Shield tunnel long-term stability evaluation method and system based on hybrid drive and computer readable storage medium

Abstract

The invention discloses a shield tunnel long-term stability evaluation method and system based on hybrid drive and a computer readable storage medium, and the evaluation method comprises the following steps: S1, obtaining shield tunnel features and tunnel stability grades, and building a shield tunnel long-term stability database; s2, preprocessing the data; s3, constructing a causal relationship graph, and identifying a feature Q influencing the stability level of the tunnel; s4, intervening the feature Q, observing the influence of the intervened feature on the stability level of the tunnel, and removing irrelevant features by using conditional independence; s5, performing standardization processing on the data, and dividing the data into a training set, a verification set and a test set; s6, training the physical deep learning model and the extreme learning machine model based on difference grey wolf optimization by using the training set to obtain a hybrid drive global model; and S7, inputting the test set into the hybrid drive global model to obtain a predicted tunnel stability level, and judging whether an error between an index and an actual value meets a threshold value or not.

Description

Technical Field

[0001] The present invention relates to the technical field of long-term stability evaluation of shield tunnels under soft rock creep effects, and in particular to a hybrid-drive-based long-term stability evaluation method, system, and computer-readable storage medium. Background Art

[0002] With the continuous expansion of cities, the scenarios for constructing rail transit using the shield method are becoming increasingly complex. Shield tunnels inevitably penetrate soft rock and soil layers. Soft rock is characterized by low strength, weak lithology, easy deformation, and strong rheological properties. This makes construction prone to problems such as large deformation and cracking of initial support structures. Furthermore, after tunnel construction, the tunnel structure is susceptible to the creep effects of the soft rock. The stresses on the structure can easily reach their limits, leading to cracking, crushing, and falling blocks in the lining. This poses significant risks to driving safety and poses a serious safety hazard. Therefore, it is necessary to conduct a theoretical analysis and establish a calculation method for the long-term stability of shield tunnels under the influence of soft rock creep.

[0003] In addition, a reasonable and effective safety assessment method is the basis for conducting tunnel structure safety assessments. Many scholars have conducted extensive research on long-term safety assessment methods for tunnel structures. Some researchers focus on studying problems encountered during the operation of tunnel structures, analyzing factors affecting their long-term safety, identifying key influencing indicators, summarizing lining damage and structural safety conditions, establishing qualitative and quantitative relationships, and conducting simple structural safety assessments. Other researchers, based on the characteristics of operational tunnel structural problems, have studied the effects of factors such as lining cavities, material degradation, lining fractures, and water leakage on the mechanical properties and safety factors of tunnel structures, thereby establishing subjective safety assessment methods. The above traditional theoretical systems mainly support knowledge-based decision-making methods for the long-term safety assessment of tunnel structures. However, in the case of soft rock tunnels, there is significant spatiotemporal uncertainty in hazard distribution and structural response, which poses a challenge to the accurate assessment of the long-term safety status of tunnel structures using existing assessment methods and indicator systems. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for evaluating the long-term stability of a shield tunnel based on hybrid drive, comprising the following steps:

[0005] S1. Obtain multiple sets of shield tunnel characteristics and corresponding tunnel stability levels, and establish a long-term stability database for shield tunnels;

[0006] S2. Preprocess the data in the shield tunnel long-term stability database to remove data segments containing invalid values and singular values;

[0007] S3. Based on the structural causal model, a causal relationship diagram is constructed between each feature and between each feature and the tunnel stability level. The causal relationship between each feature and the tunnel stability level is clarified, and the feature Q that affects the tunnel stability level is identified.

[0008] S4. Based on counterfactual reasoning, we intervene in feature Q, observe the impact of the intervened features on the tunnel stability level, and use the conditional independence test to remove irrelevant features.

[0009] S5. Standardize the data in the database and divide the processed data into training set, validation set and test set;

[0010] S6. Use the training set to train the physical deep learning model and the extreme learning machine model based on differential grey wolf optimization, and use the validation set to verify the hyperparameters of the physical deep learning model and the extreme learning machine model, thereby obtaining a hybrid drive global model;

[0011] S7, inputting the test set into the hybrid drive global model to obtain the predicted tunnel stability level;

[0012] If the error between the evaluation index and the actual value is less than the preset threshold, the construction of the hybrid drive global model is completed and the process proceeds to step S8), otherwise it returns to step S1);

[0013] S8. Obtain the characteristics of the shield tunnel to be evaluated and input them into the hybrid drive global model to obtain the tunnel stability level.

[0014] Furthermore, the shield tunnel long-term stability database includes sub-database A, sub-database B, sub-database C, sub-database D and sub-database E.

[0015] The sub-database A includes several tunnel stability levels. The tunnel stability levels are divided into four evaluation levels: safe, basically safe, with hidden dangers, and dangerous according to the degree of tunnel damage.

[0016] The sub-database B includes several tunnel structural characteristics, including the initial support form and the secondary lining form.

[0017] The sub-database C includes several tunnel structural mechanical properties, including initial supporting steel arch stress, secondary lining axial force and secondary lining bending moment.

[0018] The sub-database D includes several tunnel rock mass properties, including rock mass integrity coefficient, uniaxial compression, rock mass quality index, ultimate creep strain and rock mass bedding inclination.

[0019] The sub-database E includes several other influencing factors, including tunnel burial depth, lateral pressure coefficient, underground water storage conditions and tunnel operation time.

[0020] Furthermore, in step S3, a structural causal model is constructed based on the causal Bayesian network and the structural equation model.

[0021] The causal relationship graph is a directed acyclic graph, in which the nodes represent the features in the structural causal model, and the directed edges represent the direction of causal influence.

[0022] Each node Y i The causal function of is as follows:

[0023] Y i =f i (Pa(Y i ),U i ) (1)

[0024] Where:

[0025] Pa(Y i ) represents feature Y i The parent node of

[0026] U i For feature Y i Related interference factors;

[0027] f i is a function that describes the causal mechanism.

[0028] Furthermore, in step S4, the d-separation rule is applied to determine whether the feature node in the causal relationship graph is an irrelevant feature.

[0029] Furthermore, in step S5, the standardization processing includes performing one-hot processing on discrete data and performing maximum and minimum value normalization processing on numerical continuous data.

[0030] Furthermore, in step S6, the method for establishing a physical deep learning model includes the following steps:

[0031] S6.1.1. Establish the dynamic equation of the dynamic system. The dynamic equation is as follows:

[0032]

[0033] Where: u(t,x) represents the potential solution, i.e., the tunnel stability evaluation level; t is time; x is the input vector; N(u) is the parameterized nonlinear operator;

[0034] S6.1.2. Based on the dynamics equations, construct a neural network based on physical information. The input layer of the neural network is the eigenvalues after removing irrelevant features in step S4. The hidden layer is a fully connected layer. The output layer is the potential solution. All layers use the Reluctant Lu function as the activation function.

[0035] The neural network f(t,x) based on physical information is as follows:

[0036]

[0037] S6.1.3. Based on the dynamics equation, construct a deep neural network that approximates the potential solution u(t,x) of the dynamics equation. The input layer of the deep neural network is the observation vector x and time t, the hidden layer is a fully connected layer, and the output layer is the numerical solution approximating u(t,x). All layers use the Swish function as the activation function.

[0038] S6.1.4. Use automatic differentiation to process the terms of the dynamical equations of the two neural networks, computing the partial derivatives of the potential solution with respect to the input vector and time, so that the physics-based neural network and the deep neural network approximating the potential solution have the same network parameters.

[0039] S6.1.5. Construct a loss function to optimize the shared parameters of the two neural networks and feed the calculation results back to the neural networks to update the parameters and form the optimal local model.

[0040] The loss function MSE is as follows:

[0041] MSE=MSE u +MSE f (4)

[0042]

[0043] Where:

[0044] MSE u is the data error of the neural network;

[0045] MSE f is the physical information error of the neural network;

[0046] N u is the number of numerical solutions that approximate u(t,x);

[0047] is the i-th numerical solution predicted by the neural network approximating u(t,x), i = 1, 2, ..., N u ;

[0048] u i is the i-th potential solution;

[0049] N f is the number of predicted potential solutions;

[0050] is the i-th potential solution predicted by the neural network f(t,x), i=1,2,…,N f .

[0051] Furthermore, in step S6, the method for establishing an extreme learning machine model based on differential grey wolf optimization includes the following steps:

[0052] S6.2.1. Initialize the weights W from the input layer to the hidden layer of the extreme learning machine, the bias b, and the population parameters in the differential gray wolf optimization algorithm;

[0053] S6.2.2. Select the ReLU function as the activation function of the extreme learning machine, and set the number of neurons N in the hidden layer of the extreme learning machine and the maximum number of iterations T. max ;

[0054] S6.2.3. Use the training set to build an extreme learning machine model and calculate the average sum of squared errors of the sample data in the training set as the fitness of each gray wolf. The fitness calculation formula is:

[0055]

[0056] Where:

[0057] Fitness i is the fitness of the i-th gray wolf in the gray wolf population;

[0058] N is the total number of sample data in the training set;

[0059] y k is the actual target value of the kth sample data;

[0060] The predicted output result of the k-th sample data in the extreme learning machine model;

[0061] S6.2.4. Select the three wolves ranked in the top three by fitness, mark their individual positions, and use them as the decision layer of the extreme learning machine model.

[0062] S6.2.5. Determine whether the current number of iterations T is greater than the maximum number of iterations T max Is it true? If so, return the global optimal position as the optimal weight W from the input layer to the hidden layer of the extreme learning machine model * and bias b * Otherwise, update the position of each individual gray wolf in the gray wolf population, set T = T + 1, and return to step S6.2.3;

[0063] S6.2.6. Using weight W * and bias b * The extreme learning machine model is trained to output the deep feature representation H of the hidden layer for predicting the tunnel stability level and the weight β of the output layer to obtain the optimal local model;

[0064] The function of deep feature representation H for tunnel stability grade prediction is as follows:

[0065] H=f(W*X+b * ) (8)

[0066] The calculation formula of the output layer weight β is as follows:

[0067] β=(H T H) -1 H T Y (9)where: H is the deep feature representation for tunnel stability grade prediction;

[0068] X is the input feature matrix;

[0069] H T is the transposed matrix of H;

[0070] Y is the predicted tunnel stability grade.

[0071] Furthermore, in step S6, the method for establishing the hybrid drive global model includes the following steps:

[0072] S6.3.1. Based on edge intelligence technology, use the training set to train and optimize the physical deep learning model and the extreme learning machine based on differential grey wolf optimization;

[0073] S6.3.2. The cloud center sends the initially trained hybrid drive model to an edge node, i.e., the aggregation server. Other edge nodes participate in the model training of the aggregation server.

[0074] S6.3.3. Use the data set in the database to train the local model to obtain the optimal local model. The edge node sends the updated local model to the aggregation server to obtain the optimal hybrid drive global model.

[0075] The second object of the present invention is to provide a system applying the shield tunnel long-term stability evaluation method, comprising a data acquisition unit, a data preprocessing unit, a model building and training unit and a tunnel stability evaluation unit.

[0076] The data acquisition unit is used to collect tunnel information and conduct causal analysis, input tunnel stability assessment indicators, and build relevant sub-databases.

[0077] The data preprocessing unit is used to perform data standardization operations and shuffle the data set into a training set, a validation set, and a test set.

[0078] The model building and training unit is used to establish a physical deep learning model and an extreme learning machine based on differential grey wolf optimization, and is used to build a physical-data-knowledge-driven hybrid model for predicting the stability level of shield tunnels. Edge intelligence technology is used to use the training set to perform model training and hyperparameter optimization to obtain the optimal hybrid driving model.

[0079] The tunnel stability evaluation unit tests the optimal hybrid drive model using a test set and outputs a predicted tunnel stability level.

[0080] The third object of the present invention is to provide a computer-readable storage medium, which stores a computer program. When the computer program runs on an electronic device, the electronic device executes the above-mentioned method for evaluating the long-term stability of a shield tunnel based on hybrid drive.

[0081] The technical effects of the present invention are unquestionable, and the beneficial effects of the present invention are as follows:

[0082] 1. The hybrid-drive-based long-term stability evaluation method of a shield tunnel of the present invention fully considers the causal relationship between the tunnel's structural characteristics, structural mechanical characteristics, rock mass characteristics, and other influencing factors and stability indicators, constructs a shield tunnel stability evaluation level, and provides reliable protection for the long-term stability monitoring of shield tunnels.

[0083] 2. The use of structural causal models to test the reliability and causal relationship of parameters selected based on literature surveys, numerical analysis and expert experience is of great significance to enhancing the interpretability of the model and ensuring the scientificity, rationality and correctness of the model parameter input.

[0084] 3. In the hybrid-driven long-term stability evaluation method for shield tunnels of the present invention, data standardization processing is adopted to accelerate the convergence performance of model training; a physical deep learning model and an extreme learning machine based on differential grey wolf optimization are used to construct a physical-data-knowledge-driven hybrid model for predicting the stability level of shield tunnels, which further improves the fitting accuracy of the system variable characteristics and enhances the robustness and applicability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is the overall flow chart of the operation of the present invention.

[0086] Figure 2 This is a schematic diagram of a neural network structure based on physical information. DETAILED DESCRIPTION

[0087] The present invention will be further described below with reference to the following examples, but it should not be understood that the scope of the present invention is limited to the following examples. Without departing from the above technical ideas of the present invention, various substitutions and modifications can be made according to common technical knowledge and customary means in the art, and all should be included in the scope of protection of the present invention.

[0088] Example 1:

[0089] A method for evaluating the long-term stability of a shield tunnel based on hybrid drive includes the following steps:

[0090] S1. Obtain multiple sets of shield tunnel characteristics and corresponding tunnel stability levels, and establish a long-term stability database for shield tunnels;

[0091] S2. Preprocess the data in the shield tunnel long-term stability database to remove data segments containing invalid values and singular values;

[0092] S3. Based on the structural causal model, a causal relationship diagram is constructed between each feature and between each feature and the tunnel stability level. The causal relationship between each feature and the tunnel stability level is clarified, and the feature Q that affects the tunnel stability level is identified.

[0093] S4. Based on counterfactual reasoning, we intervene in feature Q, observe the impact of the intervened features on the tunnel stability level, and use the conditional independence test to remove irrelevant features.

[0094] S5. Use the training set to train the physical deep learning model and the extreme learning machine model based on differential grey wolf optimization, and use the validation set to verify the hyperparameters of the physical deep learning model and the extreme learning machine model, thereby obtaining a hybrid drive global model;

[0095] S6. Use the training set and validation set to train the physical deep learning model and the extreme learning machine model based on differential grey wolf optimization to obtain a hybrid drive global model;

[0096] S7, inputting the test set into the hybrid drive global model to obtain the predicted tunnel stability level;

[0097] If the error between the evaluation index and the actual value is less than the preset threshold, the construction of the hybrid drive global model is completed and the process proceeds to step S8), otherwise it returns to step S1);

[0098] S8. Obtain the characteristics of the shield tunnel to be evaluated and input them into the hybrid drive global model to obtain the tunnel stability level.

[0099] Example 2:

[0100] The main structure of this embodiment is the same as that of embodiment 1. Furthermore, the shield tunnel long-term stability database includes sub-database A, sub-database B, sub-database C, sub-database D and sub-database E.

[0101] The sub-database A includes several tunnel stability levels. The tunnel stability levels are divided into four evaluation levels: safe, basically safe, with hidden dangers, and dangerous according to the degree of tunnel damage.

[0102] The sub-database B includes several tunnel structural characteristics, including the initial support form and the secondary lining form.

[0103] The sub-database C includes several tunnel structural mechanical properties, including initial supporting steel arch stress, secondary lining axial force and secondary lining bending moment.

[0104] The sub-database D includes several tunnel rock mass properties, including rock mass integrity coefficient, uniaxial compression, rock mass quality index, ultimate creep strain and rock mass bedding inclination.

[0105] The sub-database E includes several other influencing factors, including tunnel burial depth, lateral pressure coefficient, underground water storage conditions and tunnel operation time.

[0106] Example 3:

[0107] The main structure of this embodiment is the same as any one of Embodiments 1 to 2. Furthermore, in step S3, a structural causal model is constructed based on the causal Bayesian network and the structural equation model.

[0108] The causal relationship graph is a directed acyclic graph, in which the nodes represent the features in the structural causal model, and the directed edges represent the direction of causal influence.

[0109] Each node Y i The causal function of is as follows:

[0110] Y i =f i (Pa(Y i ),U i ) (1)

[0111] Where:

[0112] Pa(Y i ) represents feature Y i The parent node of

[0113] U i For feature Y i Related interference factors;

[0114] fi is a function that describes the causal mechanism.

[0115] Example 4:

[0116] The main structure of this embodiment is the same as any one of Embodiments 1 to 3. Furthermore, in step S4, the d-separation rule is applied to determine whether the feature node in the causal relationship graph is an irrelevant feature.

[0117] Example 5:

[0118] The main structure of this embodiment is the same as any one of Embodiments 1 to 4. Furthermore, in step S5, the standardization processing includes performing one-hot processing on discrete data and performing maximum and minimum value normalization processing on numerical continuous data.

[0119] Example 6:

[0120] The main structure of this embodiment is the same as any one of Embodiments 1 to 5. Furthermore, in step S6, the method for establishing a physical deep learning model includes the following steps:

[0121] S6.1.1. Establish the dynamic equation of the dynamic system. The dynamic equation is as follows:

[0122]

[0123] Where: u(t,x) represents the potential solution, i.e., the tunnel stability evaluation level; t is time; x is the input vector; N(u) is the parameterized nonlinear operator;

[0124] S6.1.2. Based on the dynamics equations, construct a neural network based on physical information. The input layer of the neural network is the eigenvalues after removing irrelevant features in step S4. The hidden layer is a fully connected layer. The output layer is the potential solution. All layers use the Reluctant Lu function as the activation function.

[0125] The neural network f(t,x) based on physical information is as follows:

[0126]

[0127] S6.1.3. Based on the dynamics equation, construct a deep neural network that approximates the potential solution u(t,x) of the dynamics equation. The input layer of the deep neural network is the observation vector x and time t, the hidden layer is a fully connected layer, and the output layer is the numerical solution approximating u(t,x). All layers use the Swish function as the activation function.

[0128] S6.1.4. Use automatic differentiation to process the terms of the dynamical equations of the two neural networks, computing the partial derivatives of the potential solution with respect to the input vector and time, so that the physics-based neural network and the deep neural network approximating the potential solution have the same network parameters.

[0129] S6.1.5. Construct a loss function to optimize the shared parameters of the two neural networks and feed the calculation results back to the neural networks to update the parameters and form the optimal local model.

[0130] The loss function MSE is as follows:

[0131] MSE=MSE u +MSE f (4)

[0132]

[0133]

[0134] Where:

[0135] MSE u is the data error of the neural network;

[0136] MSE f is the physical information error of the neural network;

[0137] N u is the number of numerical solutions that approximate u(t,x);

[0138] is the i-th numerical solution predicted by the neural network approximating u(t,x), i = 1, 2, ..., N u ;

[0139] u i is the i-th potential solution;

[0140] N f is the number of predicted potential solutions;

[0141] is the i-th potential solution predicted by the neural network f(t,x), i=1,2,…,N f .

[0142] Example 7:

[0143] The main structure of this embodiment is the same as any one of Embodiments 1 to 6. Furthermore, in step S6, the method for establishing an extreme learning machine model based on differential grey wolf optimization includes the following steps:

[0144] S6.2.1. Initialize the weights W from the input layer to the hidden layer of the extreme learning machine, the bias b, and the population parameters in the differential gray wolf optimization algorithm;

[0145] S6.2.2. Select the ReLU function as the activation function of the extreme learning machine, and set the number of neurons N in the hidden layer of the extreme learning machine and the maximum number of iterations T. max ;

[0146] S6.2.3. Use the training set to build an extreme learning machine model and calculate the average sum of squared errors of the sample data in the training set as the fitness of each gray wolf. The fitness calculation formula is:

[0147]

[0148] Where:

[0149] Fitness i is the fitness of the i-th gray wolf in the gray wolf population;

[0150] N is the total number of sample data in the training set;

[0151] y k is the actual target value of the kth sample data;

[0152] The predicted output result of the k-th sample data in the extreme learning machine model;

[0153] S6.2.4. Select the three wolves ranked in the top three by fitness, mark their individual positions, and use them as the decision layer of the extreme learning machine model.

[0154] S6.2.5. Determine whether the current number of iterations T is greater than the maximum number of iterations T max Is it true? If so, return the global optimal position as the optimal weight W from the input layer to the hidden layer of the extreme learning machine model * and bias b * Otherwise, update the position of each individual gray wolf in the gray wolf population, set T = T + 1, and return to step S6.2.3;

[0155] S6.2.6. Using weight W * and bias b * The extreme learning machine model is trained to output the deep feature representation H of the hidden layer for predicting the tunnel stability level and the weight β of the output layer to obtain the optimal local model;

[0156] The function of deep feature representation H for tunnel stability grade prediction is as follows:

[0157] H=f(W * X+b* ) (8)

[0158] The calculation formula of the output layer weight β is as follows:

[0159] β=(H T H) -1 H T Y (9)where: H is the deep feature representation for tunnel stability grade prediction;

[0160] X is the input feature matrix;

[0161] H T is the transposed matrix of H;

[0162] Y is the predicted tunnel stability grade.

[0163] Example 8:

[0164] The main structure of this embodiment is the same as any one of Embodiments 1 to 7. Furthermore, in step S6, the method for establishing the hybrid drive global model includes the following steps:

[0165] S6.3.1. Based on edge intelligence technology, use the training set to train and optimize the physical deep learning model and the extreme learning machine based on differential grey wolf optimization;

[0166] S6.3.2. The cloud center sends the initially trained hybrid drive model to an edge node, i.e., the aggregation server. Other edge nodes participate in the model training of the aggregation server.

[0167] S6.3.3. Use the data set in the database to train the local model to obtain the optimal local model. The edge node sends the updated local model to the aggregation server to obtain the optimal hybrid drive global model.

[0168] Example 9:

[0169] The main structure of this embodiment is the same as any one of Embodiments 1 to 8. Furthermore, a method for evaluating the long-term stability of a shield tunnel based on hybrid drive includes constructing a long-term stability database for a shield tunnel, constructing a hybrid drive model, and predicting and evaluating the long-term stability of a shield tunnel. The details are as follows:

[0170] S1. The evaluation standard for tunnel structure stability is established in the following steps:

[0171] S1.1. Qualitative description of the tunnel structure stability level:

[0172] The stability of tunnel structures is qualitatively divided into four levels using a four-level classification method. Level 1 is a safe state, based on the absence of damage or minor cracks on the tunnel surface; Level 2 is basically safe, based on the presence of minor cracks on the tunnel structure surface and no water seepage; Level 3 is a hidden danger, based on the presence of many cracks, damage or minor water seepage on the tunnel surface; Level 4 is a dangerous state, based on the presence of extensive damage or falling rocks that prevent normal use of the tunnel structure.

[0173] S1.2. Quantitatively describe the tunnel structure stability level:

[0174] Based on the classification limits for the technical condition assessment of civil structures in the Technical Specifications for Highway Tunnel Maintenance (JTG H12-2015), and taking into account actual conditions, the functional status of civil structures was divided into four levels. A normalization method was used to quantify each level and assign it to a corresponding quantification interval. The quantification intervals were then used to distinguish the stability levels of tunnel structures.

[0175] S2. Construction of shield tunnel long-term stability database is mainly based on Python language. The steps are as follows:

[0176] S2.1 Construction of long-term stability database of shield tunnels:

[0177] The construction includes: sub-database A consisting of quantified tunnel stability grades; sub-database B consisting of tunnel structural characteristics consisting of initial support form and secondary lining form; sub-database C consisting of tunnel structural mechanical characteristics consisting of initial support steel arch stress, secondary lining axial force and secondary lining bending moment; sub-database D consisting of tunnel rock mass characteristics consisting of rock mass integrity coefficient, uniaxial compression, rock mass quality index, ultimate creep strain and rock mass bedding inclination; sub-database E consisting of other influencing factors consisting of tunnel burial depth, lateral pressure coefficient, groundwater storage conditions and tunnel operation time; and data preprocessing to eliminate data segments containing invalid values and singular values.

[0178] S2.2 Perform causal analysis on text data information:

[0179] In conjunction with the long-term stability database of shield tunnels, the structural causal model, based on domain knowledge and data analysis, constructs a causal relationship diagram (directed acyclic graph, DAG) between features. This causal relationship diagram clarifies the causal relationship between features, identifies features that directly or indirectly affect the target variable, and evaluates the causal effect of each feature on the target variable through counterfactual reasoning. Conditional independence tests are used to remove irrelevant features. The structural causal model is used to select features at both the causal relationship and effect significance levels, verifying the reliability of parameters selected based on literature research, numerical analysis, and expert experience, and understanding the causal relationship between tunnel parameters and stability indicators.

[0180] S2.3 Data standardization and data set division:

[0181] The data types of the shield tunnel long-term stability database are defined as discrete data and numerical continuous data. The discrete data are processed as one-hot, and the numerical continuous data are normalized to the maximum and minimum values. Sub-database A is defined as the output parameter, and sub-databases B, C, D, and E are defined as input parameters. The two sets of data are merged to obtain a set of text information data, which is shuffled and divided into training set, validation set, and test set in a ratio of 8:1:1.

[0182] S3. Construction of a hybrid-driven shield tunnel long-term stability evaluation model based on Python language, including:

[0183] S3.1 Building a physical deep learning model:

[0184] Establish the dynamic equations of the dynamic system, construct a neural network f(t,x) based on physical information and a deep neural network that approximates the potential solution u(t,x) of the dynamic equation. At the same time, use automatic differentiation to process each term of the partial differential equation and calculate the partial derivatives of the potential solution with respect to the observation vector and time, so that the neural network f(t,x) based on physical information and the deep neural network that approximates the potential solution have the same network parameters but different activation functions; construct a loss function to optimize the shared parameters of the two neural networks, and feed the calculation results back to the optimizer to achieve parameter updates and form the optimal local model.

[0185] S3.2 Establish an extreme learning machine based on differential grey wolf optimization:

[0186] Initialize the weights and biases from the input layer to the hidden layer of the extreme learning machine and the population in the differential gray wolf optimization algorithm, set hyperparameters such as the number of hidden layer neurons and activation function, use gray wolf optimization to simulate the hunting behavior of wolves to search for the optimal solution, optimize the weights and biases of the ELM model, and enhance the exploration ability through differential operations in the iterative process to avoid falling into the local optimum, obtain the optimal parameters of the extreme learning machine, and finally obtain the optimal local model through alternating optimization and training.

[0187] S3.3 Build a physics-data-knowledge driven hybrid model:

[0188] Edge intelligence technology is used to train and optimize the physical deep learning model and the extreme learning machine based on differential grey wolf optimization. The cloud center sends the initially trained hybrid drive model in its entirety to an edge node, namely the aggregation server (AS). Other edge nodes participate in the model training of the AS and use local data to train local models to obtain the optimal local model. The edge node then sends the updated local model to the AS to obtain the optimal hybrid drive global model.

[0189] S4. Prediction and evaluation of long-term stability of shield tunnels, including:

[0190] The optimal hybrid drive model is validated using the test set, and the predicted tunnel stability level is output.

[0191] Example 10:

[0192] The main structure of this embodiment is the same as that of Example 9. Furthermore, a qualitative description is made of the stability level of the tunnel structure. Based on the qualitative description of the degree of tunnel damage, a four-level grading method is used to divide the shield tunnel structure stability level into four evaluation levels: safe, basically safe, with hidden dangers and dangerous. Level one is a safe state, and the qualitative judgment is based on the fact that the tunnel surface has no damage or slight cracks; level two is basically safe, and the qualitative judgment is based on the fact that the tunnel structure surface has slight cracks and no water seepage; level three is a hidden danger, and the qualitative judgment is based on the fact that the tunnel surface has many cracks, damage or slight water seepage; level four is a dangerous state, and the qualitative judgment is based on the fact that the tunnel structure has a lot of damage or falling rocks and cannot be used normally.

[0193] Tunnel structural stability levels were quantitatively described. Based on the threshold values for the assessment of the technical condition of civil structures in the Technical Specification for Highway Tunnel Maintenance (JTG H12-2015), the assessment criteria for tunnel portals, portals, lining structures, lining leakage, pavement, maintenance roads, and drainage facilities were considered. Four grades were assigned to each grade using a normalization method. Each grade was quantified and assigned to a corresponding quantization interval (Table 1). To avoid data being zero during the normalization process, the normalization interval was set to [0.1, 1.1].

[0194] Table 1 Tunnel stability grade evaluation standards

[0195]

[0196] A long-term stability database for shield tunnels was constructed, where sub-database A represents the quantified tunnel stability grade, sub-database B represents the tunnel structural characteristics consisting of the initial support type and secondary lining type, sub-database C represents the tunnel structural mechanical characteristics consisting of the initial support steel arch stress, secondary lining axial force, and secondary lining bending moment, and sub-database D represents the tunnel rock mass characteristics consisting of rock mass integrity coefficient, uniaxial compression, rock mass quality index, ultimate creep strain, and rock mass bedding inclination. The selection of these indicators was based on the following references: 1) a literature review of factors affecting tunnel structural safety under the soft rock creep effect, 2) analysis of a large amount of field monitoring data and numerical simulation data, and 3) case experience in tunnel structural safety analysis.

[0197] Perform causal analysis on the text data information and extract the sub-database described in step 2.1 of Example 8. Furthermore, based on domain knowledge and data analysis, and in combination with the database, a structural causal model is used to construct a causal relationship diagram between features. The causal relationship between features is clarified through the causal relationship diagram, and features that directly or indirectly affect the target variable are identified. The structural causal model is a model that combines the causal Bayesian network and the structural equation model. It combines domain knowledge and data analysis to establish a directed acyclic graph to represent the causal relationship. The nodes represent the features in the model, and the directed edges indicate the direction of the causal influence; each node Y i It is represented by a structural equation that defines how the feature is affected by its parent node, in the form of: i =f i (Pa(Y i ),U i ), where Pa(Y i ) represents feature Y i The parent node, U i For feature Y i Related interference factors, f i is a function that describes the causal mechanism. Using the causal diagram and structural equation model, if a change in one node directly causes a change in another node, then the node has a direct causal relationship with the other node. If a node affects another node through multiple mediating features, then the node has an indirect causal relationship with the other node.

[0198] Furthermore, the causal effect of each feature on the target variable is evaluated through counterfactual reasoning, and finally, irrelevant features are removed using a conditional independence test. The structural causal model uses the symbol do(X=x0) to represent an intervention on feature X, forcibly modifying it to x0, and observing its impact on the target variable to evaluate the causal effect of the feature on the target variable. Combined with the constructed causal relationship graph, the d-separation rule is applied. After checking a given set of conditions, all nodes on the path are conditionalized given the value of certain features. If the path between two nodes is disconnected, the two nodes are conditionally independent under the given set of conditions, and irrelevant features are removed based on this.

[0199] Furthermore, a structural causal model was used to select features at both the causal relationship and effect significance levels. This validated the reliability of parameters selected based on literature research, numerical analysis, and expert experience, and also determined the causal relationship between tunnel parameters and stability indicators. Counterfactual reasoning and conditional independence tests within the structural causal model were used to determine the direct or indirect impact of each feature on the stability indicators, as well as the extent of that impact. Unrelated features were then removed to verify the reliability of the selected parameters, ultimately achieving feature selection.

[0200] Step 2.3: Data standardization and data set partitioning. We will use the following formula to process all numerical continuous parameters to improve the speed and accuracy of the model in finding the optimal solution:

[0201]

[0202] where x i is the i-th value in the data set, x max is the maximum value of the data set, x min is the minimum value of the data set.

[0203] Furthermore, we will process all discrete parameters to improve the convergence speed. The following is an example of the one-hot operation:

[0204]

[0205] Furthermore, sub-database A is defined as the output parameter, and sub-database B, sub-database C, sub-database D and sub-database E are defined as input parameters. The two sets of data are merged to obtain a set of text information data. Based on Python programming and random data packets, the corresponding input parameters and output parameters are shuffled and divided into training set, validation set and test set with a ratio of 8:1:1.

[0206] Step S3.1: Establish the dynamic equation of the dynamic system, which is expressed as:

[0207]

[0208] Here, u(t,x) represents the potential solution, x is the input vector, and N(u) is the parameterized nonlinear operator used to encapsulate the dynamic equations.

[0209] Furthermore, a neural network based on physical information is constructed The structural diagram is as follows Figure 2 As shown, the network input layer is the features selected as described in S2.2, the hidden layer is a fully connected layer, and the output layer is the potential solution, that is, the shield tunnel stability evaluation level. Each layer uses the ReLu function as the activation function, and the network is embedded with physical laws. In addition, a deep neural network is constructed to approximate the potential solution u(t,x) of the dynamic equation. The network structure is similar to the physical information-based neural network described above, and the activation function is the Tanh function. The goal is to approximate the solution of the partial differential equation. At the same time, automatic differentiation is used to process each term of the partial differential equation, and the partial derivatives of the potential solution with respect to the observation vector and time are calculated, so that the physical information-based neural network and the deep neural network approximating the potential solution have the same network parameters but different activation functions.

[0210] Furthermore, a loss function is constructed to optimize the shared parameters of the two neural networks, and the calculation results are fed back to the optimizer to update the parameters and form the optimal local model. The loss function is:

[0211] MSE=MSE u +MSE f

[0212]

[0213]

[0214] Among them, MSE u and MSE f They are the data error and physical information error of the neural network respectively. The calculation error is fed back to the optimizer during the training process to update the neural network parameters.

[0215] Step S3.2: Establish an extreme learning machine based on differential grey wolf optimization. Based on the Python language and the sklearn resource package, a custom extreme learning machine based on differential grey wolf optimization (DGWO-ELM) is constructed. The wolf pack positions of the DGWO algorithm are randomly initialized, represented by the weights W and bias b from the input layer to the hidden layer, and hyperparameters such as the number of hidden layer neurons N and the maximum number of iterations T are set.

[0216] Furthermore, in each iteration, the DGWO algorithm is used to update the position of the wolf pack, simulate the behavior of the wolf pack, and enhance the gray wolf's search ability through differential operations, so that it can explore the possible solution space in a larger range. The prediction error of the ELM model is measured using the sum of squared errors:

[0217]

[0218] Among them, y k is the actual target value, is the predicted output of the ELM model.

[0219] Furthermore, the optimal weight W from the input layer to the hidden layer is obtained using DGWO optimization. * and bias b * To train the ELM model, the output of the hidden layer H = f(W * X+b * ), use the least squares method to calculate the weight of the output layer β=(H T H) -1 H T Y.

[0220] Step 3.3: Build a physical-data-knowledge-driven hybrid model, and use edge intelligence technology to train and optimize the physical deep learning model and the extreme learning machine based on differential gray wolf optimization. The cloud center will send the initially trained hybrid drive model in its entirety to an edge node, namely the aggregation server (AS). Other edge nodes participate in the model training of the AS, use local data to train local models, and obtain the optimal local model. The edge node then sends the updated local model to the AS to obtain the optimal hybrid drive global model.

[0221] Step 4: After step 3.3, the optimal hybrid drive model is trained. We input the test data into the optimal model to determine the tunnel stability rating. By comparing the model with the actual values, we verify the model's effectiveness and reliability using the MAPE metric. By inputting the required tunnel parameters into the model, we can obtain a tunnel stability evaluation under the specified operating conditions.

[0222] Example 11:

[0223] A system for applying the shield tunnel long-term stability evaluation method described in any one of Examples 1-10 includes a data acquisition unit, a data preprocessing unit, a model building and training unit, and a tunnel stability evaluation unit.

[0224] The data acquisition unit is used to collect tunnel information and conduct causal analysis, input tunnel stability assessment indicators, and build relevant sub-databases.

[0225] The data preprocessing unit is used to perform data standardization operations and shuffle the data set into a training set, a validation set, and a test set.

[0226] The model building and training unit is used to establish a physical deep learning model and an extreme learning machine based on differential grey wolf optimization, and is used to build a physical-data-knowledge-driven hybrid model for predicting the stability level of shield tunnels. Edge intelligence technology is used to use the training set to perform model training and hyperparameter optimization to obtain the optimal hybrid driving model.

[0227] The tunnel stability evaluation unit tests the optimal hybrid drive model using the test set and outputs a predicted tunnel stability level.

[0228] Example 12:

[0229] A computer-readable storage medium stores a computer program. When the computer program is run on an electronic device, the electronic device executes the hybrid drive-based long-term stability evaluation method for a shield tunnel as described in any one of Examples 1 to 10.

[0230] Example 13:

[0231] The main structure of this embodiment is the same as any one of Embodiments 1 to 12. Furthermore, a method for evaluating the long-term stability of a shield tunnel based on hybrid drive includes:

[0232] S1. Establishment of evaluation standards for tunnel structure stability: A combination of qualitative and quantitative analysis methods is used to qualitatively describe the tunnel structure stability level, and then quantitatively describe it. The long-term stability of tunnel structures is qualitatively divided into four levels. Each level is quantified based on the "Technical Specification for Highway Tunnel Maintenance" (JTG H12-2015) and actual conditions, and corresponding quantitative ranges are assigned.

[0233] S2. Construction of a long-term stability database for shield tunnels: A database containing tunnel stability levels, structural characteristics, structural mechanical properties, rock mass properties, and other influencing factors was constructed on edge nodes with stable computing resources. Causal analysis was used to perform a series of processes, including feature selection, and the database was shuffled and divided into training, validation, and test sets.

[0234] S3. Construction of a hybrid-driven shield tunnel long-term stability evaluation model: Edge intelligence technology is used for model training and optimization. The cloud center sends the initially trained physical deep learning model and the extreme learning machine based on differential gray wolf optimization to an edge node, namely the aggregation server (AS). Other edge nodes participate in the model training of the AS and use local data to train local models. The physical deep learning model and the extreme learning machine based on differential gray wolf optimization are constructed and the hyperparameters are initialized. The training set is used as the input of the above model to train and optimize the model to obtain the optimal local model. The edge node sends the updated local model to the AS to obtain the updated global model.

[0235] S4. Prediction and evaluation of long-term stability of shield tunnels: The trained hybrid drive model is verified using a test set, and the predicted tunnel stability level is output.

[0236] Furthermore, in step S1, the tunnel structure stability grade evaluation standard adopts a four-level grading method, qualitatively classifying the stability of the tunnel structure into four grades, and quantitatively quantifying each grade and assigning corresponding quantification intervals according to the Technical Specifications for Highway Tunnel Maintenance (JTG H12-2015) and actual conditions.

[0237] Furthermore, in step S1, the method for establishing the tunnel structure stability grade evaluation standard is as follows:

[0238] S1.1 Qualitative description of tunnel structure stability: A four-level classification method is used to qualitatively divide the stability of tunnel structures into four levels. Level 1 is a safe state, based on the absence of damage or minor cracks on the tunnel surface; Level 2 is basically safe, based on the presence of minor cracks on the tunnel structure surface and no water seepage; Level 3 is a potential hazard, based on the presence of numerous cracks, damage, or minor water seepage on the tunnel surface; Level 4 is a dangerous state, based on the presence of extensive damage or rockfall that renders the tunnel unusable.

[0239] S1.2 Quantitatively describe the tunnel structure stability level: Based on the classification limits for the technical condition assessment of civil structures in the "Technical Specifications for Highway Tunnel Maintenance" (JTGH12-2015), and in light of actual conditions, the functional status of the civil structures is divided into four levels. A normalization method is used to quantify each level and assign it to a corresponding quantification interval. The quantification interval is used to distinguish the stability level of the tunnel structure.

[0240] Furthermore, in step S2, the long-term stability database of typical monitoring sections of shield tunnels includes text data information reflecting tunnel construction, operation information, rock mechanical properties, and groundwater storage conditions, mainly based on the shield tunnel stability grade, structural characteristics, structural mechanical properties, rock properties, and other influencing factors.

[0241] Furthermore, in step S2, the method for constructing the shield tunnel long-term stability database is as follows:

[0242] S2.1 Construct a long-term stability database for shield tunnels, including: Sub-database A consisting of quantified tunnel stability grades; Sub-database B consisting of tunnel structural characteristics consisting of initial support forms and secondary lining forms; Sub-database C consisting of tunnel structural mechanical characteristics consisting of initial support steel arch stress, secondary lining axial force and secondary lining bending moment; Sub-database D consisting of tunnel rock mass characteristics consisting of rock mass integrity coefficient, uniaxial compression, rock mass quality index, ultimate creep strain and rock mass bedding inclination; Sub-database E consisting of other influencing factors consisting of tunnel burial depth, lateral pressure coefficient, groundwater storage conditions and tunnel operation time; Pre-process the data to remove data segments containing invalid values and singular values; The reference basis for selecting the above indicators includes: 1) literature survey on factors affecting tunnel structural safety under the effect of soft rock creep, 2) analysis of a large amount of field monitoring data and numerical simulation data, and 3) case experience in tunnel structural safety analysis;

[0243] S2.2 Conduct causal analysis of text data information, using structural causal models to explore the causal relationship between different features, implement feature selection, verify the reliability of parameters selected based on literature surveys, numerical analysis, and expert experience, and understand the causal relationship between tunnel parameters and stability indicators;

[0244] S2.3 performs data standardization and data set division into training set, validation set and test set.

[0245] Furthermore, in step S2.2, the text data information causal analysis includes:

[0246] In conjunction with the long-term stability database of shield tunnels, the structural causal model, based on domain knowledge and data analysis, constructs a causal relationship diagram (directed acyclic graph, DAG) between features. This causal relationship diagram clarifies the causal relationship between features, identifies features that directly or indirectly affect the target variable, and evaluates the causal effect of each feature on the target variable through counterfactual reasoning. Conditional independence tests are used to remove irrelevant features. The structural causal model is used to select features at both the causal relationship and effect significance levels, verifying the reliability of parameters selected based on literature research, numerical analysis, and expert experience, and understanding the causal relationship between tunnel parameters and stability indicators.

[0247] Furthermore, in step S2.3, the data standardization processing and data set division include:

[0248] The data types of the shield tunnel long-term stability database are defined as discrete data and numerical continuous data. The discrete data are processed as one-hot, and the numerical continuous data are normalized to the maximum and minimum values. Sub-database A is defined as the output parameter, and sub-databases B, C, D, and E are defined as input parameters. The input and output parameters are shuffled and divided into training set, validation set, and test set in a ratio of 8:1:1.

[0249] Furthermore, in step S3, constructing a hybrid-driven shield tunnel long-term stability evaluation model includes:

[0250] S3.1 establishes a physical deep learning model, uses the training set as input to the physical deep learning model, and trains and optimizes the local model at the edge node;

[0251] S3.2 establishing an extreme learning machine based on differential grey wolf optimization, using the training set as input to the extreme learning machine model based on differential grey wolf optimization, and training and optimizing the local model at the edge node;

[0252] The S3.3 physical deep learning model and the extreme learning machine based on differential grey wolf optimization construct a physical-data-knowledge driven hybrid model for the long-term stability evaluation method of shield tunnels. The optimal local model obtained by the edge node is sent to the aggregation server to obtain the optimal hybrid-driven global model.

[0253] Furthermore, establishing a physical deep learning model includes: establishing the dynamic equations of the dynamic system, constructing a neural network f(t,x) based on physical information and a deep neural network that approximates the potential solution u(t,x) of the dynamic equation. At the same time, using automatic differentiation to process each term of the partial differential equation, calculating the partial derivatives of the potential solution with respect to the observation vector and time, so that the neural network f(t,x) based on physical information and the deep neural network that approximates the potential solution have the same network parameters but different activation functions; constructing a loss function to optimize the shared parameters of the two neural networks, and feeding the calculation results back to the optimizer to achieve parameter updates and form the optimal local model.

[0254] Building an extreme learning machine based on differential gray wolf optimization includes: initializing the weights and biases from the input layer to the hidden layer of the extreme learning machine and the population in the differential gray wolf optimization algorithm, setting hyperparameters such as the number of hidden layer neurons and activation function, using gray wolf optimization to simulate wolf hunting behavior to search for the optimal solution, optimizing the weights and biases of the ELM model, and enhancing the exploration ability through differential operations in the iterative process to avoid falling into local optimality, obtaining the optimal parameters of the extreme learning machine, and finally obtaining the optimal local model through alternating optimization and training.

[0255] Building a physical-data-knowledge-driven hybrid model includes: using edge intelligence technology to train and optimize the physical deep learning model and the extreme learning machine based on differential grey wolf optimization. The cloud center will send the initially trained hybrid drive model in its entirety to an edge node, namely the aggregation server (AS). Other edge nodes participate in the model training of the AS, use local data to train local models, and obtain the optimal local model. The edge node then sends the updated local model to the AS to obtain the optimal hybrid drive global model.

[0256] Furthermore, in step S4, the shield tunnel long-term stability prediction and evaluation includes: using the test set to verify the optimal hybrid drive model and outputting a predicted tunnel stability level.

[0257] A hybrid-drive-based shield tunnel long-term stability evaluation system is used to drive and execute the above-mentioned hybrid-drive-based shield tunnel long-term stability evaluation method. The evaluation system includes:

[0258] Data acquisition unit, used for tunnel information collection and cause-effect analysis, tunnel stability assessment index entry, and construction of relevant sub-databases;

[0259] The data preprocessing unit is used to perform operations such as data standardization and shuffle the data set into training, validation, and test sets;

[0260] A model building and training unit is used to establish a physical deep learning model and an extreme learning machine based on differential grey wolf optimization to construct a physical-data-knowledge driven hybrid model for shield tunnel stability grade prediction. The model is trained and hyperparameters are optimized using edge intelligence technology using the training set to obtain the optimal hybrid driving model.

[0261] The tunnel stability evaluation unit verifies the optimal hybrid drive model using the test set and outputs a predicted tunnel stability level.

Claims

1. A method for evaluating the long-term stability of a shield tunnel based on hybrid drive, characterized in that: The following steps are involved: S1. Obtain multiple sets of shield tunnel characteristics and corresponding tunnel stability levels, and establish a long-term stability database for shield tunnels; S2. Preprocess the data in the shield tunnel long-term stability database to remove data segments containing invalid values and singular values; S3. Based on the structural causal model, a causal relationship diagram is constructed between each feature and between each feature and the tunnel stability level. The causal relationship between each feature and the tunnel stability level is clarified, and the feature Q that affects the tunnel stability level is identified. S4. Based on counterfactual reasoning, we intervene in feature Q, observe the impact of the intervened features on the tunnel stability level, and use the conditional independence test to remove irrelevant features. S5. Standardize the data in the database and divide the processed data into training set, validation set and test set; S6. Use the training set to train the physical deep learning model and the extreme learning machine model based on differential grey wolf optimization, and use the validation set to verify the hyperparameters of the physical deep learning model and the extreme learning machine model, thereby obtaining a hybrid drive global model; S7, inputting the test set into the hybrid drive global model to obtain the predicted tunnel stability level; If the error between the evaluation index and the actual value is less than the preset threshold, the construction of the hybrid drive global model is completed and the process proceeds to step S8), otherwise it returns to step S1); S8. Obtain the characteristics of the shield tunnel to be evaluated and input them into the hybrid drive global model to obtain the tunnel stability level.

2. The method for evaluating the long-term stability of a shield tunnel based on hybrid drive according to claim 1, characterized in that: The shield tunnel long-term stability database includes sub-database A, sub-database B, sub-database C, sub-database D and sub-database E; The sub-database A includes several tunnel stability levels; the tunnel stability levels are divided into four evaluation levels: safe, basically safe, with hidden dangers and dangerous according to the degree of tunnel damage; The sub-database B includes several tunnel structural characteristics, including the initial support form and the secondary lining form; The sub-database C includes several tunnel structural mechanical properties, including initial support steel arch stress, secondary lining axial force and secondary lining bending moment; The sub-database D includes several tunnel rock mass properties, including rock mass integrity coefficient, uniaxial compression, rock mass quality index, ultimate creep strain and rock mass bedding inclination; The sub-database E includes several other influencing factors, including tunnel burial depth, lateral pressure coefficient, underground water storage conditions and tunnel operation time.

3. The method for evaluating the long-term stability of a shield tunnel based on hybrid drive according to claim 1, characterized in that: In step S3, a structural causal model is constructed based on the causal Bayesian network and the structural equation model; The causal relationship graph is a directed acyclic graph, in which the nodes represent the features in the structural causal model and the directed edges represent the direction of causal influence; Each node Y i The causal function of is as follows: Yes i =f i (Pa(Y i ),U i ) (1) Where: Pa(Y i ) represents feature Y i The parent node of . U i For feature Y i Related interference factors; f i is a function that describes the causal mechanism.

4. The method for evaluating the long-term stability of a shield tunnel based on hybrid drive according to claim 1, characterized in that: In step S4, the d-separation rule is applied to determine whether the feature node in the causal relationship graph is an irrelevant feature.

5. The method for evaluating the long-term stability of a shield tunnel based on hybrid drive according to claim 1, characterized in that: In step S5, the standardization process includes performing one-hot processing on discrete data and performing maximum and minimum value normalization processing on numerical continuous data.

6. The method for evaluating the long-term stability of a shield tunnel based on hybrid drive according to claim 1, characterized in that: In step S6, the method for establishing a physical deep learning model includes the following steps: S6.1.

1. Establish the dynamic equation of the dynamic system. The dynamic equation is as follows: Where: u(t,x) represents the potential solution, i.e., the tunnel stability evaluation level; t is time; x is the input vector; N(u) is the parameterized nonlinear operator; S6.1.

2. Based on the dynamics equations, construct a neural network based on physical information. The input layer of the neural network is the eigenvalues after removing irrelevant features in step S4. The hidden layer is a fully connected layer. The output layer is the potential solution. All layers use the Reluctant Lu function as the activation function. The neural network f(t,x) based on physical information is as follows: S6.1.

3. Based on the dynamics equation, construct a deep neural network that approximates the potential solution u(t,x) of the dynamics equation. The input layer of the deep neural network is the observation vector x and time t, the hidden layer is a fully connected layer, and the output layer is the numerical solution approximating u(t,x). All layers use the Swish function as the activation function. S6.1.

4. Use automatic differentiation to process the terms of the dynamical equations of the two neural networks, computing the partial derivatives of the potential solution with respect to the input vector and time, so that the physics-based neural network and the deep neural network approximating the potential solution have the same network parameters. S6.1.

5. Construct a loss function to optimize the shared parameters of the two neural networks and feed the calculation results back to the neural networks to update the parameters and form the optimal local model. The loss function MSE is as follows: MSE=MSE u +MSE f (4) Where: MSE u is the data error of the neural network; MSE f is the physical information error of the neural network; N u is the number of numerical solutions that approximate u(t,x); is the i-th numerical solution predicted by the neural network approximating u(t,x), i = 1, 2, ..., M u ; u i is the i-th potential solution; N f is the number of predicted potential solutions; is the i-th potential solution predicted by the neural network f(t,x), i=1,2,…,N f .

7. The method for evaluating the long-term stability of a shield tunnel based on hybrid drive according to claim 1, characterized in that: In step S6, the method for establishing an extreme learning machine model based on differential grey wolf optimization includes the following steps: S6.2.

1. Initialize the weights W from the input layer to the hidden layer of the extreme learning machine, the bias b, and the population parameters in the differential gray wolf optimization algorithm; S6.2.

2. Select the ReLU function as the activation function of the extreme learning machine, and set the number of neurons N in the hidden layer of the extreme learning machine and the maximum number of iterations T. max ; S6.2.

3. Use the training set to build an extreme learning machine model and calculate the average sum of squared errors of the sample data in the training set as the fitness of each gray wolf. The fitness calculation formula is: Where: Fitness i is the fitness of the i-th gray wolf in the gray wolf population; N is the total number of sample data in the training set; y k is the actual target value of the kth sample data; The predicted output result of the k-th sample data in the extreme learning machine model; S6.2.

4. Select the three wolves ranked in the top three by fitness, mark their individual positions, and use them as the decision layer of the extreme learning machine model. S6.2.

5. Determine whether the current number of iterations T is greater than the maximum number of iterations T max Is it true? If so, return the global optimal position as the optimal weight W from the input layer to the hidden layer of the extreme learning machine model * and bias b * Otherwise, update the position of each individual gray wolf in the gray wolf population, set T = T + 1, and return to step S6.2.3; S6.2.

6. Using weight W * and bias b * The extreme learning machine model is trained to output the deep feature representation H of the hidden layer for predicting the tunnel stability level and the weight β of the output layer to obtain the optimal local model; The function of deep feature representation H for tunnel stability grade prediction is as follows: H=f(W * X+b * ) (8) The calculation formula of the output layer weight β is as follows: β=(H T H) -1 H T Y (9) Where: H is the deep feature representation of tunnel stability grade prediction; X is the input feature matrix; H T is the transposed matrix of H; Y is the predicted tunnel stability grade.

8. The method for evaluating the long-term stability of a shield tunnel based on hybrid drive according to claim 1, characterized in that: In step S6), the method for establishing a hybrid drive global model includes the following steps: S6.3.

1. Based on edge intelligence technology, use the training set to train and optimize the physical deep learning model and the extreme learning machine based on differential grey wolf optimization; S6.3.

2. The cloud center sends the initially trained hybrid drive model to an edge node, i.e., the aggregation server. Other edge nodes participate in the model training of the aggregation server. S6.3.

3. Use the data set in the database to train the local model to obtain the optimal local model. The edge node sends the updated local model to the aggregation server to obtain the optimal hybrid drive global model.

9. A system using the shield tunnel long-term stability evaluation method according to any one of claims 1 to 8, characterized in that: It includes data acquisition unit, data preprocessing unit, model building and training unit and tunnel stability evaluation unit. The data acquisition unit is used to collect tunnel information and conduct causal analysis, enter tunnel stability assessment indicators, and build relevant sub-databases; The data preprocessing unit is used to perform data standardization operations and shuffle the data set into a training set, a validation set, and a test set; The model building and training unit is used to establish a physical deep learning model and an extreme learning machine based on differential grey wolf optimization, and is used to build a physical-data-knowledge driven hybrid model for shield tunnel stability grade prediction. Edge intelligence technology is used to perform model training and hyperparameter optimization using a training set to obtain the optimal hybrid driving model; The tunnel stability evaluation unit tests the optimal hybrid drive model using a test set and outputs a predicted tunnel stability level.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is run on an electronic device, the electronic device executes the hybrid drive-based long-term stability evaluation method for a shield tunnel according to any one of claims 1 to 8.

Images