Field-simulation-intelligence based method for evaluating the state of railway tunnel structure in operation

Abstract

The application discloses a field-simulation-intelligent-based operation railway tunnel structure state evaluation method and belongs to the technical field of tunnel engineering safety evaluation. The method comprises the following steps: acquiring tunnel initial parameters, establishing a tunnel model through three-dimensional numerical simulation, and verifying the model by using field monitoring data; periodically collecting field monitoring data, combining the verified three-dimensional numerical simulation data to construct a database, and training an artificial intelligence model; correcting the three-dimensional numerical simulation model by using the latest field monitoring data, acquiring the mechanical characteristic parameters of the whole tunnel, and predicting the corresponding deformation characteristic parameters through the trained artificial intelligence model; performing data level and feature level fusion on the whole parameter, calculating a comprehensive evaluation value; and determining the tunnel safety state and countermeasures. The method realizes accurate evaluation of the whole-time state of the operation railway tunnel structure through the deep fusion of field monitoring, three-dimensional numerical simulation and the artificial intelligence model, and provides support for maintenance decision-making.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering safety assessment technology, specifically a method for assessing the structural status of operating railway tunnels based on on-site simulation and intelligence. Background Technology

[0002] Railway tunnels are an important component of transportation infrastructure. With the increasing scale of railway tunnel construction and operation, if the tunnel structure is not properly maintained, its safety performance will gradually decrease due to external environmental factors and dynamic loads, such as vibration and aerodynamic loads. When the structural condition of a railway tunnel falls below a critical state, safety accidents such as lining collapse can occur at any time during tunnel operation, leading to a series of adverse consequences.

[0003] Accurate assessment of the safety status of tunnel structures is of significant guiding value for the development of maintenance plans and the reporting of maintenance projects during tunnel operation and maintenance. Traditional assessment methods have limitations:

[0004] 1. Current assessment methods based on field monitoring data or numerical simulation-monitoring fusion primarily assess tunnel structure monitoring points or the most unfavorable points, failing to achieve data acquisition and status assessment across the entire tunnel structure.

[0005] 2. Current evaluation methods lack deep integration between 3D numerical simulation and field monitoring data. In other words, field monitoring data does not effectively guide 3D numerical simulation analysis, nor does it enrich field monitoring data through 3D numerical simulation. The integration of 3D numerical simulation and field monitoring data is only superficial, often remaining at the level of data overlay.

[0006] 3. Data fusion methods mainly fall into three levels: data-level, feature-level, and decision-level fusion, resulting in a limited range of fusion approaches. Among these, data-level fusion involves a large computational load, while feature-level fusion tends to overlook some data. Summary of the Invention

[0007] In view of the above problems, this invention is proposed to provide a field-simulation-intelligence-based method for assessing the structural status of operating railway tunnels, which overcomes or at least partially solves the above problems. This method can achieve full-domain and full-time status assessment of tunnel structures, aiming to accurately evaluate the structural status of currently operating railway tunnels and propose corresponding treatment strategies based on the evaluation results, thereby ensuring the structural safety of railway tunnels and ultimately achieving full-life and long-life operation and maintenance of railway tunnels.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for assessing the structural condition of operational railway tunnels based on on-site simulation and intelligence, comprising the following steps: S1. Determining the initial state of the tunnel: Obtain the initial parameters of the tunnel structure and surrounding environment, establish a tunnel model through three-dimensional numerical simulation, and verify the model using on-site monitoring data to ensure that the deviation between the model parameters and the actual values ​​is less than a set threshold. S2. Database construction and artificial intelligence model training: During tunnel operation, on-site monitoring data is collected periodically and combined with the verified three-dimensional numerical simulation data to construct a database. Mechanical characteristic parameters are used as inputs and deformation characteristic parameters are used as outputs to train the artificial intelligence model. S3. Acquisition of state parameters at the evaluation time: At the evaluation time, the three-dimensional numerical simulation model is corrected using the latest field monitoring data to obtain the mechanical characteristic parameters of the entire tunnel, and the corresponding deformation characteristic parameters are predicted by the trained artificial intelligence model to achieve full coverage of state parameters of the tunnel. S4. State parameter hybrid fusion: Perform data-level fusion and feature-level fusion on the parameters of the entire domain to calculate the comprehensive evaluation value; S5. Status Assessment: Based on the comprehensive assessment value and the preset level table, determine the tunnel safety status and corresponding countermeasures.

[0009] In one embodiment, in step S1, the initial state of the tunnel includes the apparent state, the internal state, and the surrounding environment state. The apparent state is obtained through handover inspection, the internal state is obtained through engineering inspection, and the surrounding environment state is determined by combining geological survey and construction exposure.

[0010] In one embodiment, in step S2, the artificial intelligence model uses a fully connected neural network with the ReLU activation function, input parameters including stress, axial force, and bending moment, and output parameters including displacement and strain.

[0011] In one embodiment, step S4, the mixing and fusion includes: Data-level fusion: The weights of the measurement points are calculated using the analytic hierarchy process (AHP), factor analysis, or principal component analysis, and the fusion value is obtained by weighted summation. Feature-level fusion: Extract feature values ​​from deformation parameters including displacement and strain, extract peak values ​​from force parameters including stress, axial force and bending moment, and calculate a weighted comprehensive value.

[0012] In one embodiment, the formula for data-level fusion in step S4 is: (1) In the formula, y Data representing displacement, strain, stress, axial force, or bending moment; x iThis refers to the displacement or strain prediction value of the i-th monitoring point on the same cross section, as predicted by the artificial intelligence model, or the simulated value of stress, axial force, or bending moment calculated by the three-dimensional numerical model after field data correction. α i Let represent the weight coefficient of the i-th monitoring point, i=1,2,...,k, where k represents the number of monitoring points.

[0013] In one embodiment, the formula for feature-level fusion in step S4 is: (2) In the formula, z represents the comprehensive assessment value of the structural safety status of the cross section after feature-level fusion; y 1 ~y 5 represents the output of formula (1); y 1 indicates the displacement value represented by the cross section obtained after data-level fusion of displacement data; y 2 indicates the cross-sectional strain value obtained after data-level fusion of strain data; y 3 indicates the stress value represented by the cross section obtained after data-level fusion of stress data; y 4 indicates that the cross-section represents the axial force value after data-level fusion of axial force data; y 5 indicates the cross-sectional bending moment value obtained after data-level fusion of bending moment data; β 1 ，β 2 ，β 3 ，β 4 ，β 5 represents the weighting coefficients of the importance of the five characteristic parameters—displacement characteristics, strain characteristics, peak stress, peak axial force, and peak bending moment—within their respective categories. δ represents the deformation characteristics, i.e., the overall importance coefficient of displacement and strain in the final evaluation; γ represents the force characteristics, i.e., the overall importance coefficient of stress, axial force, and bending moment in the final evaluation.

[0014] In one embodiment, in step S5, a preset level table divides the safety status into levels I to V, corresponding to normalized comprehensive value ranges [0,0.2), [0.2,0.4), [0.4,0.6), [0.6,0.8), and [0.8,1.0], and associates them with accident risk descriptions and countermeasures.

[0015] In one embodiment, the frequency of the on-site monitoring data collection is set by hour, day, week or month, and the monitoring points are arranged at the arch top, left and right arch shoulders and left and right arch waist positions of the tunnel cross section.

[0016] In one embodiment, the three-dimensional numerical simulation employs the finite element method, and the model boundary conditions are set to zero displacement and stress.

[0017] In one embodiment, the training of the artificial intelligence model employs an Adadelta optimizer and a mean absolute error loss function.

[0018] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages: (1) In the prior art, on-site monitoring generally only has data from some typical measuring points, and cannot achieve full-area data acquisition. Even if image monitoring technology can achieve area data acquisition, it can only approximately achieve displacement or deformation data acquisition, and cannot obtain the stress state parameters of the tunnel structure. Therefore, the current technology has limitations or defects. However, the present invention can achieve synchronous acquisition of deformation and stress state parameters of the entire tunnel structure.

[0019] (2) Current assessment techniques have not yet deeply integrated field monitoring data with data obtained from three-dimensional numerical models. That is, the field monitoring data is not used to bind the three-dimensional numerical model, nor is the three-dimensional numerical model used to enrich the field monitoring data. The two types of data are simply superimposed using correlation coefficients, without achieving deep integration. This invention can achieve deep integration of field monitoring data and three-dimensional numerical simulation. That is, the three-dimensional numerical model is verified using field monitoring data, and then the field monitoring data is enriched using the verified three-dimensional numerical model, such as enriching the structural stress data. Finally, an artificial intelligence model is used to predict the parameters of areas outside the field monitoring data points, thereby achieving the acquisition of full-domain data of the tunnel structure.

[0020] (3) Current data fusion mainly involves data-level, feature-value, and decision-level fusion, and no hybrid fusion strategy has been proposed. To this end, this invention fuses feature data and peak data, and then proposes a hybrid fusion method to make up for the defects or deficiencies of single fusion. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a flowchart of the on-site-simulation-intelligence-based method for assessing the structural status of operating railway tunnels provided in this embodiment of the invention. Figure 2This is a flowchart of the on-site-simulation-intelligence-based operational railway tunnel structural condition assessment provided in this embodiment of the invention; Figure 3a This is a cross-sectional view of the measuring point layout for on-site monitoring provided in this embodiment of the invention; Figure 3b This is a longitudinal profile of the monitoring point layout provided in this embodiment of the invention. Figure 4 This is a structural diagram of a fully connected neural network model provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The terms used in this invention are explained as follows: On-site: is an abbreviation for on-site monitoring; Simulation: is an abbreviation for three-dimensional numerical simulation; Intelligence: is an abbreviation for artificial intelligence.

[0025] Tunnel structure: mainly refers to the tunnel lining structure, i.e., the secondary lining structure.

[0026] The entire area refers to all surface regions of the tunnel structure. Full time: refers to the time elapsed from the start of tunnel structure operation to the time of evaluation.

[0027] This invention discloses a method for assessing the structural condition of operational railway tunnels based on on-site simulation and intelligence. This method deeply integrates on-site monitoring data with a three-dimensional numerical simulation model using an artificial intelligence model. First, the parameters and settings of the three-dimensional numerical model are verified using on-site monitoring data. Then, the on-site monitoring data is enriched using the three-dimensional numerical model, thereby achieving a comprehensive and real-time condition assessment of the tunnel structure. A brief description of the specific process is provided below. Figure 1 As shown, it includes the following steps: S1. Determining the Initial State of the Tunnel: Obtain the initial parameters of the tunnel structure and surrounding environment, establish a tunnel model through three-dimensional numerical simulation, and verify the model using on-site monitoring data to ensure that the deviation between the model parameters and the actual values ​​is less than a set threshold. This step uses on-site monitoring data to verify and invert the three-dimensional numerical model, thereby obtaining accurate model parameters and settings.

[0028] S2. Database construction and artificial intelligence model training: During tunnel operation, on-site monitoring data is collected periodically and combined with the verified three-dimensional numerical simulation data to construct a database. Mechanical characteristic parameters are used as inputs and deformation characteristic parameters are used as outputs to train the artificial intelligence model. This step involves combining on-site monitoring data with data obtained from 3D numerical simulations to create a training dataset for the artificial intelligence model. The structural force parameters obtained from the 3D numerical simulations, such as stress, axial force, and bending moment, serve as input parameters to the intelligent model; the monitored structural deformation parameters, such as displacement and strain, serve as output parameters. The artificial intelligence model is then trained using this training dataset until it meets the requirements.

[0029] S3. Acquisition of state parameters at the evaluation time: At the evaluation time, the three-dimensional numerical simulation model is corrected using the latest field monitoring data to obtain the mechanical characteristic parameters of the entire tunnel, and the corresponding deformation characteristic parameters are predicted by the trained artificial intelligence model to achieve full coverage of state parameters of the tunnel. In this step, data is monitored for one cycle to obtain on-site monitoring data, such as displacement and strain. Then, the three-dimensional numerical model is corrected and inverted using displacement and strain. Finally, the structural force parameters of other evaluation areas outside the monitoring points, such as stress, axial force, and bending moment, are obtained through the three-dimensional numerical model.

[0030] S4. State parameter hybrid fusion: Perform data-level fusion and feature-level fusion on the parameters of the entire domain to calculate the comprehensive evaluation value; By using structural force parameters such as stress, axial force, and bending moment in other assessment areas outside the on-site monitoring points as input parameters for the artificial intelligence model, the corresponding output parameters, such as displacement and stress, can be obtained. This enables the acquisition of structural deformation data in other areas outside the on-site monitoring points, ultimately achieving the acquisition of state parameters for the entire tunnel structure.

[0031] S5. Status Assessment: Based on the comprehensive assessment value and the preset level table, determine the tunnel safety status and corresponding countermeasures.

[0032] By repeating steps S1 to S5, the real-time state parameters of the tunnel structure, i.e., the comprehensive evaluation value, can be obtained by monitoring data from typical measuring points on site. Then, the tunnel safety status and corresponding countermeasures can be determined by referring to the preset level table.

[0033] The following is combined with Figure 2 The following is a further explanation of each of the above steps: Step 1: Determine the initial state of the tunnel; The initial state of a tunnel encompasses two main categories: the tunnel structure and the surrounding environment. The initial state of the tunnel structure is further divided into its external and internal states. The external state of the tunnel structure is primarily obtained through post-construction and pre-operation handover inspections, such as detecting surface cracks and honeycomb-like defects. The internal state of the tunnel structure is mainly obtained through engineering inspections, such as detecting lining thickness, voids behind the lining, and voids between the initial support and the secondary lining; additionally, it includes internal defects, cavities, and porosity between the initial support and the secondary lining. The state of the surrounding environment is primarily obtained through geological surveys and construction exposure. For example, tunnel depth and topography are obtained through geological surveys, while the properties of the surrounding rock and adverse geological conditions are determined through construction exposure. The tunnel's cross-sectional dimensions and structural parameters can be determined through design documents, such as the tunnel's cross-sectional shape, lining thickness, and the materials and parameters used in the lining structure. To monitor tunnel operation, monitoring equipment is installed at typical locations during construction to monitor the tunnel structure's lateral displacement and strain. Typical locations on the cross-section include the arch crown, left and right arch shoulders, and arch waist—a total of five locations. The longitudinal profiles are mainly arranged in sections with varying cross sections and adverse geological conditions. To facilitate subsequent analysis, for example, the typical locations of the longitudinal profiles are selected at intervals of 10m, and the typical locations of the cross profiles are the arch crown, the left and right arch shoulders, and the left and right arch waists, for a total of five locations.

[0034] After determining the initial state parameters of the tunnel, the initial state of the tunnel was reproduced using three-dimensional numerical simulation technology. The model parameters of the numerical simulation were then verified using field monitoring data to ensure that the simulated parameters accurately reflect the actual situation of the project. This also lays the foundation for subsequent numerical analysis. The three-dimensional numerical simulation process uses the finite element method, which is briefly described below: ① Discretize the tunnel structure into a finite number of unit cells; ② Perform force analysis on each discretized unit; ③ The elements are connected by nodes, and the node displacements satisfy the deformation compatibility equation; ④ Solve the force equilibrium equations and nodal displacement compatibility equations for a finite number of units one by one; ⑤ Obtain the deformation and stress characteristics of each unit; ⑥ Based on the deformation and stress characteristics of a finite number of unit cells, the overall deformation and stress characteristics of the tunnel structure are analyzed.

[0035] After verifying the three-dimensional numerical model using on-site monitoring data, deformation and mechanical characteristic values ​​of typical longitudinal and transverse measuring points in the tunnel were obtained, and these characteristic values ​​were used as the initial state values ​​for tunnel operation. Among them, the deformation characteristic values ​​mainly include two categories: displacement and strain; the mechanical characteristic values ​​mainly include three categories: stress, axial force, and bending moment.

[0036] Step 2: Construction of tunnel operation database and training of artificial intelligence model; Starting from the year after the tunnel is completed and put into operation, if the tunnel has been in operation for h years, then the on-site monitoring data will be collected for h years. The data collection frequency during construction is m, and the unit can be hour, day, week, month, etc., so the number of data collections n = h / m. Furthermore, it can be seen that n data points were collected for displacement and strain at typical locations during on-site monitoring.

[0037] For each data acquisition, the tunnel's state at the time of acquisition is reproduced through three-dimensional numerical simulation, and verified using on-site monitoring data to obtain a numerical simulation that accurately reflects the monitoring moment. After obtaining the correct numerical model, mechanical characteristic values ​​such as stress, axial force, and bending moment at typical locations can be extracted through numerical simulation.

[0038] A database was constructed using deformation feature values ​​from typical monitoring locations in the field and mechanical feature values ​​obtained from 3D numerical simulation. Then, a fully connected network of an artificial intelligence model was selected as the training model, with the mechanical feature values ​​serving as the model's input parameters and the deformation feature values ​​as the model's output parameters. The ReLU function was chosen as the activation function in the model.

[0039] Step 3: Obtaining state parameters at the tunnel assessment time; Assume the tunnel assessment time is n+1 years. Then, at year n+1, deformation characteristic values ​​obtained from typical locations in the field are reproduced through three-dimensional numerical simulation, and verified using field monitoring data to obtain a numerical simulation that accurately reflects the monitoring time. After obtaining the correct numerical model, the mechanical characteristic values ​​of the entire area are then obtained. One measuring point is placed at the tunnel cross-section arch, and Q measuring points are placed on the left and right sides of the tunnel, for a total of K measuring points on the cross-section, where K=2. Q+1. If a monitoring section is set up at intervals T meters along the longitudinal profile of the tunnel, then a total of W monitoring sections are monitored, W = L / T (where L represents the tunnel length). Therefore, the total number of monitoring points acquired is K. W.

[0040] By inputting the mechanical characteristic values ​​into the aforementioned trained artificial intelligence model, the displacement and strain characteristic values ​​at the corresponding locations can be output. This allows for the acquisition of the deformation and mechanical characteristic values ​​across the entire tunnel at the moment of evaluation.

[0041] Step 4: Hybridization and fusion of state parameters at the tunnel assessment time; (1) Data-level fusion of deformation and mechanical characteristic values ​​of tunnel cross-section monitoring points. First, the weights of the tunnel cross-section monitoring points are calculated using methods such as analytic hierarchy process (AHP), factor analysis, or principal component analysis. The weights range from 0 to 1; the smaller the weight, the lower the importance of the monitoring point, and vice versa. The formula for calculating the characteristic values ​​after cross-section data-level fusion can be expressed as follows: (1) In the formula, y Data representing displacement, strain, stress, axial force, or bending moment; x i This refers to the displacement or strain prediction value of the i-th monitoring point on the same cross section, as predicted by the artificial intelligence model, or the simulated value of stress, axial force, or bending moment calculated by the three-dimensional numerical model after field data correction. α i Let represent the weight coefficient of the i-th monitoring point, i=1,2,...,k, where k represents the number of monitoring points.

[0042] The fused data y is divided into two main categories. For displacement and strain, relevant theoretical models, such as cloud models, fuzzy theory, and grayscale theory, are used to extract the corresponding feature parameters y1 and y2. For stress, axial force, and bending moment, peak parameters y3, y4, and y5 are directly extracted. Finally, the feature parameters and peak parameters are fused together, and the calculation formula is as follows: (2) In the formula, z represents the comprehensive assessment value of the structural safety status of the cross section after feature-level fusion; y 1 ~y 5 represents the output of formula (1); y 1 indicates the displacement value represented by the cross-section obtained after data-level fusion of displacement data; unit: millimeters, mm; y 2 indicates that the cross-sectional strain value obtained after data-level fusion of strain data is dimensionless; y 3 indicates the stress value of the cross section obtained after data-level fusion of stress data, in megapascals (MPa). y 4 indicates the axial force value represented by the cross section after data-level fusion of axial force data, in kilonewtons (kN). y 5 indicates the cross-sectional representative bending moment value obtained after data-level fusion of bending moment data, in units of kilonewton-meter (kN·m).

[0043] β 1 ，β 2 ，β 3 ，β 4 ，β5 represents the weighting coefficients of the importance of the five characteristic parameters—displacement characteristics, strain characteristics, peak stress, peak axial force, and peak bending moment—within their respective categories. δ represents the deformation characteristics, i.e., the overall importance coefficient of displacement and strain in the final evaluation; γ represents the force characteristics, i.e., the overall importance coefficient of stress, axial force, and bending moment in the final evaluation.

[0044] Formulas (1) and (2) are connected and progressive in a logical relationship, together forming the two core steps of hybrid integration.

[0045] Formula (1) is a data-level fusion: its purpose is to compress the data dimension. It weights and sums the same type of parameters of k measuring points on the same cross section of the tunnel at the same evaluation time, and finally outputs a fused single representative value (y) for the cross section.

[0046] Formula (2) is feature-level fusion: its purpose is to integrate multiple types of features to form the final comprehensive evaluation index. It takes the output result y of formula (1), i.e., the cross-sectional representative value, as input. Specifically: First, the displacement of the cross-section representative value for deformation type. y 1. Strain y 2. Further processing is performed to extract characteristic parameters that reflect the overall distribution characteristics, such as the expectation, entropy, and hyperentropy calculated using the cloud model.

[0047] At the same time, for the representative values ​​of the cross-sections subjected to stress, stress y 3. Axial force y 4. Bending moment y 5， Extract its "peak parameters" directly.

[0048] Finally, the extracted deformation characteristic parameters and peak stress parameters are assigned new weights according to their importance in the safety assessment. β 1- β 5) The overall contribution of deformation and stress (δ, γ) is distinguished and weighted again to calculate a single numerical index z that comprehensively reflects the safety status of the cross-section. z is a dimensionless value and is the direct basis for the final safety level determination.

[0049] Step 5: Status assessment at the time of tunnel assessment; (1) The tunnel structure status level is determined as shown in the table below.

[0050] Table 1. Tunnel Structural Status Levels and Countermeasures

[0051] After calculating the eigenvalue fusion value z, it is normalized, and then corresponding countermeasures are taken according to Table 1.

[0052] Let's take a railway tunnel on an operating line as an example to illustrate this further: Step 1: Determine the initial state of the tunnel; A railway tunnel is 500.0 m long and has a net cross-sectional area of ​​100.0 m². 2 The maximum burial depth is 650m. The initial support is C30 concrete with a shotcrete thickness of 0.35m. The secondary lining is a C35 reinforced concrete structure with a thickness of 0.5m. During construction from 2002 to 2026, displacement and strain sensors were installed at five locations on the tunnel cross-section: the arch crown, left and right arch shoulders, and left and right arch waists. A monitoring section was established every 100.0m along the tunnel longitudinal section, totaling six monitoring sections and 60 measuring points. Initial values ​​were collected from these sensors at these points, with a monitoring frequency of once a week. The measuring point layout is as follows... Figures 3a-3b As shown. Among them, Figure 3a The black circles in the middle represent displacement sensors, and the rectangles represent strain sensors. Figure 3b The entrance is on the left and the exit is on the right. The arch is 9.5m above the ground, the shoulder is 6.5m above the ground, and the waist is 3.0m above the ground.

[0053] After its completion in 2008, pre-operation inspections were conducted, primarily focusing on the tunnel lining surface. A longitudinal crack was discovered at the arch crown, 200m from the tunnel entrance, measuring 2.0m long, 0.25mm wide, and 0.3mm deep, without honeycomb or pitting. Further engineering inspections of the tunnel structure revealed a cavity at the left arch waist, 300m from the tunnel entrance, measuring 2.0m long, 3.0m wide, and 3.2m thick, without any signs of separation between the primary support and secondary lining.

[0054] Based on geological surveys and construction findings, it was discovered that the surrounding rock within 200m of the tunnel entrance is mainly composed of Class III granite, while the surrounding rock from 200m of the entrance to the exit is mainly composed of Class IV diorite. In the longitudinal middle of the tunnel, there is a fault fracture zone with a thickness of 25m, which is mainly composed of severely weathered diorite, has no water accumulation, and is not connected to the surface.

[0055] A three-dimensional numerical model was constructed by combining data from the tunnel's cross-sectional dimensions, structural parameters, surrounding rock properties, handover inspections, and engineering inspections. The model length was five times the tunnel's cross-sectional width to reduce the model's size effect. The tunnel structure was then discretized into a finite number of element units, and stress analysis was performed on each discretized element. Element units were connected by nodes, and node displacements satisfied the deformation compatibility equations. Finally, the force equilibrium equations and node displacement compatibility equations for each finite number of element units were solved individually, based on the boundary conditions (i.e., zero boundary displacements and stresses), thus obtaining the deformation and stress characteristics of each element unit.

[0056] Based on the calculation results of the three-dimensional numerical model, displacement and strain parameter values ​​for the on-site monitoring points are proposed. The calculation results of the three-dimensional numerical simulation are compared and analyzed with those of the on-site monitoring. If the deviation between the numerical simulation and the on-site monitoring is less than 5.0%, the parameters and settings of the three-dimensional numerical model are considered to accurately reflect the supporting project. If the deviation is greater than 5.0%, the model parameters are adjusted, such as reducing the cohesion and internal friction angle of the surrounding rock, until the deviation is less than 5.0%.

[0057] Step 2: Construction of tunnel operation database and training of artificial intelligence model; Based on the data collection frequency of the on-site monitoring equipment, data is collected weekly, and the specific operational details of the tunnel are recorded each time, such as train speed, vehicle weight, whether there are any new construction projects near the tunnel, and whether any special geological disasters have occurred. After five years of tunnel operation, data is collected weekly, and the tunnel operation status is recorded once a week. Assuming four collections per month and 48 collections per year, a total of 240 data points will be collected over five years to record the tunnel's operational status at the time of collection.

[0058] Based on the three-dimensional numerical model constructed in the first step, the operational status at each data collection moment during tunnel operation is added to the three-dimensional numerical model. For example, train weight is simulated by adding vertical force, and train speed is simulated by adding longitudinal wind speed. Then, the newly built three-dimensional numerical model is solved, and the calculation results of the three-dimensional numerical simulation are compared and analyzed with the field monitoring until the deviation between the numerical simulation and the field monitoring is less than 5.0%. This process is repeated until the three-dimensional numerical models for all 240 data points at the collection moments are calculated and meet the requirements.

[0059] For the locations of 60 field monitoring points, displacement and strain data from the field monitoring were combined with stress, axial force, and bending moment data from the same locations derived from 3D numerical simulations to construct a corresponding database. A fully connected neural network model was selected as the artificial intelligence model, with stress, axial force, and bending moment data as input parameters and displacement and strain data as output parameters. The ReLU activation function was chosen, with two hidden layers, the Adadelta algorithm as the optimizer, and the mean absolute error loss function as the loss function. The fully connected neural network model was then trained using the aforementioned database. The fully connected neural network model is as follows: Figure 4 As shown.

[0060] I. Algorithm Structure: 1. Network type and topology: This method employs a feedforward fully connected neural network, suitable for learning complex nonlinear mappings between input features and output targets. The network consists of an input layer, at least one hidden layer, and an output layer connected sequentially. Each neuron in each layer is connected to all neurons in the next layer, forming a fully connected network. Data is unidirectionally transferred from the input layer to the output layer, with no feedback or recurrent connections.

[0061] 2. Functions and configurations of each layer: 1) Input layer: Number of neurons: equal to the dimension of the input feature vector. The input features are mechanical parameters, namely stress, axial force, and bending moment. For a data sample, the input layer has 3 neurons, each receiving the stress value, axial force value, and bending moment value at a specific location.

[0062] Function: To receive raw or pre-processed mechanical characteristic data.

[0063] 2) Hidden layer: Number of layers and neurons: For example, 2 layers. The number of neurons in each layer is a key hyperparameter. It needs to be set according to the complexity of the problem during implementation. For example, a "gradual reduction" principle can be adopted: the number of neurons in the first hidden layer can be set to a multiple of the input dimension, such as 16 or 32; the number in the second layer is reduced accordingly, such as 8 or 16, aiming to gradually extract and compress features.

[0064] Core calculation: The output of each hidden layer neuron is given by the formula a = g(W The function is calculated as x + b). Here, x is the output vector of all neurons in the previous layer, W is the weight vector of this neuron, b is the bias term, and g is the activation function.

[0065] 3) Output layer: Number of neurons: equal to the number of predicted targets. Output targets are deformation parameters, i.e., displacement and strain. The output layer should have two neurons, one outputting the predicted displacement value and the other the predicted strain value.

[0066] Activation function: For regression tasks predicting continuous values, the output layer uses a linear activation function, g(z) = z, to ensure that the output value can be any real number. The hidden layer uses the ReLU function. ReLU(x) = max(0, x).

[0067] The ReLU function effectively alleviates the vanishing gradient problem in deep networks. It is computationally fast, introduces nonlinearity, and enables the network to fit complex mapping relationships. Compared to the Sigmoid or Tanh functions, its gradient is always 1 in the positive interval, which is beneficial for gradient propagation in deep networks.

[0068] II. Training Process: Training data sources and construction: The data comes from the second step, "Construction of Tunnel Operation Database." This database is a paired collection of on-site monitoring data and three-dimensional numerical simulation data.

[0069] Data pairing logic: For the same "data acquisition time" and the same "monitoring point location", the database contains: Input sample (X): The mechanical characteristic values ​​of this point calculated by the validated three-dimensional numerical model, including stress, axial force, and bending moment.

[0070] Output label (Y): Deformation characteristic values, displacement, and strain of this point actually collected by the field monitoring equipment.

[0071] Data scale: If the operation lasts for h years and the data collection frequency is once a week (m=1 week), then the total number of samples n = 52h Number of monitoring points. For example, with 60 monitoring points over 5 years, theoretically approximately 52 [measurement points] can be established. 5 60 = 15,600 training sample pairs.

[0072] Data preprocessing: Due to the significant differences in physical units and numerical magnitudes of the input features, including stress, axial force, and bending moment, directly inputting them into the network would lead to training instability. Output values, including displacement and strain, also require normalization. Therefore, both input and output data are standardized. For example, for each feature dimension, such as the stress values ​​of all samples, its mean (μ) and standard deviation (σ) are calculated, and then transformed as: x_normalized = (x - μ) / σ. This ensures that each feature dimension follows a standard normal distribution with a mean of 0 and a standard deviation of 1, accelerating model convergence.

[0073] Loss function design: Using Mean Absolute Error Loss: MAE Loss = (1 / N) Σ|y true - y pred | where N is the batch size. MAE loss measure predicted value y pred Compared with the true value y true The average of the absolute differences between the values. Compared to mean squared error, MAE is less sensitive to outliers, making the training process more robust and more directly reflecting the average error level of the prediction. Its physical meaning is also easier to interpret.

[0074] The optimization algorithm chosen is the Adadelta optimizer. It eliminates the need for manually setting the global learning rate; instead, it adaptively adjusts the learning rate for each parameter by considering the decaying average of the squared historical gradients. At its core, it maintains two state variables: the exponential moving average of the squared gradients E[g²] and the exponential moving average of the squared parameter updates E[Δx²].

[0075] III. Parameter Settings: 1. Network structure hyperparameters: Number of hidden layers: 2. For problems of moderate complexity, such as mapping from 3 mechanical features to 2 deformation features, 2 hidden layers are sufficient to capture the nonlinear relationship while avoiding overfitting and excessive computational cost.

[0076] Number of neurons per layer: Start with an empirical rule, for example, the number of neurons in the first hidden layer can be 4-8 times the input dimension (12-24), and the number in the second layer can be halved (6-12). Cross-validation can be used to search in grids such as [8, 16, 32] and [4, 8, 16] to select the combination with the minimum MAE on the validation set.

[0077] Activation function: Fixed at ReLU. Its sparse activation property helps the network learn more effective feature representations.

[0078] 2. Training hyperparameters: Optimizer parameters: Adadelta's rho is typically set to 0.95, and epsilon is set to 1e. -7 These are commonly used default values ​​that ensure good performance.

[0079] Batch size: Batch size affects training speed and gradient estimation noise. For datasets potentially containing tens of thousands of data points, a medium batch size, such as 32, 64, or 128, can be chosen. Larger batch sizes (e.g., 128) make training more stable, while smaller batch sizes (e.g., 32) may result in better generalization performance.

[0080] Training epochs: Early stopping is employed. The validation set loss is monitored, and training is stopped when it no longer decreases over several consecutive training epochs (e.g., 10 or 20 epochs). This effectively prevents overfitting and automatically determines the required number of training epochs.

[0081] Weight initialization: Use "He normal initialization" for fully connected layers. This method is designed to work with the ReLU activation function and can keep the variance of each layer's output stable in the early stages of training, thus accelerating convergence.

[0082] IV. Technical Effects: Model performance quantification metrics: Performance is reported on a separate test set, reserved from historical data, and not used in training and validation.

[0083] Key metric: Mean Absolute Error, which directly corresponds to the loss function and reflects the average deviation of the prediction.

[0084] Auxiliary indicators: Root mean square error and coefficient of determination R² can be added. The closer the R² value is to 1, the stronger the model's ability to explain data variation.

[0085] Example: "After testing, the trained AI model achieved a displacement prediction MAE of 0.15 mm and an R² of 0.92 on an independent test set; and a strain prediction MAE of 12 με and an R² of 0.88. This demonstrates that the model can accurately invert deformation parameters based on mechanical parameters." Step 3: Obtaining state parameters at the tunnel assessment time; Five years and one month after tunnel operation, the condition of the tunnel structure is assessed. Based on four on-site monitoring data points within the first month after five years, the operational status at the four data acquisition points is added to the three-dimensional numerical model constructed in the first step. The newly constructed three-dimensional numerical model is then solved, and the calculation results of the three-dimensional numerical simulation are compared and analyzed with the on-site monitoring data until the deviation between the numerical simulation and the on-site monitoring data is less than 5.0%.

[0086] To obtain full-area tunnel data, a measuring point was set up at 6m intervals along the tunnel cross-section, for a total of 9 numerical simulation measuring points. One measuring point was located at the tunnel crown, 4 on the left side of the tunnel cross-section, and 4 on the right side. A monitoring section was also set up at 5m intervals along the tunnel longitudinal section, for a total of 500 / 5+1=101 monitoring sections, totaling 9... 101 = 909 measuring points. The smaller the distance between the measuring points on the tunnel cross section and the longitudinal section, the more measuring points are required.

[0087] Based on the verification of the three-dimensional numerical model, stress, axial force and bending moment data values ​​were extracted from the above 909 measuring points. Then, the stress, axial force and bending moment data values ​​were input into the aforementioned fully connected network model of artificial intelligence for training. The model input parameters displacement and stress can be obtained. At this point, the deformation parameters including displacement and strain, and the force parameters including stress, axial force and bending moment of the 909 measuring points on the entire surface of the tunnel structure are all determined.

[0088] Step 4: Hybridization and fusion of state parameters at the tunnel assessment time; (1) The weight allocation calculation of the 9 measuring points of the tunnel cross section is performed using the analytic hierarchy process. The steps are as follows: The comparison judgment matrix P is constructed using the "1-9 scale method" as follows: (3) Maximum eigenvalue λ max =9.58, Consistency Index CI=(λ) max -n) / (n-1)=0.0725, consistency coefficient CR=CI / RI=0.05<0.1, satisfying the consistency condition. The weight distribution of the 9 measuring points on the tunnel cross section is as follows (selected counterclockwise from the right arch foot of the tunnel): [0.12, 0.102, 0.125, 0.203, 0.051, 0.069, 0.098, 0.058, 0.174]. Here, RI represents the average consistency index closely related to the order of the judgment matrix, and its values ​​are shown in Table 2; n represents the order of the matrix.

[0089] Table 2

[0090] Substitute the nine data values ​​of the tunnel cross section and their corresponding weights into formula (1) to calculate the value of any index of the tunnel cross section after data fusion. Thus, each cross section has only one deformation (strain and displacement) and force (stress, axial force and bending moment) characteristic value, and a total of 505 data fusion values ​​for 101 monitoring sections.

[0091] The five types of data mentioned above are divided into two main categories. For deformation data, which characterizes strain and displacement, the characteristic parameters are solved using cloud model theory, and then the weight allocation for strain and displacement is calculated using the analytic hierarchy process (AHP), yielding values ​​of 0.4 and 0.6, respectively. For force data, which characterizes stress, axial force, and bending moment, peak parameters are directly selected, and the corresponding weight allocation is calculated using the AHP, yielding values ​​of 0.3, 0.3, and 0.4, respectively. Furthermore, the weight allocation for both deformation and force data is set to 0.5.

[0092] The steps for solving feature parameters using a cloud model are briefly described below: (1) By xi Calculate the sample mean of this data set: First-order sample absolute central moments Sample variance: ; (2) From (1), we can obtain the expected value: ; (3) At the same time, the entropy can be obtained from the sample mean: ; (4) From the sample variance in (1) and the entropy in (3), we can obtain: .

[0093] Step 5: Status assessment at the time of tunnel assessment; Multiply the above weights by the solution characteristic parameters and peak parameters, and substitute them into formula (2) to solve for parameter z. After normalizing it, it can be seen that the maximum value of z for the 101 monitoring sections of the tunnel longitudinal section is 0.38 and the minimum value is 0.22, all of which are between 0.2 and 0.4. It can be seen that the overall status level of the tunnel structure is Level II, the risk accident is small and basically acceptable, and the countermeasures are to maintain the tunnel normally and strengthen monitoring.

[0094] The present invention provides a method for assessing the structural condition of operational railway tunnels based on on-site monitoring, simulation, and intelligence. This method overcomes the spatial limitations of traditional monitoring and enables the synchronous acquisition of condition parameters across the entire tunnel surface. Through the deep integration of on-site monitoring, 3D numerical simulation, and artificial intelligence models, it achieves accurate assessment of the overall, real-time condition of operational railway tunnel structures, providing support for maintenance decisions and reducing the risk of missed alarms.

[0095] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0096] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for assessing the structural condition of operational railway tunnels based on on-site simulation and intelligence, characterized in that: Includes the following steps: S1. Determining the initial state of the tunnel: Obtain the initial parameters of the tunnel structure and surrounding environment, establish a tunnel model through three-dimensional numerical simulation, and verify the model using on-site monitoring data to ensure that the deviation between the model parameters and the actual values ​​is less than a set threshold. S2. Database construction and artificial intelligence model training: During tunnel operation, on-site monitoring data is collected periodically and combined with the verified three-dimensional numerical simulation data to construct a database. Mechanical characteristic parameters are used as inputs and deformation characteristic parameters are used as outputs to train the artificial intelligence model. S3. Acquisition of state parameters at the evaluation time: At the evaluation time, the three-dimensional numerical simulation model is corrected using the latest field monitoring data to obtain the mechanical characteristic parameters of the entire tunnel, and the corresponding deformation characteristic parameters are predicted by the trained artificial intelligence model to achieve full coverage of state parameters of the tunnel. S4. State parameter hybrid fusion: Perform data-level fusion and feature-level fusion on the parameters of the entire domain to calculate the comprehensive evaluation value; S5. Status Assessment: Based on the comprehensive assessment value and the preset level table, determine the tunnel safety status and corresponding countermeasures.

2. The method as described in claim 1, characterized in that, In step S1, the initial state of the tunnel includes the apparent state, the internal state, and the surrounding environment state. The apparent state is obtained through handover inspection, the internal state is obtained through engineering inspection, and the surrounding environment state is determined by geological survey and construction exposure.

3. The method as described in claim 1, characterized in that, In step S2, the artificial intelligence model uses a fully connected neural network with the ReLU activation function. The input parameters include stress, axial force, and bending moment, and the output parameters include displacement and strain.

4. The method as described in claim 1, characterized in that, In step S4, the mixing and fusion includes: Data-level fusion: The weights of the measurement points are calculated using the analytic hierarchy process (AHP), factor analysis, or principal component analysis, and the fusion value is obtained by weighted summation. Feature-level fusion: Extract feature values ​​from deformation parameters including displacement and strain, extract peak values ​​from force parameters including stress, axial force and bending moment, and calculate a weighted comprehensive value.

5. The method as described in claim 4, characterized in that, In step S4, the formula for data-level fusion is: （1） In the formula, y Data representing displacement, strain, stress, axial force, or bending moment; x i This refers to the displacement or strain prediction value of the i-th monitoring point on the same cross section, as predicted by the artificial intelligence model, or the simulated value of stress, axial force, or bending moment calculated by the three-dimensional numerical model after field data correction. α i Let represent the weight coefficient of the i-th monitoring point, i=1,2,...,k, where k represents the number of monitoring points.

6. The method as described in claim 5, characterized in that, In step S4, the formula for feature-level fusion is: （2） In the formula, z represents the comprehensive assessment value of the structural safety status of the cross section after feature-level fusion; y 1 ~y 5 represents the output of formula (1); y 1 indicates the displacement value represented by the cross section obtained after data-level fusion of displacement data; y 2 indicates the cross-sectional strain value obtained after data-level fusion of strain data; y 3 indicates the stress value represented by the cross section obtained after data-level fusion of stress data; y 4 indicates that the cross-section represents the axial force value after data-level fusion of axial force data; y 5 indicates the cross-sectional bending moment value obtained after data-level fusion of bending moment data; β 1 ，β 2 ，β 3 ，β 4 ，β 5 represents the weighting coefficients of the importance of the five characteristic parameters—displacement characteristics, strain characteristics, peak stress, peak axial force, and peak bending moment—within their respective categories. δ represents the deformation characteristics, i.e., the overall importance coefficient of displacement and strain in the final evaluation; γ represents the force characteristics, i.e., the overall importance coefficient of stress, axial force, and bending moment in the final evaluation.

7. The method as described in claim 1, characterized in that, In step S5, a preset level table divides the safety status into levels I to V, corresponding to normalized comprehensive value ranges [0,0.2), [0.2,0.4), [0.4,0.6), [0.6,0.8), and [0.8,1.0], and associates them with accident risk descriptions and countermeasures.

8. The method as described in claim 1, characterized in that, The on-site monitoring data collection frequency is set by hour, day, week or month, and the monitoring points are arranged at the arch top, left and right arch shoulders and left and right arch waist of the tunnel cross section.

9. The method as described in claim 1, characterized in that, The three-dimensional numerical simulation uses the finite element method, and the model boundary conditions are set to zero displacement and stress.

10. The method as described in claim 1, characterized in that, The artificial intelligence model is trained using the Adadelta optimizer and the mean absolute error loss function.

Images