A system for calculating the critical depth of shallow-buried tunnels in loess soil

By combining DBN neural network with multi-parameter optimization to calculate the critical depth of tunnels, the problem of inaccurate calculation of critical depth of shallow loess tunnels in existing technologies has been solved, achieving a more accurate division of the boundary between shallow and deep tunnels, and improving construction safety and material utilization efficiency.

CN123116756APending Publication Date: 2026-11-17XI'AN UNIVERSITY OF ARCHITECTURE AND TECHNOLOGY +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411228186.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies lack precision in calculating the critical depth of shallow-buried tunnels in loess areas, making it difficult to effectively delineate the boundaries between shallow and deep burial, which affects construction safety and material utilization efficiency.

Method used

A deep belief network (DBN) neural network was used to construct a neural network model by combining parameters such as ground slope, support parameters, excavation method and surface load. The method for calculating the critical depth of the tunnel was optimized and the Xie Jiaxiao maximum formula was used for accurate calculation.

Benefits of technology

It improves the accuracy and applicability of critical depth calculation for shallow-buried tunnels in loess areas, provides more reasonable suggestions for the boundary between shallow and deep tunnels, and reduces construction risks and material waste.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116756A_ABST
    Figure CN123116756A_ABST
Patent Text Reader

Abstract

This invention discloses a system for calculating the critical depth of shallow-buried loess tunnels, relating to the field of tunnel construction. By employing a deep neural network (DBN), a critical depth optimization system for shallow-buried loess tunnels is established. This system includes a neural network model subsystem, a critical depth optimization subsystem, and a data storage subsystem. The critical depth optimization subsystem is used for human-computer interaction, data input and preprocessing, and accessing data from the storage system. The neural network model subsystem is used for the construction, training, parameter input, processing, and output of the DBN neural network model. The storage system stores the input and output parameters of the neural network model. The coordinated calculation of these multiple subsystems enables a reasonable division of the range of tunnel depth limits, improving the applicability of existing calculation methods for shallow-buried loess tunnels.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of tunnel construction technology, specifically relating to a system for calculating the critical depth of shallow-buried tunnels in loess. Background Technology

[0002] The distinction between deep and shallow tunnels can be divided into two main viewpoints. One is based on the theory of surrounding rock relaxation, which classifies deep and shallow tunnels according to whether a collapse arch can be formed. The other is based on the mechanics of continuous soil and rock media, which defines the boundary value between deep and shallow tunnels as the minimum depth at which the ground surface is not affected by the relaxation of the surrounding rock during the tunnel excavation process.

[0003] The stress characteristics of the support structure and the distribution of surrounding rock stress differ between deep-buried and shallow-buried tunnels. Appropriate excavation methods and support types can improve the safety and economic efficiency of tunnel construction, effectively utilizing the mechanical properties of the support structure while shortening the construction period and saving materials. Besides soil properties and tunnel cross-section, topography, the hydrogeological and engineering geological conditions of the overlying soil, and the disturbance of the surrounding rock during construction also affect the depth of the tunnel.

[0004] Therefore, questions arise regarding whether tunnel lining needs reinforcement, the appropriate length of reinforcement, and how to calculate the ground pressure on the reinforced lining section. There is a certain correlation between tunnel depth and surface cracks and settlement characteristics. Therefore, accurately understanding tunnel depth is crucial for construction safety and material conservation, as it allows for more effective control of surface cracks and settlement during tunnel construction.

[0005] Therefore, it is urgent to design a system for calculating the critical depth of shallow-buried tunnels in loess to solve the above problems, define a reasonable range of the boundary values ​​between shallow and deep tunnels, and improve the applicability of existing calculation methods to shallow-buried tunnels in loess. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a system for calculating the critical depth of shallow-buried loess tunnels. This system employs a deep neural network (DBN) to establish an optimization system for the critical depth of shallow-buried loess tunnels. The system includes a neural network model subsystem, a critical depth optimization subsystem, and a data storage subsystem. The critical depth optimization subsystem handles human-computer interaction, data input and preprocessing, and access to data from the storage system. The neural network model subsystem is used for constructing, training, inputting, processing, and outputting parameters of the DBN neural network model. The storage system stores the input and output parameters of the neural network model and performs post-processing on the database. The resulting optimized critical depth is more accurate than the Xiejiaxiao critical depth, improving the applicability of existing calculation methods for shallow-buried loess tunnels.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A system for calculating the critical depth of shallow-buried tunnels in loess soil is characterized by comprising: a neural network model subsystem, a critical depth optimization subsystem, and a data storage subsystem.

[0009] The neural network model subsystem uses the data storage subsystem as its database and the manually entered parameters processed by the critical depth optimization subsystem as its input parameters. It then obtains the final optimized critical depth result and returns it to the critical depth optimization subsystem and the data storage subsystem.

[0010] The neural network model subsystem includes a DBN neural network model construction module, a network model training module, and a DBN neural network application module. The network model training module trains the DBN neural network model built by the DBN neural network model construction module and transmits the trained model to the DBN neural network application module. The DBN neural network application module uses the trained DBN neural network model to calculate the critical depth optimization value of the tunnel parameters manually entered in real time. It then calculates the critical depth optimization value with the critical depth of the shallow tunnel calculated by Xie Jiaxiao's maximum formula to obtain the final optimized critical depth and output it.

[0011] The critical depth optimization subsystem includes a visualization interface display module, an input parameter data processing module, and a software interface conversion module. The input parameter data processing module is used to input tunnel parameters, process these parameters, and upload the processed data to the DBN neural network application module. Simultaneously, it uploads the critical depth results of shallow-buried tunnels to the visualization interface display module. The input tunnel parameters include: tunnel height h, tunnel span B, surrounding rock unit weight γ, and internal friction angle. Cohesion c, ground slope R, surface load P, excavation method W, support structure S;

[0012] The data storage subsystem is used to store data from the neural network model subsystem and the critical depth optimization subsystem, and to perform post-processing on the sample data in the database.

[0013] Preferably, the steps for constructing a DBN neural network include:

[0014] S1. Establish the energy function of the RBM stochastic neural network;

[0015] S2. Based on the energy function, the joint distribution probability of neurons in the explicit and hidden layers of the RBM stochastic neural network is solved;

[0016] S3. Based on the joint distribution probability, the edge probabilities of neurons in the visible and hidden layers are obtained respectively;

[0017] S4. Based on the edge probability values, obtain the state probability values ​​of neurons in the visible and hidden layers;

[0018] S5. Based on the state probability values ​​of neurons in the visible and hidden layers, the optimized RBM stochastic neural network is obtained;

[0019] S6. Based on the optimized RBM stochastic neural network, establish the DBN neural network model.

[0020] Preferably, the energy function expression for the RBM stochastic neural network is:

[0021]

[0022] In the formula, W is the connection weight matrix of neurons in each layer, v is the state vector of neurons in the visible layer, and v = (v1, v2, ..., v n h is the state vector of the hidden layer neurons, and h = (h1, h2, ..., h2). m ), where a is the bias vector of the visible layer neurons, and a = (a1, a2, ..., a2). n Let b be the bias vector of the hidden layer neurons, and b = (b1, b2, ..., bb2). m ), where n, m, i, and j are all constants.

[0023] Preferably, the formula for solving the joint distribution probability of neurons in the visible and hidden layers is:

[0024]

[0025] Where E(v,h) is the model energy function.

[0026] Preferably, the marginal probabilities of neurons in the visible and hidden layers are calculated as follows:

[0027]

[0028]

[0029] in, The integral of the hidden layer neurons in the joint probability. This refers to the integral of the joint probability (Gibbs distribution) of neurons in the visible layer.

[0030] Preferably, the formulas for calculating the probability values ​​of visible and hidden layers are as follows:

[0031]

[0032] Wherein, P(v j =1|h) represents the data features received by the hidden layer neurons; P(h) j =1|v) represents the data characteristics received by the visible layer neurons; ω ij These are the connection weights of neurons in each layer.

[0033] Preferably, the DBN neural network constructed through training optimizes the connection parameters of the hidden and visible neurons based on the error between the input data and the mapping data, so that the extracted data information better matches the characteristics of the original data.

[0034] Preferably, the step of optimizing the connection parameters of hidden and visible layer neurons includes:

[0035] S1. Find the maximum value of the logarithm of the likelihood function;

[0036] S2. Based on the maximum value of the logarithm of the likelihood function, combined with gradient iteration of the optimization algorithm, the optimal connection parameters are solved.

[0037] Preferably, the formula for finding the maximum value of the logarithm of the likelihood function is:

[0038]

[0039] Where, θ * The maximum value of the logarithm of the likelihood function; L s (θ) is the maximum likelihood function, P(x) t |θ) is the joint probability of a sample and its corresponding code;

[0040] Optimal connection parameter θ i+1 The solution is:

[0041]

[0042] Where, θ i These are model data connection parameters. It is the connection parameter θ i The updated gradient.

[0043] Preferably, the data post-processing in the data storage subsystem includes: anomaly data processing and sensitivity analysis;

[0044] Among them, anomaly processing uses box plots to clean up abnormal data and optimize the database; sensitivity analysis can intuitively show the degree to which the critical depth is affected by different factors.

[0045] The beneficial effects of this invention are: This invention discloses a system for calculating the critical depth of shallow-buried tunnels in loess. Compared with the prior art, the improvement of this invention lies in:

[0046] Compared with existing technologies, the critical depth optimization system for shallow buried tunnels in loess, as described in this invention, is based on a constructed and trained DBN neural network. It considers four parameters: ground slope R, support parameters S, excavation method W, and surface load P. This system calculates an optimized critical depth value for the shallow buried tunnel. This optimized value is then compared with the critical depth calculated using the Xie Jiaxiao maximum formula to arrive at the final optimized critical depth. This final optimized critical depth is more accurate than the Xie Jiaxiao critical depth. Furthermore, the system stores the input and output data from each calculation to supplement the system's database, and optimizes the database through post-processing to further improve the computational performance of the neural network. Attached Figure Description

[0047] Figure 1 This invention relates to a critical depth optimization system for shallow buried tunnels in loess.

[0048] Figure 2 This is a basic structure diagram of the DBN of this invention;

[0049] Figure 3 A box plot for handling outlier values ​​of variables is entered for this invention;

[0050] Figure 4 This is the principle behind the construction of the box-type diagram of the present invention;

[0051] Figure 5 This is a basic structural diagram of the RBM in this invention;

[0052] Figure 6 This is a flowchart of the critical depth optimization method for shallow buried tunnels in loess.

[0053] Figure 7 This is a simulation diagram of the system of the present invention;

[0054] Figure 8 This is a schematic diagram illustrating the damage to the shallow-buried tunnel of the present invention;

[0055] Figure 9 This is a diagram illustrating the relationship between the tunnel span and the optimized critical depth value of this invention.

[0056] Figure 10 The input interface for the critical depth optimization subsystem;

[0057] Figure 11 The output interface of the critical depth optimization subsystem is optimized. Detailed Implementation

[0058] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0059] Example 1:

[0060] See attached document Figure 1 The critical depth optimization system for shallow buried tunnels in loess shown includes: a neural network model subsystem, a critical depth optimization subsystem, and a data storage subsystem;

[0061] The neural network model subsystem uses the data storage subsystem as its database and the manually entered parameters processed in the critical depth optimization subsystem as its input parameters. It then obtains the critical depth optimization results and returns them to the critical depth optimization subsystem.

[0062] The critical depth optimization subsystem is used for human-computer interaction, allowing manual input of tunnel-related parameters (tunnel height h, tunnel span B, surrounding rock unit weight γ, internal friction angle) in real time. The critical depth optimization subsystem takes the input parameters (cohesion c, ground slope R, surface load P, excavation method W, support structure S) as input, performs data standard processing, uploads the processed data to the neural network model subsystem and the data storage subsystem, displays the critical depth result of the shallow buried tunnel calculated by the neural network model subsystem, and calls the data storage subsystem to store the returned result.

[0063] The data storage subsystem is used to store the input data of the critical depth optimization subsystem and the result data obtained by the neural network model subsystem, and to perform post-processing on the sample data in the database.

[0064] The neural network model subsystem includes: a DBN neural network model construction module, a network model training system module, and a DBN neural network application module. The DBN neural network model construction module is used to build the DBN neural network model, and the network model training system is used to train the built DBN neural network model. The training dataset comes from the database of the data storage subsystem. The DBN neural network application module uses the trained DBN neural network model to calculate the critical depth optimization value of the tunnel parameters manually entered in real time. It then calculates the critical depth optimization value with the critical depth of shallow tunnels calculated by Xie Jiaxiao's maximum formula to obtain the final optimized critical depth, which is then output to the critical depth optimization subsystem for display.

[0065] Furthermore, this embodiment also provides the theoretical basis for introducing critical depth optimization parameters (R, P, W, S);

[0066] The critical depth obtained by Xie Jiaxiao's maximum formula is a range, but when the tunnel cross-section is large or there are special geological conditions, the range of critical depth values ​​is too wide to be reliable. Therefore, to make the range of critical depth values ​​for shallow tunnels more accurate and improve the applicability of the formula, this invention introduces four parameters—ground slope R, surface load P, excavation method W, and support structure S—through a neural network and combines them with Xie Jiaxiao's maximum method to calculate the final optimized critical depth of shallow tunnels, thus obtaining a more accurate critical depth for shallow tunnels. Practical application has shown that the system of this invention provides reasonable results for classifying loess tunnels into deep and shallow burial depths, with the actual fracture angle close to the theoretical fracture angle. It also provides a more accurate suggestion for the tunnel depth corresponding to the point of maximum loosening pressure in the surrounding rock as the boundary between deep and shallow burial depths. Figure 8 This is a schematic diagram illustrating the damage to the shallow-buried tunnel of the present invention;

[0067] (1) Excavation method W

[0068] Excavation methods for shallow-buried tunnels in loess include shield tunneling and drill-and-blast methods. Different excavation methods will disturb the surrounding rock to varying degrees, causing groundwater loss and resulting in reconsolidation and settlement of the surrounding rock. When considering the critical depth of shallow-buried tunnels, the disturbance to the strata during construction cannot be ignored.

[0069] (2) Ground slope R

[0070] Ground slope affects tunnel bias and the critical location of tunnel entrances and exits. For example, shallow tunnels located near mountains and water will directly affect the stress form of the support structure and the economy and safety of tunnel construction. In this case, ground slope has a significant impact on the critical depth of shallow tunnels with bias.

[0071] (3) Surface load P

[0072] The influence of surface load on critical depth is mainly due to the presence of structures on the surface above shallow-buried tunnels. For example, subway tunnels built in cities are generally shallow-buried, and their construction is affected by surface urban structures. From the perspective of their mutual influence, the impact of structures on tunnel construction manifests as increased settlement of the tunnel arch and increased horizontal convergence of the arch waist; expansion of the plastic zone between the foundation of the structures and the tunnel; and exacerbation of uneven surface settlement.

[0073] (4) Support structure S

[0074] The influence of the support structure on the critical depth is mainly manifested in the construction time, stiffness, and degree of contact with the surrounding rock. The initial support structure of composite lining generally allows for a certain degree of deformation of the surrounding rock (New Austrian Tunneling Method) to fully utilize the self-supporting capacity of the surrounding rock. The surrounding rock pressure consists of deformation pressure and loosening pressure; therefore, the construction time, stiffness, and degree of contact with the surrounding rock of the support structure directly affect the deformation and loosening of the surrounding rock.

[0075] Furthermore, the method for obtaining the critical depth optimization parameters (R, P, W, S) is given:

[0076] 1. Excavation method W

[0077] Tunnel excavation methods are mainly divided into three categories: drill-and-blast method, shield tunneling method, and tunnel boring machine (TBM) method. The drill-and-blast method, also known as the mining method, involves excavating the tunnel cross-section through drilling and blasting operations. This method is highly adaptable to various geological conditions, has low excavation costs, and is particularly suitable for constructing tunnels in hard rock. Current drill-and-blast construction techniques include the New Austrian Tunneling Method (NATM) and the New Italian Method (NITM). Depending on the method of tunnel cross-section excavation, drill-and-blast excavation can be further divided into:

[0078] ① Full-face excavation method;

[0079] ② Step method: including a. Long step method: L>5B; b. Short step method: (1~1.5)B≤L≤5B; c. Ultra-short step method: step length 3~5m;

[0080] ③ Sectional excavation method: including a. core soil reserved method; b. single-side wall pilot tunnel method; c. double-side wall pilot tunnel method; d. central pilot tunnel method; e. central diaphragm wall method (CD method); f. cross central diaphragm wall method (CRD method), etc.

[0081] Full-face excavation is a single-stage method with fast construction progress and minimal disturbance to the surrounding rock. The bench method is currently the most widely applicable construction method; the support closure time and bench length can be adjusted according to changes in surrounding rock conditions and the efficiency of construction machinery. Sectional excavation is suitable for soil or easily collapsible soft surrounding rock. For tunnels with high requirements for surface subsidence, sectional excavation should be given priority. The disadvantages of sectional excavation are slow progress, multiple sections, poor ventilation, and poor economic efficiency.

[0082] Tunnel boring machines (TBMs) are suitable for soil tunnels, while tunneling machines (TBMs) are suitable for rock tunnels. Higher water content and greater burial depth result in greater disturbance to the surrounding rock during TBM excavation. The disturbance to the surrounding rock varies depending on the TBM construction parameters (total thrust, jacking speed, cutterhead torque, screw conveyor torque) and the type of TBM. TBMs offer significant advantages over drill-and-blast methods in terms of speed and minimal disturbance to the surrounding rock, but they are not suitable for tunnels in adverse geological conditions. Excavation and lining cannot be carried out simultaneously; secondary lining must be performed after the TBM excavation is completed, resulting in an excessively long construction period.

[0083] Therefore, from the perspective of the disturbance to the surrounding rock during construction, the excavation method W is obtained using an expert scoring method. The value of W ranges from (0, 100). The reference value range for the full-face excavation method and the TBM shield tunneling method is (0-50), the reference value range for the bench method is (25-75), and the reference value range for the sectional excavation method is (75-100). The larger the value, the greater the disturbance to the surrounding rock caused by the construction method.

[0084] 2. Ground slope R

[0085] Ground slope directly affects the bias pressure of a tunnel. Uneven stress in the tunnel structure can lead to stress concentration and excessive deformation. Analyzing bias pressure based on ground slope involves directly converting the angle between the ground and the horizontal plane into a numerical value, which is the value of the ground slope R.

[0086] 3. Surface load P

[0087] The spatial relationship between tunnels and structures involves distance and orientation. Orientation-wise, tunnels can be categorized as oblique lateral crossings, oblique underpasses, orthogonal underpasses, and parallel lateral crossings. To simplify calculations, the loads on surface structures or load-bearing objects can be simplified to concentrated loads using the following simplification method:

[0088]

[0089] In the formula, P is the load affecting the tunnel; P' is the actual concentrated load; and L is the closest distance between the load and the tunnel structure.

[0090] 4. Support structure S

[0091] Tunnel support linings can be mainly classified into integral linings, composite linings, anchor-sprayed linings, and prefabricated linings. Integral linings are suitable for shallow buried sections at tunnel entrances, while prefabricated linings are used for tunnels excavated using the shield method. The initial support structure of composite linings uses anchor-sprayed linings, and the secondary lining uses cast-in-place concrete or shotcrete as the secondary lining structure. The consideration of the support structure here involves two aspects: (1) the stiffness of the support structure is reflected by the final deformation; (2) the proportion of vertical loosening pressure to the surrounding rock pressure (deformation pressure + loosening pressure + impact pressure + expansion pressure), that is, the support structure S includes Δ and Ψ. Determining the support structure S requires determining the values ​​of Δ and Ψ; among which, the formulas for calculating the values ​​of Δ and Ψ are:

[0092] (1) Ignoring the directionality of the deformation of the support structure, the final deformation Δ

[0093]

[0094] In the formula, Δ i It represents the displacement of different parts of a certain cross section (or multiple cross sections of a certain segment) of a tunnel, including the crown, shoulder, waist, foot, and bottom of the arch, where the unit of Δ is mm.

[0095] (2) The proportion of vertical loosening pressure to surrounding rock pressure Ψ

[0096]

[0097] In the formula, q sThe vertical pressure q for shallow-buried tunnels is determined according to the "Railway Tunnel Design Code" (TB10003). t It is the vertical pressure of the surrounding rock acting on the support structure as measured on site;

[0098] Specifically, the DBN neural network model construction process is as follows:

[0099] DBN Basic Structure

[0100] A typical deep belief network (DBN) model consists of multiple restricted Boltzmann machines (RBMs) "cascaded" together, such as... Figure 2 As shown, during the transmission process from the lower layer to the upper layer, the hidden layer of the lower RBM is also the visible layer of the upper RBM. When the sample data is input into the visible layer of the first RBM, the hidden layer of the RBM that receives the data becomes the input layer of the next RBM, thus forming a learning network framework containing multiple RBMs.

[0101] The DBN neural network model uses a Restricted Boltzmann Machine (RBM stochastic neural network) as its learning algorithm, following the maximum probability principle to generate the joint distribution form of the model. The DBN neural network model consists of three parts: an input layer, hidden layers, and an output layer. The hidden layers are composed of multiple RBM stochastic neural networks "cascaded" together, such as... Figure 2 As shown.

[0102] Furthermore, any RBM stochastic neural network consists of two parts: a visible layer and a hidden layer. The visible layer is the input layer for data samples of the stochastic neural network, and the hidden layer is the data feature extraction layer, which can be one or more layers. There are no connections between neurons in the hidden and visible layers. The neurons in each layer of the hidden layer are interconnected, and the connections are non-directional, that is, the connections between neurons in the layers are bidirectional, and data features can be transmitted in both forward and backward directions.

[0103] The basic structure of an RBM is as follows: Figure 5 As shown in the figure, the state vector of the visible layer neurons is represented as v = (v1, v2, ..., v...). n The hidden layer neuron state vector is represented as h = (h1, h2, ..., h...). m The bias vector of the visible layer neurons is represented as a = (a1, a2, ..., a...). n The bias vector of the hidden layer neurons is represented as b = (b1, b2, ..., bb2). m );W n×m The connection weight matrix for each layer of neurons;

[0104] S1. Establish the energy function of the RBM stochastic neural network;

[0105] Since neurons in an RBM stochastic neural network have only two states—activated and inactive—and each state is interpreted randomly using probabilistic methods (represented by "1" for activated states and "0" for inactive states), the probability expression between two connected neurons follows a Bernoulli distribution. Therefore, the energy function of an RBM stochastic neural network can be expressed as:

[0106]

[0107] In the formula: W is the connection weight matrix of neurons in each layer, w ij Let n be the value in the i-th row and j-th column of the weight matrix W, where n, m, i, and j are all constants.

[0108] S2. Based on the energy function, the joint distribution probability of neurons in the explicit and hidden layers of the RBM stochastic neural network is solved;

[0109] The energy value of the RBM random neural network is the main parameter to consider for the error of mutual conversion between the data features of the visible layer and the hidden layer. The joint distribution probability of the visible layer neurons and the hidden layer neurons can be solved by the energy function, as shown in Equation (2).

[0110]

[0111] Where E(v,h) is the model energy function.

[0112] S3. Based on the joint distribution probability, the edge probabilities of neurons in the visible and hidden layers are obtained respectively;

[0113] The transmission of data features from hidden and visible neurons in an RBM stochastic neural network is represented by a probability distribution, among which the marginal probability is more representative. Therefore, the marginal probability distribution expressions for visible and hidden neurons can be obtained from equation (2):

[0114]

[0115] in, The integral of the hidden layer neurons in the joint probability. The integral of the joint probability (ReLU distribution) of neurons in the visible layer;

[0116] S4. Based on the edge probability values, obtain the state probability values ​​of neurons in the visible and hidden layers;

[0117] In an RBM (Randomized Benchmarking Network), neuron inputs and outputs are transmitted in binary form (0 for sleep, 1 for activation). Therefore, given the known states of neurons in the visible (hidden) layers, the characteristics of the data received by neurons in the hidden (visible) layers can be represented by a conditional probability distribution, i.e., P(v...). j =1|h) or P(h) j=1|v), using ReLU, the most commonly used activation function in DBN neural networks, to represent the state probability values ​​of neurons in the visible and hidden layers:

[0118]

[0119] Where, ω ij These are the connection weights of neurons in each layer.

[0120] S5. Based on the probability values ​​of neurons in the visible and hidden layers, the optimized RBM stochastic neural network is obtained;

[0121] S6. Based on the optimized RBM stochastic neural network, establish the DBN neural network model.

[0122] RBM stochastic neural network training:

[0123] In the RBM stochastic neural network, each neuron in the visible layer maps the data features to the hidden layer neurons according to Equation (5), and then maps the received data from the hidden layer neurons to the visible layer neurons according to Equation (6). Therefore, the training of the RBM stochastic neural network is based on the error between the input data and the mapped data, and optimizes the connection parameters of the hidden and visible layer neurons so that the extracted data information is more consistent with the original data features. At this time, the data obtained by the hidden layer is considered to have better representativeness.

[0124] The above analysis shows that by repeatedly adjusting the connection parameters of the visible and hidden neurons to minimize the difference between the original input data and the reconstructed data, i.e., by optimizing the connection parameters, the training of the RBM random neural network can be achieved; for example... Figure 6 As shown,

[0125] S1. Find the maximum value of the logarithm of the likelihood function;

[0126] Minimizing the difference between two sets of data (i.e., the ReLU distribution represented by the RBM stochastic neural network and the distribution of the input samples) means that the likelihood function... Maximize, however, directly solve the maximum likelihood function The process is quite complex, so it is obtained indirectly by solving for the maximum value of the logarithm of the likelihood function, that is:

[0127]

[0128] Where, θ * The maximum value of the logarithm of the likelihood function; L s (θ) is the maximum likelihood function, P(x) t |θ) is the joint probability of a sample and its corresponding code, i.e., the probability density function of the ReLU distribution of the RBM network.

[0129] S2. Based on the maximum value of the logarithm of the likelihood function, combined with gradient iteration of the optimization algorithm, the optimal connection parameters are solved. Specifically, through θ * Worth going to L s (θ), and using L s (θ) further yields the optimal parameter θ i+1 ;

[0130] The optimal parameters can be further solved through gradient iteration of an optimization algorithm:

[0131]

[0132] Where, θ i+1 The updated connection parameters, i.e., the optimal parameters, θ i These are the connection parameters before the update. It is the connection parameter θ i The updated gradient is θ = (W, a, b);

[0133] The training performance and accuracy of the DBN model are closely related to the setting of model parameters. For parameter setting, parameters are first initially determined based on empirical methods, and then controlled variables are analyzed using experimental methods. The empirically determined model parameters are as follows: the number of hidden layer nodes is set to 10, the learning rate is set to 0.001, the number of RBM training iterations and the number of BP backpropagation iterations are set to 500 and 100, respectively.

[0134] Comparative evaluation criteria were set for each parameter, using mean absolute error (MAE), mean relative error (MRE), mean squared error (MSE), and coefficient of determination R. 2 Mean absolute error is insensitive to outliers and better reflects the actual situation of prediction errors; mean relative error reflects the magnitude of relative errors; mean squared error is sensitive to outliers and reflects the distribution of prediction errors; the coefficient of determination is an indicator used to evaluate the goodness of fit of a regression model.

[0135]

[0136] In the formula, y i It is the true value, y′ i It is a predicted value. It is the average of all true values, and N is the number of samples.

[0137] The optimized critical depth is calculated using the trained RBM stochastic neural network. This critical depth is then compared with the critical depth derived from Xie Jiaxiao's maximum formula to obtain the final optimized critical depth. This process is performed within the DBN neural network application module. The calculation formula is as follows:

[0138] H′ L =H L +ΔH L (13)

[0139] In the formula, H′ L It refers to the optimized critical depth, that is, the final optimized critical depth of shallow-buried tunnels; H L The critical depth of shallow-buried tunnels is calculated using Xie Jiaxiao's maximum formula; ΔH L It is the critical depth optimization value obtained by neural network calculation of the four introduced parameters (R, W, P, S).

[0140] Among them, the critical depth H obtained using Xie Jiaxiao's maximum formula is... L As per existing technology, the calculation formula is:

[0141]

[0142] In the formula, B is the tunnel excavation height; h is the tunnel excavation height; λ is the lateral pressure coefficient of the surrounding rock; Θ is the friction angle between the two sides of the soil column at the top of the tunnel;

[0143] The following data need to be analyzed during model training:

[0144] (1) Analysis of the number of input and output nodes

[0145] The number of input and output nodes depends on the input and output parameters; the number of input nodes is 4, and the number of output nodes is 1.

[0146] (2) Hidden Layer Number Analysis

[0147] During the analysis of the number of hidden layers, other parameters were kept constant during training, and only the number of hidden layers was adjusted. The number of hidden layers was set to 1 to 10, with a total of 10 sets of data.

[0148] (3) Hidden Layer Node Count Analysis

[0149] If the number of hidden layers is inappropriate, the RBM's ability to extract data features will be either too low or too high, resulting in a decrease in the training effect of the neural network. During the RBM training process, the number of hidden layers was set to 20 groups of 1-15, 20, 30, 40, 50, and 60, respectively, for comparative analysis.

[0150] (4) Batch size analysis

[0151] The block size will affect the computational cost and time of the model. It should be adjusted according to the total number of training samples. For example, if the total number of training samples is 300, the block size can be set to 1, 2, 5, 10, 15, 30, 50, 60, 100, 150, or 300.

[0152] (5) Learning rate analysis

[0153] The learning rate directly affects the iteration speed of RBM: if the learning rate is too large, the iteration process will be unstable and the model will not converge; if the learning rate is too small, the model may get stuck in local optimization. During RBM training, the learning rates were set to 0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, and 0.5 respectively.

[0154] (6) Analysis of other parameters

[0155] The initial weights are based on a normal distribution N(0,0.01); the initial bias vector is set to 0; and the momentum parameter is set to 0 to 1.

[0156] Furthermore, the critical depth optimization subsystem includes: a visualization interface display module, an input parameter data processing module, and a software interface conversion module. The input parameter data processing module is used to input tunnel parameters (ground slope R, surface load P, excavation method W, support structure S), preprocess the input tunnel parameters, and upload the processed data to the DBN neural network application module of the neural network model subsystem for calculating the critical depth optimization results. The critical depth optimization results and evaluation indicators are then uploaded to the visualization interface display module. The software interface conversion module is used to enable the neural network model subsystem to operate normally in Python.

[0157] Specifically, data preprocessing includes:

[0158] Data preprocessing is essentially data standardization. Specifically, the purpose of data transformation in data cleaning is to standardize the data of each variable before model training, ensuring that the data values ​​are analyzed under the same standard. Z-Score standardization is used to standardize the mean and standard deviation of the original data. The advantage of Z-Score standardization is that it preserves useful information from outliers, making the algorithm less sensitive to them and indirectly avoiding the influence of outliers and extreme values ​​through centering. The processed data conforms to a standard normal distribution, i.e., a mean of 0 and a standard deviation of 1. The transformation function is:

[0159]

[0160] Where, x i Let μ be the data sequence; μ be the mean of the data sequence; and σ be the standard deviation of the data sequence.

[0161] Post-processing of data in the data storage subsystem specifically includes:

[0162] (1) Abnormal data handling:

[0163] Use box plots (e.g.) Figure 3 As shown, the database of the data storage subsystem undergoes anomaly processing to clean up abnormal data and optimize the database. Specifically, box plots are used for anomaly processing during data cleaning. Box plots can describe the distribution characteristics of collected data and visually represent anomaly data points. Therefore, a box plot is drawn by calculating the median, 0.25 and 0.75 quantiles, and upper and lower boundary values ​​of data of the same type. Data points within the box plot range are valid data that conform to the data distribution characteristics, while data points outside the boundaries are anomaly data points. Figure 4 Based on the principles of box plot construction, the formula for calculating the boundary of data points in a box plot is as follows:

[0164] Upper boundary value = Q3 + 1.5IQR (16)

[0165] Lower boundary value = Q1 - 1.5IQR (17)

[0166] Where: Q3 is the 0.75 quantile of the same type of data; Q1 is the 0.25 quantile of the same type of data; IQR is the quantile difference, calculated as IQR = Q3 - Q1;

[0167] (2) Sensitivity analysis

[0168] Sensitivity analysis analyzes and calculates the sensitivity of the output results to multiple uncertain factors, identifying variables that significantly influence the output results. These uncertain factors include the tunnel span B, lateral pressure coefficient λ, friction angle Θ on both sides of the soil column at the tunnel top, tunnel span B, height h, unit weight of surrounding rock γ, and internal friction angle. By fitting the relationship diagram between each parameter and the degree of optimization of the critical depth of shallow tunnels, the correlation influence of each input variable on the optimized value of the critical depth of shallow tunnels can be obtained.

[0169] Sensitivity analysis is an analysis of stored data. It provides a relatively intuitive view of the degree to which the critical depth is affected by different factors, offering important reference value for further research on the critical depth of shallow-buried tunnels. Among these, Figure 9 The figure shows the relationship between tunnel span and optimal critical depth. The straight line in the figure is the fitted curve of the relationship between tunnel span and optimal critical depth.

[0170] Example 2:

[0171] Example 2 uses the system from Example 1 to verify the feasibility of the DBN neural network with existing data and calculates the optimized value for the critical depth of shallow-buried tunnels. The feasibility of the DBN neural network can be verified through the calculation of parameter data; wherein, Figure 10 The input interface for the critical depth optimization subsystem. Figure 11 Optimize the output interface of the critical depth subsystem; Figure 7 This is a simulation diagram of the system of the present invention;

[0172] Specifically, Table 1 shows the evaluation indicators for neural network training. Analysis of Table 1 shows that the predicted values ​​of the neural network training results fit the true values ​​well.

[0173] Table 1 Evaluation metrics for neural network training

[0174] MAE MRE MSE <![CDATA[R 2 ]]> Lower bound of optimization value 0.4585 0.6148 0.2807 0.9721 Optimization value upper limit 0.4412 0.1145 0.2801 0.9791

[0175] Table 2 shows the dispersion of the DBN neural network calculation results. Through the analysis of the dispersion of the upper and lower limits in Table 2, it is proved that the DBN neural network in this embodiment 1 is effective in optimizing the critical depth of Xie Jiaxiao, with basically no outliers, and the average optimization value of the critical depth limit in this study is about 2.5m.

[0176] Table 2. Discrepancy of Neural Network Calculation Results

[0177]

[0178] Table 3 Data from the system calculation process

[0179]

[0180]

[0181] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A system for calculating the critical depth of shallow-buried tunnels in loess soil, characterized in that: include: Neural network model subsystem, critical depth optimization subsystem, and data storage subsystem; The neural network model subsystem uses the data storage subsystem as its database and the manually entered parameters processed by the critical depth optimization subsystem as its input parameters. It then obtains the final optimized critical depth result and returns it to the critical depth optimization subsystem and the data storage subsystem. The neural network model subsystem includes a DBN neural network model construction module, a network model training module, and a DBN neural network application module. The network model training module trains the DBN neural network model built by the DBN neural network model construction module and transmits the trained model to the DBN neural network application module. The DBN neural network application module uses the trained DBN neural network model to calculate the critical depth optimization value of the tunnel parameters manually entered in real time. It then calculates the critical depth optimization value with the critical depth of the shallow tunnel calculated by Xie Jiaxiao's maximum formula to obtain the final optimized critical depth and outputs it to the critical depth optimization subsystem for display. The critical depth optimization subsystem includes a visualization interface display module, an input parameter data processing module, and a software interface conversion module. The input parameter data processing module is used to input tunnel parameters, process the input tunnel parameters, and upload the processed data to the DBN neural network application module. At the same time, it uploads the final optimized critical depth result to the visualization interface display module.

2. The system for calculating the critical depth of shallow-buried tunnels in loess according to claim 1, characterized in that: The steps for building a DBN neural network using the DBN neural network model building module include: S1. Establish the energy function of the RBM stochastic neural network; S2. Based on the energy function, the joint distribution probability of neurons in the explicit and hidden layers of the RBM stochastic neural network is solved; S3. Based on the joint distribution probability, the edge probabilities of neurons in the visible and hidden layers are obtained respectively; S4. Based on the edge probability values, obtain the state probability values ​​of neurons in the visible and hidden layers; S5. Based on the state probability values ​​of neurons in the visible and hidden layers, the optimized RBM stochastic neural network is obtained; S6. Based on the optimized RBM stochastic neural network, establish the DBN neural network model.

3. The system for calculating the critical depth of shallow-buried tunnels in loess according to claim 2, characterized in that: The energy function expression for an RBM stochastic neural network is: In the formula, W is the connection weight matrix of neurons in each layer, v is the state vector of neurons in the visible layer, and v = (v1, v2, ..., v n h is the state vector of the hidden layer neurons, and h = (h1, h2, ..., h2). m ), where a is the bias vector of the visible layer neurons, and a = (a1, a2, ..., a2). n Let b be the bias vector of the hidden layer neurons, and b = (b1, b2, ..., bb2). m ), where n, m, i, and j are all constants.

4. The system for calculating the critical depth of shallow-buried tunnels in loess according to claim 2, characterized in that: The formula for the joint probability distribution of neurons in the visible and hidden layers is: Where E(v,h) is the model energy function.

5. The system for calculating the critical depth of shallow-buried tunnels in loess according to claim 2, characterized in that: The formulas for calculating the marginal probabilities of neurons in the visible and hidden layers are as follows: in, The integral of the hidden layer neurons in the joint probability. This refers to the integral of the joint probability (Gibbs distribution) of neurons in the visible layer.

6. The system for calculating the critical depth of shallow-buried tunnels in loess according to claim 2, characterized in that: The formulas for calculating the probability values ​​of visible and hidden layers are: Wherein, P(v j =1|h) represents the data features received by the hidden layer neurons; P(h) j =1|v) represents the data characteristics received by the visible layer neurons; ω ij These are the connection weights of neurons in each layer.

7. The system for calculating the critical depth of shallow-buried tunnels in loess according to claim 1, characterized in that: The DBN neural network constructed through training optimizes the connection parameters of hidden and visible neurons based on the error between the input data and the mapped data, so that the extracted data information better matches the characteristics of the original data.

8. The system for calculating the critical depth of shallow-buried tunnels in loess according to claim 7, characterized in that: The steps to optimize the connection parameters of hidden and visible neurons include: S1. Find the maximum value of the logarithm of the likelihood function; S2. Based on the maximum value of the logarithm of the likelihood function, combined with gradient iteration of the optimization algorithm, the optimal connection parameters are solved.

9. A system for calculating the critical depth of shallow-buried tunnels in loess soil according to claim 8, characterized in that: The formula for finding the maximum value of the logarithm of the likelihood function is: Where, θ * The maximum value of the logarithm of the likelihood function; L s (θ) is the maximum likelihood function, P(x) t |θ) is the joint probability of a sample and its corresponding code; Optimal connection parameter θ i+1 The solution is: Where, θ i These are model data connection parameters. It is the connection parameter θ i The updated gradient.

10. A system for calculating the critical depth of shallow-buried tunnels in loess according to claim 1, characterized in that: The input tunnel parameters include: tunnel height h, tunnel span B, surrounding rock unit weight γ, and internal friction angle. Cohesion c, ground slope R, surface load P, excavation method W, support structure S.