stress mechanism analysis method based on physical-data dual drive is that the theoretical process of data-driven neural network model comparison analysis is as follows:

[0039] (1) Data collection: Obtain model training samples based on monitoring data, geological structures, physical formulas, and tunnel structures;

[0040] (2) Data division: Filter the collected data to form a data set. Set the test data volume to 20, 40, 60, 80, 100, 200, 400, and 600 respectively, and conduct multiple groups of tests to analyze the generalization performance, fitting, and model stability of the model;

[0041] (3) Data preprocessing: The sample data is preprocessed using the StandardScaler method, which scales the features to conform to the standard normal distribution with a mean of 0 and a standard deviation of 1, so that different features have the same scale during the training process, thereby improving the training effect and stability of the model;

[0042] (4) Experimental grouping: The other methods were set as the control group, and the water-rich tunnel stress mechanism analysis method based on physical-data dual drive was set as the experimental group for control experiments.

[0043] (5) Experimental method: 80% of the samples are used as the training set of the model, and the remaining 20% ​​are used as the test set to verify the performance of the model, and the prediction results, loss value and accuracy are output;

[0044] (6) Result analysis: By drawing a line graph of data volume, model loss, and accuracy, the performance of the two models is intuitively displayed, and the curve fluctuations are analyzed to further evaluate the model stability; the curve trend is analyzed to evaluate the model learning ability, and different data volumes are tested to verify the advantages of this method over deep neural networks in fitting training data and test data.

[0045] Beneficial Effects

[0046] The present invention outputs prediction results based on data by adding physical constraints-mechanical equations to the neural network model. The results are more in line with physical laws and improve the interpretability of the model. At the same time, the neural network can automatically capture nonlinear relationships and complex patterns in the data, further overcoming the problem that traditional physical models are limited by model assumptions and parameter uncertainties.

[0047] During the training process, the model aims to achieve dual optimization: 1) on the one hand, the mean square error between the predicted value and the target observed value is minimized to ensure high-precision fitting at the data level; 2) on the other hand, the physical residual is innovatively introduced as a key component of the loss function. This residual quantifies the deviation between the result obtained after the model prediction value is substituted into the physical equation and the true solution of the equation, which not only deepens the model's learning depth of the data, but more importantly, it prompts the model to actively follow the physical laws during the optimization process, realizing the deep integration of data-driven and physical constraints. This dual-oriented training mechanism provides the model with a self-correction ability, so that while pursuing the best fit of the data, it continuously adjusts the predicted output to be more in line with the physical laws, thereby comprehensively and deeply learning the essential characteristics of the physical system, significantly improving the accuracy and reliability of the model in predicting the stress mechanism of the water-rich tunnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the groundwater seepage network diagram;

[0049] Figure 2 is the water pressure distribution diagram;

[0050] Figure 3 Schematic diagram of neural network;

[0051] Figure 4This is a flow chart of the analysis method of the stress mechanism of water-rich tunnels based on physical-data dual drive;

[0052] Figure 5 This is a structural diagram of the intelligent analysis method for the stress mechanism of water-rich tunnels based on physical-data dual drive. DETAILED DESCRIPTION

[0053] In combination with the appended Figure 1-5 The invention is further described, but it should not be understood that the above subject scope of the present invention is limited to the following embodiments. Without departing from the above idea of ​​the present invention, various substitutions and changes are made according to common technical knowledge and customary means in the field, which should all be included in the protection scope of the present invention.

[0054] A method for analyzing the stress mechanism of a water-rich tunnel based on physical and data dual drive, the method mainly includes the following steps:

[0055] Step 1: Analytical Method Theory

[0056] They are data collection, data aggregation, data preprocessing, data splitting, model training, model testing, and model evaluation;

[0057] During data collection, mechanical sensor data in the water-rich tunnel is collected. ω , consult the design drawings and exploration data to obtain tunnel structural data r, r0, r1, r2, r3, R and geological data k1, k2, k3, k4, P i ; where P ω is the water pressure, in MPa; r is the radial seepage radius of groundwater, in m; r0 is the inner radius of the secondary lining of the tunnel, in m; r1 is the inner radius of the initial support, in m; r2 is the inner radius of the reinforcement circle, in m; r3 is the outer radius of the reinforcement circle, in m; R is the outer radius of the study area, in m; k1 is the permeability coefficient of the secondary lining, in m / d; k2 is the permeability coefficient of the initial support, in m / d; k3 is the permeability coefficient of the reinforcement circle, in m / d; k4 is the permeability coefficient of the surrounding rock, in m / d; P i is the water pressure of the original seepage field, in MPa;

[0058] Step 2: Physical Modeling

[0059] The physical model is based on the following basic assumptions:

[0060] 1) In high head areas, the circular tunnel is surrounded by groundwater;

[0061] 2) The surrounding rock is a homogeneous, isotropic, continuous medium;

[0062] 3) Seepage potential is caused only by changes in water pressure;

[0063] 4) The water load is axisymmetrically distributed;

[0064] Groundwater seepage network Figure 1 As shown; water pressure distribution is as Figure 2 As shown;

[0065] The physical model is further derived into a tunnel force calculation formula, expressed as:

[0066]

[0067] In the formula, a and b are calculation parameters, which can be expressed as:

[0068]

[0069] According to formulas (1), (2), and (3), the water pressure behind the secondary lining, initial support, and reinforcement ring of the tunnel can be calculated under basic assumptions. However, this physical model cannot accurately simulate the complex geological conditions and physical relationships in actual situations. Therefore, it is necessary to add a deep neural network to learn monitoring data and further improve the model.

[0070] Step 3: Neural network model establishment

[0071] The input layer of the deep learning neural network model is the 11 parameters r, r0, r1, r2, r3, R, k1, k2, k3, k4, P in step 1. i There are two hidden layers, each layer consists of 16 neurons. Each neuron performs a weighted summation on the output of the previous layer and performs a nonlinear transformation through the ReLU activation function to generate the output of this layer. The connections between the hidden layers form the depth of the neural network. The neural network diagram is shown in Figure 3 As shown; the last layer generates the network output, namely the water pressure P at the radial seepage radius r of the groundwater ω ; ReLU stands for Rectified Linear Unit.

[0072] The neural network training process consists of the following four parts:

[0073] 1) Forward propagation: Starting from the input layer, the output of each layer of neurons is calculated in sequence until the result of the output layer is obtained;

[0074] 2) Back propagation: Use the chain rule to calculate the gradient of the loss function with respect to each weight and bias term. These gradients indicate how to adjust the weights to reduce the loss.

[0075] 3) Optimization algorithm: Adopt the Adam optimization algorithm and use the gradient calculated by back-propagation to update the weights and biases, and gradually iterate and optimize the model; the full name of Adam is AdaptiveMomentEstimation;

[0076] 4) Loss function: The mean square error (MSE) is used to measure the gap between the model prediction result and the true label. The full name of MSE is Mean Squared Error.

[0077] Step 4: Establishing a physical-data dual-driven model

[0078] The model is based on previously successful deep learning models and physical models, and innovatively integrates physical knowledge with machine learning methods. The core of this innovation lies in the construction of a custom loss function, which not only includes the mean square error of the data, but also creatively introduces physical formulas as part of the loss function. Specifically, the equation describing the relationship between the water pressure of the tunnel structure and the geological structure and tunnel structure is first derived, and then the physical equation is applied to evaluate the consistency between the model prediction value and the actual physical law. The specific loss function is as follows:

[0079]

[0080] Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample; y i1 is the calculated value of the i-th sample; λ is the weight of the data-driven loss value, which is adjusted according to different working conditions.

[0081] This analysis method is based on the DNN deep neural network model, where DNN stands for Deep Neural Networks.

[0082] The model is used to explore the complex relationship between features and targets; and the model performance is represented by the model's loss function; the loss function is expressed as:

[0083]

[0084] Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample.

[0085] The mean square error between the predicted value and the true value is used to further reduce the impact of noise and disturbance on the model and improve the interpretability of the model. The traditional physical model is incorporated into it to enhance the robustness and interpretability of the model.

[0086] A physical loss term is introduced into the loss function of the model to achieve this goal; the loss function after introduction is expressed as:

[0087]

[0088] Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample; y i1 is the calculated value of the i-th sample; λ is the weight of the data-driven loss value, which is adjusted according to different working conditions.

[0089] Example 1

[0090] like Figure 4 The present invention is a method for analyzing the stress mechanism of a water-rich tunnel based on physical-data dual-drive. This implementation case is analyzed by comparing the method of the present invention with a data-driven neural network model and is elaborated in detail.

[0091] The data-driven neural network model loss function is expressed as:

[0092]

[0093] Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample;

[0094] The loss function of the force mechanism analysis method of water-rich tunnel based on physical-data dual drive is shown in formula (4);

[0095] The comparative analysis of the method of the present invention and the data-driven neural network model mainly includes the following steps:

[0096] (1) Data collection: Obtain model training samples based on monitoring data, geological structures, physical formulas, and tunnel structures;

[0097] (2) Data division: Filter the collected data to form a data set. Set the test data volume to 20, 40, 60, 80, 100, 200, 400, and 600 respectively, and conduct multiple groups of tests to analyze the generalization performance, fitting, and model stability of the model;

[0098] (3) Data preprocessing: The sample data is preprocessed using the StandardScaler method, which scales the features to conform to the standard normal distribution with a mean of 0 and a standard deviation of 1, so that different features have the same scale during the training process, thereby improving the training effect and stability of the model;

[0099] (4) Experimental grouping: The data-driven neural network was set as the control group, and the water-rich tunnel stress mechanism analysis method based on physical-data dual drive was set as the experimental group for control experiments.

[0100] (5) Experimental method: 80% of the samples are used as the training set of the model, and the remaining 20% ​​are used as the test set to verify the performance of the model, and the prediction results, loss value and accuracy are output;

[0101] (6) Result analysis: By drawing a line graph of data volume, model loss, and accuracy, the performance of the two models is intuitively displayed, and the curve fluctuations are analyzed to further evaluate the model stability; the curve trend is analyzed to evaluate the model learning ability, and different data volumes are tested to verify the advantages of this method over deep neural networks in fitting training data and test data.

[0102] Based on the deep neural network model, this method innovatively incorporates physical mechanisms and constructs a physical constraint framework for the model. This constraint framework not only effectively reduces the learning space of the model, guides it to focus on the solution set that conforms to physical principles, significantly improves the generalization ability of the model when facing unknown conditions, but also enhances its learning efficiency under limited data. Therefore, this method can handle the complexity of different geological conditions and the diversity of tunnel structures, achieve more accurate and reliable tunnel force prediction, and show its application potential in complex engineering environments.

[0103] Example 2

[0104] The difference between this embodiment and embodiment 1 is that embodiment 1 shows that this method can enhance the performance of the neural network model when the data volume is small and has stronger generalization ability, while embodiment 2 shows that this method has stronger processing ability for complex systems than physical models.

[0105] This example analysis includes the following steps:

[0106] (1) Data collection: same as (1) in Example 1;

[0107] (2) Data division: same as (2) in Example 1;

[0108] (3) Data preprocessing: The physical model does not require data preprocessing. The preprocessing of the test group is the same as (3) in Example 1.

[0109] (4) Experimental grouping: The physical model was used as the control group, and the force mechanism analysis method of the water-rich tunnel based on the dual drive of physics and data was used as the experimental group, and a control experiment was conducted.

[0110] The physical model does not require data preprocessing, and the preprocessing of the test group is the same as (4) in Example 1;

[0111] (5) Experimental method: Since the physical model does not need to be trained, the test data can be directly input, the calculation results are output, and the error value with the true value is calculated to obtain the mean square error value. The calculation formula can be expressed as:

[0112]

[0113] In the formula, η is the mean square error value; n is the amount of data; y c Calculated value for the physical model; y a is the actual value.

[0114] The experimental method of the experimental group was the same as (5) in Example 1;

[0115] (6) Result analysis: By drawing a line graph of data volume, model loss, and accuracy, the performance of the two models is intuitively displayed, and the curve fluctuations are analyzed to further evaluate the model stability; the curve trend is analyzed to evaluate the model learning ability, and different data volumes are tested to verify the advantages of this method over deep neural networks in fitting training data and test data.

[0116] Compared with the traditional physical model analysis method, this invention innovatively integrates deep neural networks with physical laws, which not only retains the rigor of the physical model, but also makes full use of the excellent ability of deep neural networks in capturing complex nonlinear relationships and dealing with uncertainties. This combination greatly improves the simulation accuracy of the physical model for the stress conditions of water-rich tunnels, making the prediction results more consistent with the actual recorded stress values, providing accuracy and reliability for structural stress analysis in tunnel construction and operation, and providing data reference for engineering decision-making.

[0117] The above is a preferred embodiment of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described above, and these modifications or additions should also be regarded as within the protection scope of the present invention.

Claims

1. A method for analyzing the stress mechanism of a water-rich tunnel based on physical and data dual drive, characterized in that: The method mainly includes the following steps: Step 1: Comparative analysis of data-driven neural network models They are data collection, data division, data preprocessing, experimental grouping, experimental methods, and model evaluation; During data collection, mechanical sensor data in the water-rich tunnel is collected. ω , consult the design drawings and exploration data to obtain tunnel structural data r, r0, r1, r2, r3, R and geological data k1, k2, k3, k4, P i ; where P ω is the water pressure, in MPa; r is the radial seepage radius of groundwater, in m; r0 is the inner radius of the secondary lining of the tunnel, in m; r1 is the inner radius of the initial support, in m; r2 is the inner radius of the reinforcement circle, in m; r3 is the outer radius of the reinforcement circle, in m; R is the outer radius of the study area, in m; k1 is the secondary lining permeability coefficient, in m / d; k2 is the initial support permeability coefficient, in m / d; k3 is the reinforcement ring permeability coefficient, in m / d; k4 is the surrounding rock permeability coefficient, in m / d; P i is the water pressure of the original seepage field, in MPa; Step 2: Physical Modeling The physical model is based on the following basic assumptions: 1) In high head areas, the circular tunnel is surrounded by groundwater; 2) The surrounding rock is a homogeneous, isotropic, continuous medium; 3) Seepage potential is caused only by changes in water pressure; 4) The water load is axisymmetrically distributed; the physical model is further derived into the tunnel force calculation formula, which is expressed as: In the formula, a and b are calculation parameters, which can be expressed as: According to formulas (1), (2), and (3), the water pressure behind the secondary lining, initial support, and reinforcement ring of the tunnel can be calculated under basic assumptions. However, this physical model cannot accurately simulate the complex geological conditions and physical relationships in actual situations. Therefore, it is necessary to add a deep neural network to learn monitoring data and further improve the model. Step 3: Neural network model establishment The input layer of the deep learning neural network model is the 11 parameters r, r0, r1, r2, r3, R, k1, k2, k3, k4, P in step 1. i There are two hidden layers, each consisting of 16 neurons. Each neuron performs a weighted summation of the output of the previous layer and performs a nonlinear transformation through the ReLU activation function to generate the output of this layer. The connection between the hidden layers forms the depth of the neural network. The last layer generates the network output, that is, the water pressure P at the radial seepage radius r of the groundwater. ω ; ReLU stands for Rectified Linear Unit. Step 4: Establishing a physical-data dual-driven model The model is based on previously successful deep learning models and physical models, and innovatively integrates physical knowledge with machine learning methods. The core of this innovation lies in the construction of a custom loss function, which not only includes the mean square error of the data, but also creatively introduces physical formulas as part of the loss function. Specifically, the equation describing the relationship between the water pressure of the tunnel structure and the geological structure and tunnel structure is first derived, and then the physical equation is applied to evaluate the consistency between the model prediction value and the actual physical law. The specific loss function is as follows: Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample; y i1 is the calculated value of the i-th sample; λ is the weight of the data-driven loss value, which is adjusted according to different working conditions.

2. According to claim 1, a method for analyzing the stress mechanism of a water-rich tunnel based on physical-data dual drive is characterized in that: This analysis method is based on the DNN deep neural network model, where DNN stands for Deep Neural Networks. The model is used to explore the complex relationship between features and targets; and the model performance is represented by the model's loss function; the loss function is expressed as: Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample.

3. According to claim 2, a method for analyzing the stress mechanism of a water-rich tunnel based on physical-data dual drive is characterized in that: The mean square error between the predicted value and the true value is used to further reduce the impact of noise and disturbance on the model and improve the interpretability of the model. The traditional physical model is incorporated into it to enhance the robustness and interpretability of the model.

4. According to claim 2, a method for analyzing the stress mechanism of a water-rich tunnel based on physical-data dual drive is characterized in that: A physical loss term is introduced into the loss function of the model to achieve this goal; the loss function after introduction is expressed as: Where n is the amount of data; y i1 is the true value of the i-th sample; y i2 is the predicted value of the i-th sample; y i1 is the calculated value of the i-th sample; λ is the weight of the data-driven loss value, which is adjusted according to different working conditions.

5. According to claim 1, a method for analyzing the stress mechanism of a water-rich tunnel based on physical-data dual drive, characterized in that: In step 3, the neural network training process includes the following four parts: 1) Forward propagation: Starting from the input layer, the output of each layer of neurons is calculated in sequence until the result of the output layer is obtained; 2) Back propagation: Use the chain rule to calculate the gradient of the loss function with respect to each weight and bias term. These gradients indicate how to adjust the weights to reduce the loss. 3) Optimization algorithm: Adopt the Adam optimization algorithm and use the gradient calculated by back-propagation to update the weights and biases, and gradually iterate and optimize the model; the full name of Adam is AdaptiveMomentEstimation; 4) Loss function: The mean square error (MSE) is used to measure the gap between the model prediction result and the true label. The full name of MSE is Mean Squared Error.

6. The method for analyzing the stress mechanism of a water-rich tunnel based on physical-data dual drive according to claim 1 is characterized in that: The theoretical process of the analysis method is as follows: (1) Data collection: Obtain model training samples based on monitoring data, geological structures, physical formulas, and tunnel structures; (2) Data division: Filter the collected data to form a data set. Set the test data volume to 20, 40, 60, 80, 100, 200, 400, and 600 respectively, and conduct multiple groups of tests to analyze the generalization performance, fitting, and model stability of the model; (3) Data preprocessing: The sample data is preprocessed using the StandardScaler method, which scales the features to conform to the standard normal distribution with a mean of 0 and a standard deviation of 1, so that different features have the same scale during the training process, thereby improving the training effect and stability of the model; (4) Experimental grouping: The other methods were set as the control group, and the water-rich tunnel stress mechanism analysis method based on physical-data dual drive was set as the experimental group for control experiments. (5) Experimental method: 80% of the samples are used as the training set of the model, and the remaining 20% ​​are used as the test set to verify the performance of the model, and the prediction results, loss value and accuracy are output; (6) Result analysis: By drawing a line graph of data volume, model loss, and accuracy, the performance of the two models is intuitively displayed, and the curve fluctuations are analyzed to further evaluate the model stability; the curve trend is analyzed to evaluate the model learning ability, and different data volumes are tested to verify the advantages of this method over deep neural networks in fitting training data and test data.

Images