Underground cavern group surrounding rock parameter inversion method and electronic device based on fracture evolution

Abstract

The disclosure provides a method and an electronic device for inverting surrounding rock parameters of a fissure evolution underground cavern group, and relates to the technical field of engineering data processing. The method comprises the following steps: inputting first condition information and first latent space noise of the underground cavern group into a generator to output first parameter evolution information; the first parameter evolution information comprises first mechanical parameter evolution information and first fissure evolution information; calculating the time sequence displacement of the surrounding rock according to the first mechanical parameter evolution information to obtain first displacement prediction data; inputting the first displacement prediction data, first displacement observation data in the first time sequence and the first parameter evolution information into a discriminator to output comprehensive discrimination information; determining a comprehensive loss function value based on the comprehensive discrimination information to update the generator; and inputting second condition information and second latent space noise into the updated generator to output second parameter evolution information. The disclosure improves the accuracy of the inversion of the surrounding rock parameters of the underground cavern group and can reflect the dynamic evolution process of the surrounding rock.

Description

Technical Field

[0001] This disclosure relates to the field of engineering data processing technology, and more specifically, to a method and electronic equipment for inverting surrounding rock parameters of underground cavern groups that integrates fracture evolution. Background Technology

[0002] In large-scale projects such as hydropower stations, underground cavern complexes are typically designed and constructed to serve as core functional spaces for the main powerhouse, main transformer room, and tailrace tunnel. The stability of the surrounding rock of these underground cavern complexes directly determines the safety and long-term operational reliability of the project. Surrounding rock parameters are the foundation for conducting surrounding rock stability analysis, support structure design, and disaster risk early warning, while parameter inversion is a classic problem in the fields of rock mechanics and underground engineering.

[0003] In related technologies, most methods for inverting surrounding rock parameters focus on finding static parameter fields, which cannot reflect the dynamic evolution process of the surrounding rock and the time-varying characteristics of the parameters. As a result, the inversion results cannot accurately predict the surrounding rock response in subsequent engineering stages, which makes it difficult to meet the needs of dynamic engineering design and risk warning. Summary of the Invention

[0004] This disclosure provides a method and electronic equipment for inverting surrounding rock parameters of underground cavern groups that integrates fracture evolution, so as to at least partially solve the technical problem that the inversion results of surrounding rock parameters of underground cavern groups cannot reflect dynamic processes.

[0005] According to a first aspect of this disclosure, a method for inverting surrounding rock parameters of an underground cavern group by integrating fracture evolution is provided. The method includes: inputting first condition information and first potential space noise of the underground cavern group into a generator, and outputting first parameter evolution information; the first condition information is condition information of the underground cavern group within a first time series, and the first parameter evolution information includes first mechanical parameter evolution information and first fracture evolution information, representing the mechanical parameters and fracture parameters of the surrounding rock of the underground cavern group within the first time series, respectively; calculating the temporal displacement of the surrounding rock based on the first mechanical parameter evolution information to obtain first displacement prediction data; inputting the first displacement prediction data, first displacement observation data within the first time series, and the first parameter evolution information into a discriminator, and outputting comprehensive discrimination information; determining a comprehensive loss function value based on the comprehensive discrimination information, and updating the generator through the comprehensive loss function value; inputting second condition information and second potential space noise of the underground cavern group into the updated generator, and outputting second parameter evolution information.

[0006] According to a second aspect of this disclosure, a device for inverting surrounding rock parameters of an underground cavern group by integrating fracture evolution is provided. The device includes: a generation and processing module configured to input first condition information and first potential space noise of the underground cavern group into a generator, and output first parameter evolution information; the first condition information is condition information of the underground cavern group within a first time series, and the first parameter evolution information includes first mechanical parameter evolution information and first fracture evolution information, respectively representing the mechanical parameters and fracture parameters of the surrounding rock of the underground cavern group within the first time series; and a time-series displacement estimation module configured to, based on the first... The system uses the evolution information of mechanical parameters to calculate the temporal displacement of the surrounding rock, obtaining the first displacement prediction data. A discrimination processing module is configured to input the first displacement prediction data, the first displacement observation data within the first time series, and the first parameter evolution information into a discriminator, and output comprehensive discrimination information. A model update module is configured to determine the comprehensive loss function value based on the comprehensive discrimination information, and update the generator using the comprehensive loss function value. A model application module is configured to input the second condition information of the underground cavern group and the second potential space noise into the updated generator, and output the second parameter evolution information.

[0007] According to a third aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method of the first aspect described above and possible implementations thereof.

[0008] According to a fourth aspect of this disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the method of the first aspect and possible implementations thereof by executing the executable instructions.

[0009] The technical solution disclosed herein has the following beneficial effects:

[0010] On the one hand, a generative adversarial network (GAN) architecture is adopted. A generator produces parameter evolution information containing mechanical parameter evolution and fracture evolution information. A discriminator directly or indirectly discriminates the generated parameter evolution information, thereby updating and optimizing the generator. This ensures that the generator's output can completely and accurately depict the dynamic evolution process of surrounding rock parameters in underground cavern groups. This breaks the static parameter field premise of related technologies, solves the problem of information loss in the surrounding rock deterioration process caused by static inversion, and avoids the shortcomings of related technologies that only use current displacement data to infer the final state parameters while ignoring the evolution process itself. This allows the inversion results to accurately reflect the time-varying characteristics of surrounding rock parameters, enabling precise prediction of the surrounding rock response in subsequent excavation stages and meeting the needs of dynamic engineering design and risk warning. On the other hand, by introducing latent space noise, the generalization of the inversion results is improved, covering different possible surrounding rock evolution scenarios, providing reliable technical support for the safety design and risk management of deep underground engineering. Attached Figure Description

[0011] Figure 1 A flowchart is shown showing a method for inverting surrounding rock parameters of underground cavern groups based on the fusion fracture evolution in one embodiment of this disclosure;

[0012] Figure 2 This diagram illustrates a flowchart of obtaining first parameter evolution information in one embodiment of the present disclosure;

[0013] Figure 3 This diagram illustrates a flowchart of the output of comprehensive discrimination information in one embodiment of the present disclosure;

[0014] Figure 4 This diagram illustrates a flowchart of determining the value of the comprehensive loss function in one embodiment of the present disclosure;

[0015] Figure 5 This diagram illustrates a schematic of the GITPF method architecture in one embodiment of the present disclosure.

[0016] Figure 6 This diagram illustrates the parameter evolution curve in one embodiment of the present disclosure.

[0017] Figure 7 A schematic diagram of an inversion device for surrounding rock parameters of underground cavern groups that integrates fracture evolution is shown in one embodiment of this disclosure;

[0018] Figure 8 A schematic diagram of an electronic device according to one embodiment of the present disclosure is shown. Detailed Implementation

[0019] Exemplary embodiments of this disclosure will be described more fully below with reference to the accompanying drawings.

[0020] The accompanying drawings are schematic illustrations of this disclosure and are not necessarily drawn to scale. Some block diagrams shown in the drawings may be functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in hardware modules or integrated circuits, or in networks, processors, or microcontrollers. Implementations can be carried out in various forms and should not be construed as limited to the examples set forth herein. The features, structures, or characteristics described in this disclosure can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full description of the technical solutions of this disclosure. However, those skilled in the art will recognize that one or more specific details may be omitted when implementing the technical solutions provided in this disclosure, or other methods, components, apparatuses, steps, etc., may be used to replace one or more specific details.

[0021] In large-scale projects such as hydropower stations, underground cavern complexes are typically designed and constructed to serve as core functional spaces for the main powerhouse, main transformer room, and tailrace tunnel. The stability of the surrounding rock of these underground cavern complexes directly determines the safety and long-term operational reliability of the project. Surrounding rock parameters are the foundation for conducting surrounding rock stability analysis, support structure design, and disaster risk early warning, while parameter inversion is a classic problem in the fields of rock mechanics and underground engineering.

[0022] Among related technologies, physical information neural networks (PINs) are representative, incorporating physical laws such as the rock mechanics governing equations and the law of energy conservation as soft constraints into the network training. This reduces the dependence of the inversion on monitoring data and improves the physical rationality of the inversion results. However, this method still has the following problems:

[0023] 1. The inversion process implicitly assumes the existence of a "real" static parameter field in the surrounding rock, and the goal of the inversion is to find this fixed and unchanging parameter combination. However, deep rock excavation is a typical highly disturbed, nonlinear, and irreversible mechanical process, belonging to a dynamic evolution system. Simplifying it to a static parameter field essentially uses a steady-state assumption to describe a transient process, leading to a fundamental loss of process information. Thus, the inversion results can only fit the displacement monitoring data at the current moment, failing to reflect the time-varying characteristics of the parameters, and consequently, failing to predict the surrounding rock response in subsequent excavation stages, making it difficult to meet the needs of dynamic engineering design and risk warning.

[0024] 2. Related technologies only utilize displacement monitoring data at the current moment to attempt to deduce the equivalent parameters of the final state, neglecting the evolutionary process leading to this state. More importantly, different evolutionary paths may converge to similar final states, but the surrounding rock response during subsequent excavation can differ significantly. Therefore, these technologies lack a temporal dimension, failing to characterize the complete evolutionary path of surrounding rock parameters, revealing the intrinsic mechanisms of surrounding rock deterioration, and making accurate predictions of future surrounding rock responses impossible.

[0025] In view of one or more of the above-mentioned problems, this disclosure provides a method for inverting the surrounding rock parameters of underground cavern groups that integrates fracture evolution. This method can be used to invert the surrounding rock parameters of underground cavern groups in large-scale projects such as hydropower stations to obtain relevant parameters such as the mechanical properties of the rock mass. The underground cavern group can be a cavern group during the excavation and construction process.

[0026] Figure 1 An exemplary workflow for inverting surrounding rock parameters of underground cavern groups by integrating fracture evolution is shown, including the following steps:

[0027] S110, input the first condition information of the underground cavern group and the first potential space noise into the generator, and output the first parameter evolution information; the first condition information is the condition information of the underground cavern group in the first time series, and the first parameter evolution information includes the first mechanical parameter evolution information and the first fracture evolution information, which respectively represent the mechanical parameters and fracture parameters of the surrounding rock of the underground cavern group in the first time series;

[0028] S120, the temporal displacement of the surrounding rock is calculated based on the evolution information of the first mechanical parameter to obtain the first displacement prediction data;

[0029] S130, input the first displacement prediction data, the first displacement observation data in the first time series, and the first parameter evolution information into the discriminator, and output the comprehensive discriminant information;

[0030] S140, Determine the comprehensive loss function value based on the comprehensive discrimination information, and update the generator using the comprehensive loss function value;

[0031] S150 inputs the second condition information of the underground cavern group and the second potential space noise into the updated generator and outputs the second parameter evolution information.

[0032] Based on the above method, on the one hand, a generative adversarial network (GAN) architecture is adopted. A generator produces parameter evolution information containing mechanical parameter evolution information and fracture evolution information. A discriminator directly or indirectly discriminates the generated parameter evolution information, thereby updating and optimizing the generator. This ensures that the generator's output can completely and accurately depict the dynamic evolution process of the surrounding rock parameters in underground cavern groups. This breaks the static parameter field premise of related technologies, solves the problem of information loss in the surrounding rock deterioration process caused by static inversion, and avoids the defects of related technologies that only use current displacement data to infer the final state parameters while ignoring the evolution process itself. This allows the inversion results to accurately reflect the time-varying characteristics of the surrounding rock parameters, enabling precise prediction of the surrounding rock response in subsequent excavation stages and meeting the needs of dynamic engineering design and risk warning. On the other hand, by introducing latent space noise, the generalization of the inversion results is improved, covering different possible surrounding rock evolution scenarios, providing reliable technical support for the safety design and risk management of deep underground engineering.

[0033] The following describes, in conjunction with one or more embodiments and related accompanying drawings, Figure 1 Each step in the process will be explained in detail.

[0034] refer to Figure 1 In step S110, the first condition information of the underground cavern group and the first potential space noise are input into the generator, and the first parameter evolution information is output. The first condition information is the condition information of the underground cavern group in the first time series. The first parameter evolution information includes the first mechanical parameter evolution information and the first fracture evolution information, which respectively represent the mechanical parameters and fracture parameters of the surrounding rock of the underground cavern group in the first time series.

[0035] In this context, the latent space is a low-dimensional structured space used for feature processing by the generator. First conditional information and first latent space noise are used for training the generator, while second conditional information and second latent space noise are used for parameter inversion. The first conditional information and first latent space noise can belong to the same underground cavern group or different underground cavern groups. In one implementation, for an underground cavern group requiring parameter inversion, first conditional information (the first time series being a historical time series, which could represent the completed excavation stages of the underground cavern group) can be obtained from its historical information, and corresponding first latent space noise can be constructed. Conditional information containing the current time series (such as the second time series) can be obtained as second conditional information, and corresponding second latent space noise can be constructed. Corresponding to the first conditional information, first displacement observation data can also be obtained, representing displacement observation data of the surrounding rock within the first time series, which can be data monitored by displacement sensors, etc. The first displacement observation data can provide supervision information for model training.

[0036] In one implementation, the first condition information includes:

[0037] Geological background information, such as lithology, number and occurrence of initial joint groups, distribution of initial rock mass quality indicators, and original rock mass mechanical parameters (such as original elastic modulus). E Cohesion c wait);

[0038] Information on the geostress field, such as the maximum principal stress. s 1. Minimum principal stress s 3. Magnitude and direction, lateral pressure coefficient, and initial stress distribution pattern;

[0039] Excavation sequence information, such as the geometric profile of each excavation step, excavation time interval, excavation sequence, and excavation unloading rate;

[0040] Support information, such as support type (anchor bolts, shotcrete, steel bracing, etc.), construction timing, and support parameters (anchor bolt length, spacing, shotcrete thickness, etc.).

[0041] In one implementation, the first latent space noise is randomly generated Gaussian noise. For example, using... noise This represents the first latent space noise, which can be a 100-dimensional Gaussian noise vector, following the... noise~N ( 0 , I )distributed( I (The identity matrix).

[0042] By using a generator, combining the actual engineering and geological conditions of the underground cavern complex (first condition information) and random noise (first latent space noise), first parameter evolution information reflecting the temporal evolution characteristics of surrounding rock parameters is generated. For example, first condition information of the underground cavern complex within a first time series is collected and organized to ensure its completeness and accuracy. The first condition information comprehensively covers key scenario information such as geology, geostress, excavation, and support of the underground cavern complex, providing sufficient constraints for the generator. This avoids generating parameter evolution information that is out of touch with the actual engineering situation. Simultaneously, first latent space noise is generated randomly to ensure its randomness and diversity. The purpose of introducing this noise is to prevent the generator from falling into pattern collapse, i.e., to prevent the generator from generating only a single parameter evolution path and failing to cover different possible surrounding rock evolution scenarios. In one implementation, by introducing first latent space noise, the generator can generate multiple parameter evolution paths that conform to the constraints of the first condition information. Subsequent filtering and optimization by a discriminator can yield the evolution path that best fits the actual engineering situation. The prepared first condition information and the generated first latent space noise are simultaneously input into a pre-built generator. The generator performs feature extraction, temporal reasoning, and mapping transformation on the first condition information and the first latent space noise, and finally outputs the first parameter evolution information. The first parameter evolution information is the evolution path of the surrounding rock parameters derived by the generator, which includes the first mechanical parameter evolution information and the first fracture evolution information. Among them, the first mechanical parameter evolution information completely describes the variation law of mechanical parameters of the surrounding rock of the underground cavern group at each time and spatial location within the first time series, while the first fracture evolution information describes the evolution law of the surrounding rock fractures within the first time series. The two are interconnected and evolve synchronously, together constituting the dynamic evolution process of the surrounding rock parameters.

[0043] The generator can meet the needs of capturing temporal evolution features, ensuring that the output first parameter evolution information has continuous temporal correlation, and can truly reflect the dynamic deterioration process of surrounding rock parameters over time, avoiding situations such as parameter jumps and temporal breaks that do not conform to the actual evolution law.

[0044] In one implementation, the generator includes an encoder, a timing processing unit, and a decoder. (See reference) Figure 2 As shown, the first condition information of the underground cavern group and the first potential space noise are input into the generator, and the first parameter evolution information is output, including the following steps:

[0045] S210, input the first condition information and the first latent space noise into the generator;

[0046] S220, the encoder extracts the first intermediate feature from the first conditional information and the first latent space noise;

[0047] S230, the first intermediate feature is serialized and inferred by the time-series processing unit to obtain the parameter field feature;

[0048] S240 maps the parameter field features to the first parameter evolution information through the decoder.

[0049] For example, a Conditional Generative Adversarial Networks (CGAN) architecture can be used, which includes a generator and a discriminator that incorporate prior information (i.e., conditional information). The generator uses the first conditional information... C and first latent space noise noise As input, generate complete evolution information of the first mechanical parameters. Simultaneously output the evolution information of the first fracture. F ( x , t The two can be represented separately in t Time and space x Mechanical parameters at the location (such as elastic modulus) E Cohesion c internal friction angle Permeability coefficient k The generator uses the mechanical parameter field and fracture parameters to achieve preliminary coupling between the fracture evolution information and the fracture parameter field. The mathematical expression for the generator is:

[0050] (1)

[0051] The first parameter evolution information output by the generator It can be a high-dimensional spacetime tensor with dimensions of . N x × N y × N z × N t × N p ,in, N x , N y , N z These represent the spatial domain Ω (representing the rock mass space of the underground cavern complex) in... x , y , z The number of grid points in the three directions; N t For the time domain T The number of discrete-time steps; N p The number of types of output parameters (e.g., elastic modulus)E Cohesion c internal friction angle Permeability coefficient k Crack probability F Five types of parameters).

[0052] In one implementation, to ensure the temporal continuity of the generation evolution path, the generator can adopt a hybrid architecture of 3D convolution, deconvolution, and ConvLSTM (Convolutional Long Short-Term Memory), as shown in the following structure:

[0053] The encoder can consist of four 3D convolutional layers (conv3d) used to extract the first conditional information. C and first latent space noise noise The spatial features are reduced in dimensionality by sliding convolution kernels. After each convolutional layer, the ReLU activation function is used to enhance the nonlinear fitting ability of the network. After encoding, the first intermediate feature is obtained.

[0054] Temporal processing unit: This may include two ConvLSTM layers (the ConvLSTM layer combines convolutional operations to capture spatial features with the temporal memory properties of LSTM to capture temporal dependencies), used to process temporal evolution features. This ensures that the output at each time step depends not only on the latent spatial noise and conditional information of the current time step, but also on the network's hidden state from the previous time step, thereby guaranteeing that the generated evolutionary path has continuous temporal relevance. Its mathematical expression is:

[0055] (2)

[0056] in, h t-1 , h t The first t -1 moment, the t The hidden state of the network at any given moment; For the first t The parameter field features extracted at each time step; z t For the first t The latent vector at time; i G These are the trainable parameters of the generator.

[0057] The decoder can consist of four 3D deconvolutional layers. It maps the low-dimensional parameter field features extracted by the encoder into high-dimensional first parameter evolution information through progressive upsampling, and finally outputs complete first mechanical parameter evolution information. and the evolution information of the first fracture F (x , t ).

[0058] In the generator described above, the key features of the input information can be accurately extracted and the temporal evolution of the parameters can be captured through the collaborative work of the encoder, the timing processing unit and the decoder. This further improves the temporal continuity and accuracy of the first parameter evolution information, making it more consistent with the actual evolution process of the surrounding rock, and providing a more reliable basis for subsequent temporal displacement estimation and discriminator evaluation.

[0059] In one embodiment, the mechanical parameters include elastic modulus, cohesion, internal friction angle, and permeability coefficient; the fracture parameters include fracture probability. F ( x , t ) indicates in t Time and space x The probability of a rock mass being fractured is the probability that the rock mass belongs to a fractured region (a broken rock mass region). F ( x , t It can include the following characteristics:

[0060] Value range: F ( x , t For any given location ∈ [0,1], a probability threshold can be set to determine whether the rock mass at each location belongs to a fractured region. For example, a probability threshold of 0.5 can be set based on experience or specific needs. F ( x , t When the value is greater than 0.5, the location is determined to be a fracture region. F ( x , t If the value is less than or equal to 0.5, the location is considered a complete rock block area.

[0061] Timing characteristics: F ( x , t Regarding time t Monotonically increasing, that is F ( x , t+ Δ t )≥ F ( x , t )(Δ t >0), which means that the crack will not heal after it is initiated, that is, the crack is irreversible, which is consistent with the actual law of continuous crack expansion during rock mass excavation.

[0062] Spatial characteristics: F ( x , tThe spatial gradient direction reflects the orientation of the fracture, and the gradient magnitude reflects the intensity of fracture development; a larger gradient indicates a denser fracture distribution at that location. Therefore, it is possible to determine the fracture density based on the gradient direction. F ( x , t The spatial gradient determines the direction of fracture propagation (e.g., the direction of maximum spatial gradient), which is usually coordinated with the direction of principal stress in the surrounding rock, such as the direction of fracture propagation being parallel to the direction of principal stress, or there being a small angular deviation between the two (e.g., the angle between the two is less than the angle threshold).

[0063] In one implementation, the mechanical parameters and the crack parameters satisfy the following relationship:

[0064] ;

[0065] ;

[0066] ;

[0067] (3)

[0068] in, E ( x , t ), c ( x , t ), ( x , t ), k ( x , t ) respectively represent the first time series t Time and space x The elastic modulus, cohesion, internal friction angle, and permeability coefficient of the rock mass at that location; F ( x , t ) indicates in t Time and space x The probability of fractures in the rock mass at that location; E 0、 c 0、 , k 0 represents the elastic modulus, cohesion, internal friction angle, and permeability coefficient of the rock mass at the initial time (such as the starting time of the first time series, the starting time of the excavation of the underground cavern group, or the starting time of a certain process). or E , or c , , or kThese represent the correlation coefficients between elastic modulus, cohesion, internal friction angle, permeability coefficient, and fracture probability, respectively. They indicate the deterioration or enhancement coefficients of various mechanical parameters on the development of surrounding rock, and can be determined based on engineering experience and the numerical scale of various parameters. For example: or E =0.65, or c =0.70, =0.15, or k =100.

[0069] The coupling relationship in formula (3) ensures the synchronicity of fracture network evolution and parameter deterioration, which conforms to the evolution law of rock mass mechanics. In one embodiment, the first mechanical parameter evolution information is generated by a generator, and the first fracture evolution information is further calculated according to formula (3). In one embodiment, the generator includes a physical coupling layer, such as a physical coupling layer that can be set in the decoder. After the generator obtains the first mechanical parameter evolution information through processing, it can calculate the first fracture evolution information through the physical coupling layer.

[0070] By establishing the coupling relationship between mechanical parameters and fracture parameters, the intrinsic connection between fracture evolution and mechanical parameter deterioration is clarified, making the generated first parameter evolution information more consistent with the rock mass mechanical evolution law, improving the physical rationality and reliability of the inversion results, further avoiding the generation of non-physical parameter evolution paths, and providing more accurate parameter support for subsequent time-series displacement estimation and engineering applications.

[0071] Continue to refer to Figure 1 In step S120, the temporal displacement of the surrounding rock is calculated based on the evolution information of the first mechanical parameter to obtain the first displacement prediction data.

[0072] Specifically, the evolution information of the first mechanical parameters output by the generator can be mapped to the observable displacement space to obtain the first displacement prediction data, providing a quantifiable basis for subsequent comparison with the first displacement observation data and for the evaluation of the discriminator. Specifically, based on the principles of rock mechanics and combined with the dynamic information of parameter evolution reflected by the evolution information of the first mechanical parameters, a displacement forward calculation model is constructed. This model calculates the predicted values ​​of the surrounding rock displacement at each monitoring time point within the first time series, i.e., the first displacement prediction data. During the time-series displacement extrapolation process, the calculation logic of the displacement forward calculation model conforms to the physical laws of rock mass deformation, avoiding the disconnect between the first displacement prediction data and the actual displacement due to model bias. Furthermore, the dimensions of the first displacement prediction data are consistent with those of the first displacement observation data; that is, the number of monitoring times within the first time series, the number of displacement monitoring points at each monitoring time point, and the dimensions of the displacement data are all consistent, ensuring that the subsequent discriminator can directly compare and analyze the two.

[0073] In one implementation, the temporal displacement of the surrounding rock is calculated using the evolution information of the first mechanical parameter through the following formula to obtain the first displacement prediction data:

[0074] (4)

[0075] in, This represents the evolution information of the first mechanical parameter, which may include the elastic modulus. E Cohesion c internal friction angle Permeability coefficient k ; t i For the first time series i Monitoring time, M This represents the total number of monitoring moments within the first time series. for t i The dimensionality of the first displacement prediction data at time step can be the same as that of the first displacement observation data. They are identical to ensure that the discriminator can directly compare the two; The mapping function for time-series displacement calculation can be achieved using a CNN-Transformer (convolutional neural network-transformer) hybrid surrogate model, i.e. , This is the spatial encoder function for CNN (Convolutional Neural Network). This is a Transformer time encoder function. For example, it inputs the evolution information of the first mechanical parameters at each time step into a mapping function. First, convolutional encoding is performed using the CNN spatial encoder function, and then the information after convolutional encoding is processed using the Transformer temporal encoder function to realize the mapping of the evolution information of the first mechanical parameter to the observation space and output the first displacement prediction data.

[0076] In one implementation, the temporal displacement of the surrounding rock can be estimated based on the evolution information of the first mechanical parameters and the evolution information of the first fracture, thus obtaining the first displacement prediction data. For example, in the above formula (4), the input data is the evolution information of the first mechanical parameters. In addition, it also includes information on the evolution of the first fracture. F ( x , t This provides more comprehensive information about the time-series displacement calculation process.

[0077] Continue to refer to Figure 1In step S130, the first displacement prediction data, the first displacement observation data in the first time series, and the first parameter evolution information are input into the discriminator, and the comprehensive discriminant information is output.

[0078] The discriminator evaluates the first parameter evolution information output by the generator from multiple dimensions. Combining the matching degree between the first displacement prediction data and the first displacement observation data, it outputs comprehensive discriminant information, providing clear guidance for generator optimization. For example, firstly, the first displacement observation data within the first time series is processed. This data comes from actual acquisition by on-site monitoring equipment in the underground cavern complex. After preprocessing such as noise reduction and calibration, the accuracy and reliability of the data are ensured, avoiding distortion of the discriminator's evaluation results due to observation data errors. Subsequently, the first displacement prediction data, the preprocessed first displacement observation data, and the first parameter evolution information are simultaneously input into the discriminator. The discriminator comprehensively analyzes the three types of input information through its built-in multi-dimensional evaluation logic. The evaluation dimensions include the matching degree between the first displacement prediction data and the first displacement observation data, as well as the physical and temporal rationality of the first parameter evolution information. Through multi-dimensional comprehensive evaluation, the discriminator can quantify the quality of the first parameter evolution information and ultimately output comprehensive discriminant information. The output of the comprehensive discrimination information should be in the form of a quantifiable numerical value. The larger the value, the better the generated first parameter evolution information matches the engineering reality and physical laws. Conversely, the smaller the value, the more likely there is a deviation in the first parameter evolution information, and the generator needs to be optimized and adjusted through subsequent steps.

[0079] In one implementation, the method further includes the following steps:

[0080] Based on the first fracture evolution information, a fracture network diagram is constructed for multiple moments within the first time series. The nodes of the fracture network diagram are rock mass units of the underground cavern group. There are edges between two nodes corresponding to adjacent rock mass units, and the edge weight represents the mechanical correlation strength between the two nodes.

[0081] Node aggregation features are extracted from the gap network graph using graph convolutional networks.

[0082] Among them, the evolution information of the first fracture F ( x , tTypically, the fracture probability field is implicit. By constructing a fracture network graph, the evolution information of the first fracture is transformed into an explicit graph structure, facilitating the application of physical constraints and the calculation of subsequent displacement responses. A dynamic graph construction method can be used to transform the evolution information of the first fracture into a fracture network graph at multiple time points within a first time series, achieving a dynamically evolving graph structure. Topological features of the fracture network graph are then extracted using a Graph Convolution Network (GCN), achieving deep coupling between the parameter field and the fracture network. By adding the fracture network graph construction and node aggregation feature extraction stages, the topological features of fracture evolution and the mechanical relationships between rock mass units can be captured more accurately, providing richer evidence for the evaluation of the discriminator.

[0083] For example, the construction of the dynamic graph is based on the cell partitioning of the numerical model, with each time step... t Construct corresponding fracture network diagrams for each. G t =( V t , E t , W t ),in V t For a set of nodes, E t For edge set, W t This is the edge weight matrix. The specific construction steps are as follows:

[0084] Node definition: In the numerical model of an underground cavern complex, each rock mass element can be considered a node, with the total number of nodes matching the total number of elements. The feature vector of each node... H Includes: cell center coordinates ( x , y , z ), unit volume V Initial mechanical parameters (such as) E 0、 c 0、 (etc.), the probability of a fracture at the current moment. F ( x , t ) and current parameter values ​​(e.g. E ( x , t ), c ( x , t )wait).

[0085] Define an edge: for any two adjacent rock mass elements i and j Define that there are edges connecting them (i , j )∈ E t For non-adjacent rock mass elements, edge connections are not defined to ensure the geometric rationality of the graph structure.

[0086] Edge weight calculation: edge weight w ij ( t ) is used to characterize nodes i and j The mechanical correlation strength between them, taking into account geometric adjacency, fracture development degree, and parameter gradient changes, is calculated as follows:

[0087] (5)

[0088] in, w ij ( t ) represents a node i and nodes j Edge weights between them For indicator functions, if node i and nodes j If the rock mass elements are adjacent, the indicator function is set to 1; otherwise, the indicator function is set to 0. F i ( t ), F j ( t ) are nodes i、j At any moment t The probability of a crack; Δ x ij For nodes i and nodes j The distance can be the center distance; α 1. α 2. α 3 represents the weighting coefficient, which can be determined empirically or through cross-validation, for example... α 1 = 0.4 α 2 = 0.4 α 3 = 0.2. In formula (5), the first term ensures geometric adjacency, the second term reflects the influence of fracture development on the mechanical association between units, and the third term reflects the influence of parameter gradient changes caused by fractures on the mechanical association. Formula (5) comprehensively considers three key factors: geometric adjacency, degree of fracture development, and fracture probability gradient. It can more accurately quantify the mechanical association strength between rock mass units, improve the accuracy of fracture network diagrams, and make the node aggregation features extracted by graph convolutional networks more realistic. This provides a more reliable basis for discriminator evaluation and further improves the accuracy of comprehensive discrimination information and the targeting of generator optimization.

[0089] Dynamic evolution: at each time step t According to the current crack probability field F ( x , t Recalculate the edge weights. w ij ( t ), forming a dynamic graph sequence { G t} t∈T This enables dynamic updating of the fracture network topology, keeping pace with the time-varying evolution of the parameter field.

[0090] After constructing a fracture network graph at multiple time points, a Graph Convolutional Network (GCN) is used to extract the topological and mechanical features of the fracture network graph, obtaining node aggregation features. This provides support for subsequent calculation of physical constraint loss and generator regularization. The role of the GCN is to update the feature vector of each node by aggregating the features of its neighboring nodes, capturing the interaction relationships between nodes. Its inter-layer propagation formula is:

[0091] (6)

[0092] in, H (l) , H (l+1) The first l Layer, First l +1 layer node aggregation features; σ( ) is the activation function, such as the ReLU (Rectified Linear Unit) function; For adding self-loops, the adjacency matrix ( I (a unit matrix), used to preserve the feature information of the nodes themselves; for The corresponding degree matrix; W (l) For the first l Trainable weights for layered graph convolutional networks.

[0093] The number of layers in a graph convolutional network can be set according to requirements, such as three layers, as follows: the input layer node feature dimension is 10 (including cell coordinates, volume, initial parameters, current mechanical parameters, and crack probability); the feature dimensions of the middle two layers are 64 and 32 respectively, extracting core topological features through feature dimensionality reduction; the output layer feature dimension is 16, and the output result is the final node aggregation feature. The node aggregation feature can be used to calculate the physical constraint loss (such as the energy criterion for crack propagation) and also serves as the topological regularization term of the generator, ensuring that the generated crack network graph has reasonable connectivity and fractal dimension.

[0094] In one implementation, reference Figure 3 As shown, the above-mentioned input of the first displacement prediction data, the first displacement observation data within the first time series, and the first parameter evolution information into the discriminator, and the output of comprehensive discriminant information, includes the following steps:

[0095] S310, input the first displacement prediction data, the first displacement observation data, the first parameter evolution information and the node aggregation features into the discriminator;

[0096] S320, determine first discrimination information based on data matching based on first displacement prediction data and first displacement observation data;

[0097] S330, determine the second discriminant information based on energy dissipation according to the first parameter evolution information and node aggregation characteristics;

[0098] S340, based on the irreversibility and temporal smoothness of the first crack evolution information and the coordination between the crack propagation direction and the principal stress direction, determine the third discrimination information based on the rationality of crack evolution;

[0099] S350 outputs comprehensive discrimination information based on the first discrimination information, the second discrimination information, and the third discrimination information.

[0100] For example, this disclosure proposes a physical information adversarial discriminator, which consists of three sub-discriminators: a data discriminator, an energy discriminator, and an evolution discriminator. These sub-discriminators work together to achieve dual discrimination of data authenticity and physical plausibility, providing effective constraints for generator training. The three sub-discriminators are described below.

[0101] Data discriminator: Used to discriminate the evolution information of the generated first parameter. The degree of matching with the actual parameter evolution information can be determined by comparing the first displacement prediction data and the first displacement observation data. Its output expression is as follows:

[0102] (7)

[0103] in, This is the first discriminant information; M Total number of monitoring moments; i For the index of the monitoring time; exp( ) is the natural exponential function, used to map the error term within the parentheses to the interval (0,1]. for t i Predicted first displacement data at time step; for t i The first displacement observation data at time σ; uThe standard deviation of the first displacement observation data is used for normalization. It is the L2 norm. The larger the first discriminant information value, the higher the matching degree between the first displacement prediction data and the first displacement observation data, that is, the higher the matching degree between the first parameter evolution information and the real parameter evolution information, and the stronger the data authenticity.

[0104] Energy Discriminator: Based on the principle of energy dissipation, the total dissipated energy during the unloading process of surrounding rock excavation should equal the sum of plastic work and fracture surface energy, and satisfy the second law of thermodynamics. Energy discrimination calculations can be performed based on the evolution information of the first parameter and the nodal aggregation characteristics. For example, calculating the dissipated energy density. as follows:

[0105] (8)

[0106] in, for t time x Stress tensor at location (MPa); for t time x Plastic strain rate at location; for t time x The rate of change of free energy at a given location (J / m³); ":" indicates the inner product operation of tensors. In one embodiment, the stress tensor, plastic strain rate, and rate of change of free energy can be obtained simultaneously during the time-series displacement calculation. For example, based on the evolution information of the first mechanical parameter, the evolution information of the first fracture, and the node aggregation characteristics, considering the influence of the fracture network diagram's topology on the rock mass's mechanical transmission characteristics, the rock mass stress field, plastic strain rate field, and rate of change of free energy field are solved to obtain the stress tensor, plastic strain rate, and rate of change of free energy at different times and locations. In one embodiment, an energy dissipation calculation model can be constructed and trained, such as a multilayer perceptron structure. This model takes the evolution information of the first parameter and the node aggregation characteristics as input, processes them, and outputs the stress tensor, plastic strain rate, and rate of change of free energy.

[0107] The dissipation energy density expressed by formula (8) This is the core constraint of thermodynamic consistency. Based on this, the output expression of the energy discriminator is:

[0108] (9)

[0109] in, This is the second discriminant information. For dissipation energy density, t now This represents the current time (or the end time of the first time series), after integration. This represents the cumulative energy dissipation density from the initial moment to the current moment; This represents the middle value within a reasonable range of energy dissipation. This represents the error in the prediction of dissipated energy; among which This represents the energy dissipation error. The larger the value of the second discrimination information, the more the energy dissipation characteristics of the evolution information of the first parameter conform to physical laws, and the better the thermodynamic consistency.

[0110] Evolution discriminator: Used to determine the rationality of the evolution of the first crack evolution information, which can be judged from three aspects: irreversibility, smoothness, and causality.

[0111] Irreversibility refers to the probability of a crack. F ( x , t ) should be about t Monotonically increasing, meaning it satisfies:

[0112] (10)

[0113] The formula for calculating the penalty for irreversibility can be:

[0114] (11)

[0115] in, Ω represents the rate of change of the fracture probability over time; ReLU is the linear rectified function; Ω is the spatial domain (i.e., the rock mass space of the underground cavern group). t now This refers to the current moment.

[0116] Smoothness refers to whether the evolution information of the first crack changes smoothly over time. For example, the penalty formula for smoothness can be:

[0117] (12)

[0118] In formula (12), the crack probability is... F ( x , t A squared penalty is applied to the second derivative over time. The smaller the second derivative, the smoother the change in the crack probability over time. smooth The smaller the value.

[0119] Causality refers to the requirement that the crack propagation direction should be consistent with the principal stress direction. For example, the penalty formula for causality can be:

[0120] (13)

[0121] in, n F ( x ,t ) represents the spatial location at time t. x The direction of crack propagation at a given location, i.e., the spatial gradient direction of the crack probability at that location. n σ1 ( x , t ) represents the spatial location at time t. x The direction of the principal stress at a given point. The closer the two directions are to being parallel, the closer the dot product is to 1. The larger the value, the better. causal The smaller the value.

[0122] Considering the three aspects of irreversibility, smoothness, and causality, the output expression of the evolutionary discriminator is:

[0123] (14)

[0124] in, β 1. β 2. β 3 represents the penalty coefficients for irreversibility, smoothness, and causality, respectively, which can be set based on experience or specific needs. For example... β 1 = 1.0 β 2 = 0.5 β 3 = 0.5; Penalty irrev Penalty smooth Penalty causal These are the penalty values ​​for irreversibility, smoothness, and causality, respectively.

[0125] In one implementation, the first discrimination information, the second discrimination information, and the third discrimination information are fused using the following formula to obtain comprehensive discrimination information:

[0126] (15)

[0127] in, Indicates the first discriminant information, Indicates the second discriminant information, This indicates the third discriminant information; c 1. c 2. c 3 represents the weighting coefficients, used to adjust the contribution levels of the three sub-discriminators. These coefficients can be determined empirically or through grid search (i.e., pre-defined weighting coefficients). c 1. c 2. c 3. Define several candidate values ​​to form a discrete numerical combination space; traverse all possible coefficient combinations in this space, and substitute each set of coefficients into the space for verification to select the optimal weight coefficients. c 1. c 2. c 3) Determined by methods such as, for example c 1 = 0.4 c 2 = 0.3, c 3 = 0.3.

[0128] Based on the above method, node aggregation features are incorporated into the input information of the discriminator, and three discrimination dimensions—data matching, energy dissipation, and fracture evolution rationality—are introduced. This makes the evaluation of comprehensive discrimination information more comprehensive and accurate, enabling a more precise judgment of the rationality of parameter evolution information. The training objective of the generator can include maximizing comprehensive discrimination information, that is, ensuring that the generated parameter evolution information satisfies data matching, physical laws, and evolution rationality. This improves the targeting of generator optimization, allowing the optimized generator to output parameter evolution information that better conforms to engineering practice and physical laws.

[0129] Continue to refer to Figure 1 In step S140, the comprehensive loss function value is determined based on the comprehensive discrimination information, and the generator is updated using the comprehensive loss function value.

[0130] The comprehensive discrimination information output by the discriminator is transformed into a comprehensive loss function value that can be used to update the generator parameters. The generator parameters are then adjusted through the backpropagation algorithm to gradually improve the rationality and accuracy of the generator output parameter evolution information.

[0131] For example, the comprehensive loss function value can be determined first based on the comprehensive discrimination information and combined with the preset loss function calculation logic. The calculation of the comprehensive loss function value is closely related to the comprehensive discrimination information. The value corresponding to the comprehensive discrimination information and the comprehensive loss function value can be negatively correlated. That is, the larger the value of the comprehensive discrimination information, the smaller the value of the comprehensive loss function, indicating that the generator output result is better; conversely, the smaller the value of the comprehensive discrimination information, the larger the value of the comprehensive loss function, indicating that the generator output result has a greater deviation.

[0132] In one implementation, reference Figure 4 As shown, the above method for determining the comprehensive loss function value based on comprehensive discriminant information includes the following steps:

[0133] S410, determines the generated adversarial loss based on comprehensive discrimination information;

[0134] S420, determine the data matching loss based on the first displacement prediction data and the first displacement observation data;

[0135] S430, determine the physical constraint loss based on the evolution information of the first parameter;

[0136] S440, determine the latent space regularization loss based on the first latent space noise;

[0137] S450 determines the comprehensive loss function value based on generative adversarial loss, data matching loss, physical constraint loss, and latent space regularization loss.

[0138] The generative adversarial loss is used to quantify the loss corresponding to the comprehensive discriminative information of the discriminator, as shown in the following formula:

[0139] (16)

[0140] in, To generate adversarial losses, In order to comprehensively judge the information, express The expectation.

[0141] Data matching loss refers to the difference loss between the predicted data of the first displacement and the observed data of the first displacement. For example, the absolute difference between the two can be used as the data matching loss.

[0142] The physical constraint loss is used to constrain the parameter evolution information output by the generator to conform to physical laws. This disclosure does not limit the physical laws used in the physical constraint loss. For example, physical constraints can be applied at the energy dissipation level, as shown in the following expression:

[0143] (17)

[0144] The latent space regularization loss is used to constrain the distribution of noise in the latent space, preventing generator pattern collapse and ensuring output diversity and stability. For example, the latent space regularization loss can be constructed using standard Gaussian prior constraints, and its expression is as follows:

[0145] (18)

[0146] In one implementation, the expression for the comprehensive loss function is as follows:

[0147] (19)

[0148] in, To generate adversarial losses; For data matching loss; Loss due to physical constraints; For latent space regularization loss; l 1. l 2. l 3 represents the loss weighting coefficient, used to adjust the contribution of each loss term. It can be determined based on experience or through methods such as grid search. For example, taking... l 1 = 10.0 l 2 = 5.0 l 3 = 0.1.

[0149] After determining the comprehensive loss function value, the backpropagation algorithm is used to propagate the comprehensive loss function value back to each layer of the generator network, adjusting the trainable parameters of the generator so that the generator's training objective moves closer to minimizing the comprehensive loss function value. During the parameter update process, the update step size can be controlled to avoid situations where parameter updates are too fast, leading to training oscillations and failure to converge, or updates are too slow, leading to low training efficiency, ensuring that the generator can gradually converge to the optimal state.

[0150] This step can be repeated multiple times, that is, steps S110 to S140 are executed repeatedly. After each update of the generator parameters, the first parameter evolution information is regenerated, the time-series displacement is calculated, and the comprehensive loss function value is evaluated and calculated by the discriminator until the comprehensive loss function value converges to the preset threshold. At this time, the generator parameters reach the optimal state and can output parameter evolution information that conforms to engineering reality and physical laws.

[0151] Continue to refer to Figure 1 In step S150, the second condition information of the underground cavern group and the second potential space noise are input into the updated generator, and the second parameter evolution information is output.

[0152] By using the optimized generator, combined with the second condition information and the second potential space noise, the final usable second parameter evolution information is output, and the inversion of the surrounding rock parameters of the underground cavern group is completed, providing support for subsequent engineering applications.

[0153] For example, the collection and processing of the second conditional information is consistent with that of the first conditional information. It can be conditional information for subsequent time periods of the first time series, used to predict the evolution of surrounding rock parameters in subsequent time periods; or it can be conditional information for different scenarios within the same time series, used to verify the generalization ability of the generator. The generation method of the second potential space noise is consistent with that of the first potential space noise, ensuring its randomness and diversity, and avoiding the generation of singular parameter evolution information.

[0154] The second condition information and the second potential space noise are input into the updated generator. Based on the optimized parameters, the generator performs feature extraction, temporal reasoning, and mapping transformation on the input information, and finally outputs the second parameter evolution information. This second parameter evolution information has undergone multiple rounds of evaluation by the discriminator and multiple rounds of optimization by the generator. It can accurately reflect the temporal evolution law of the surrounding rock parameters of the underground cavern group, and has high rationality and accuracy. It can be directly used for engineering practices such as surrounding rock stability analysis, support structure design, and disaster risk early warning.

[0155] In one implementation, during the excavation of an underground cavern complex, as the project progresses, a first time series is established from the start of the project to the current time. First conditional information within this first time series is acquired, and corresponding first latent space noise is constructed. First displacement observation data is obtained through monitoring. The generator is updated and optimized through steps S110 to S140. Then, the first conditional information is used as second conditional information within a second time series (which may be equivalent to the first time series), and corresponding second latent space noise is constructed (which may be equivalent to or different from the first latent space noise). The updated generator processes the second conditional information and the second latent space noise, outputting second parameter evolution information as the final inversion result.

[0156] This disclosure implements the GITPF (Generative Inversion of Time-Varying Parameter Field) method architecture. Figure 5A schematic diagram of the method architecture is shown. The GITPF method uses a generative adversarial network as its core. A generator produces parameter evolution information of the surrounding rock from its initial state to the present (such as elastic modulus, cohesion, internal friction angle, and permeability coefficient) and fracture evolution information (such as the fracture probability field). The parameter evolution information is input into a displacement forward calculation model (such as a CNN-Transformer model) to map from the parameter space to the observation space, obtaining displacement prediction data (such as the first displacement prediction data). This facilitates comparison with displacement observation data (such as the first displacement observation data), providing data matching criteria for the discriminator and loss function calculation. Fracture evolution information is input into a fracture network graph convolutional unit, where a fracture network graph is constructed. Node aggregation features are extracted through the graph convolutional network. There is a bidirectional coupling between the fracture network graph and the parameter evolution information. Based on the node aggregation features, and by combining one or more of the parameter evolution information, displacement prediction data, and displacement observation data, physical constraint features are determined. Physical constraint features can be determined based on the specific physical laws employed. For example, in energy dissipation constraints, physical constraint features may include stress tensor, plastic strain rate, free energy change rate, and dissipation energy density. These physical constraint features can be determined based on node aggregation characteristics and parameter evolution information. Physical constraint features can provide physical rationality criteria for the discriminator and loss function calculation. Furthermore, fracture evolution information itself can provide evolutionary rationality criteria. The discriminator can include three sub-discriminators to judge the parameter evolution information generated by the generator from three aspects: data matching, physical rationality, and evolutionary rationality, obtaining comprehensive discriminant information. Further, a multi-scale loss fusion mechanism can be adopted to fuse multiple losses corresponding to the comprehensive discriminant information, such as generative adversarial loss, data matching loss, physical constraint loss, and latent space regularization loss, to obtain a comprehensive loss function value. The generator is updated and optimized based on the comprehensive loss function value, enabling it to perform higher-quality parameter inversion in subsequent engineering stages. This closed-loop architecture enables deep coupling between generative adversarial networks and physical laws and monitoring data, completing a paradigm shift from static parameter identification to dynamic evolution path generation, and improving the quality and accuracy of surrounding rock parameter inversion.

[0157] The embodiments of this disclosure will be further illustrated and verified through examples below.

[0158] A hydroelectric power station's underground cavern complex was selected as an engineering example to verify the effectiveness of the GITPF method in this disclosure. This underground cavern complex consists of the main powerhouse, main transformer room, tailrace tunnel, and diversion tunnel, with a total length of approximately 1200m, a maximum excavation span of 32m, and a maximum excavation height of 78m, classifying it as a typical deep, large-scale underground cavern complex. The rock mass in this project area is mainly granite, with moderate rock integrity, multiple sets of joints and fissures, high initial in-situ stress, and a maximum principal stress of [missing information]. s 1. Up to 35 MPa, minimum principal stress s3 represents 12 MPa, with a lateral pressure coefficient λ = 0.35. During the tunnel excavation, 15 key displacement monitoring points were selected (8 in the main powerhouse, 5 in the main transformer room, and 2 in the tailrace tunnel). The monitoring period was 180 days, and displacement observation data were acquired at 60 monitoring times. The monitoring accuracy was 0.01 mm, providing reliable observation data for inversion analysis. The correlation coefficients between different mechanical parameters and cracks can be taken as... or E =0.65, or c =0.70, =0.15, or k =100.

[0159] Three traditional inversion methods were selected for comparison:

[0160] 1) Static parameter inversion method based on particle swarm optimization (PSO) (PSO-SI): The Mohr-Coulomb elastoplastic model is used, and the static parameter field is optimized by using displacement monitoring data as a constraint, without considering the time-varying characteristics of the parameters.

[0161] 2) BP-SI (Backpropagation Neural Network-based Surrogate Inversion Method): Constructs a mapping relationship between parameters and displacements to invert the static parameter field, but lacks physical constraints;

[0162] 3) Static inversion method based on physical information neural network (PINN) (PINN-SI): The rock mechanics control equations are embedded in the network loss function to invert the static parameter field. It has certain physical constraints, but does not consider time-varying characteristics.

[0163] All four methods (including the GITPF method and the three traditional inversion methods mentioned above) use the same displacement observation data, condition information, and physical correlation models. The comparison indicators include: displacement prediction error, deviation between inversion parameters and field test values, accuracy of subsequent excavation displacement prediction, and computational efficiency.

[0164] The evolution path of the time-varying parameter field of the surrounding rock of the underground powerhouse cavern group was obtained by inversion using the GITPF method, and key areas of the main powerhouse arch (such as...) were selected. x =50m, y =30m, z The parameter evolution curve of (=40m) is referenced. Figure 6 As shown, the mechanical parameters of the surrounding rock exhibit obvious time-varying deterioration characteristics: elastic modulus E The pressure gradually decreased from an initial 35 GPa to 22 GPa, a decrease of approximately 37%; cohesion c The pressure decreased from an initial 2.8 MPa to 0.84 MPa, a decrease of approximately 70%; the internal friction angle... The permeability decreased from an initial 38° to 32.3°, a decrease of approximately 15%; k From the initial 1.2×10 -7 The speed increased to 1.21 × 10 m / s -5 The speed increased by approximately 100 times. The parameter evolution trend is consistent with the actual laws of rock mass excavation unloading and fracture propagation: in the early stage of excavation (0-60 days), the parameter deterioration rate is relatively fast, mainly due to the rapid accumulation of rock mass damage caused by excavation unloading; in the middle stage of excavation (60-120 days), the parameter deterioration rate slows down, and the rock mass damage enters a stable development stage; in the later stage of excavation (120-180 days), the parameters tend to stabilize, and the rock mass damage reaches its limit state.

[0165] Comparison of inversion parameters with field test values ​​(post-excavation rock mass parameters obtained from field sampling tests): E =21.5GPa, cohesion c =0.8MPa, internal friction angle =32°, permeability coefficient k =1.18×10 -5 The relative errors of the GITPF method inversion results are all less than 3% (m / s). In contrast, the relative errors of the three static inversion methods, PSO-SI, BP-SI, and PINN-SI, are all greater than 8%, as shown in Table 1. Among them, the BP-SI method has the largest relative error, reaching more than 15%, indicating that the parameter accuracy of the GITPF method inversion is significantly higher than that of the traditional static inversion method.

[0166] Table 1 Comparison of inversion results of the four methods

[0167]

[0168] Two indicators, energy dissipation characteristics and irreversibility of crack propagation, were selected. The output expressions of the above energy criterion and evolution criterion were used to verify the physical rationality of the inversion results of the GITPF method. The energy dissipation error and irreversibility penalty value of the four methods were compared and the results are shown in Table 2.

[0169] Table 2 Comparison of the physical rationality of the four methods

[0170]

[0171] As shown in Table 2, the energy dissipation error of the GITPF method is only 2.1%, and the irreversibility penalty is only 0.03, which is far lower than the three traditional static inversion methods. This indicates that the GITPF method uses physical information to counteract the constraints of the discriminator and physical constraint loss, and the generated parameter evolution information conforms to the principle of energy dissipation and the law of irreversible crack propagation, with significantly better physical rationality than traditional methods. In contrast, the BP-SI method, lacking physical constraints, has the largest energy dissipation error and irreversibility penalty, resulting in the worst physical rationality. Although the PINN-SI method introduces physical constraints, its physical rationality is still inferior to the GITPF method because it does not consider time-varying characteristics.

[0172] It is evident that the GITPF method constructed in this disclosure has good computational efficiency, can meet the needs of practical engineering applications, and provides reliable technical support for the inversion of surrounding rock parameters and subsequent stability prediction of underground cavern groups.

[0173] In summary, the embodiments disclosed herein achieve the following technical effects:

[0174] This invention achieves a dimensional leap from static point estimation to dynamic spatiotemporal fields. Traditional methods can only output a few static parameters, essentially describing transient processes using steady-state assumptions. The implementation method disclosed here fuses generative adversarial networks with ConvLSTM to output a four-dimensional parameter field containing three spatial and temporal dimensions, extending to multiple time steps in the temporal dimension. Simultaneously, through implicit representation of the fracture probability field, it achieves end-to-end modeling of fracture initiation, expansion, and penetration. The inverted fracture fractal dimension highly matches the field borehole observations, realizing a transition from parameter identification to process understanding.

[0175] A data- and physics-driven generative inversion mechanism was constructed, significantly improving inversion accuracy. Purely data-driven models often produce parameter combinations that violate the laws of mechanics. The GITPF method proposes a discriminator based on physical information adversarial mechanisms, encoding the principles of energy dissipation, fracture irreversibility, and the Mohr-Coulomb criterion as adversarial losses. This ensures that the generated path simultaneously satisfies data matching and physical rationality, and that the parameter combinations always satisfy mechanical compatibility, thus resolving the inherent problem of parameter contradictions in purely data-driven methods.

[0176] A time-series prediction framework based on evolution paths was established to accurately predict subsequent responses. Traditional methods only utilize current displacement data, failing to reflect the continuous deterioration trend of the surrounding rock and inevitably underestimating subsequent deformation. GITPF ensures the temporal relevance of the evolution path through ConvLSTM, enabling the model to grasp the complete trajectory from the initial state to the current state.

[0177] This disclosure also provides a device for inverting surrounding rock parameters of underground cavern groups that integrates fracture evolution. (Reference) Figure 7 As shown, the underground cavern group surrounding rock parameter inversion device 700 includes:

[0178] The generation and processing module 710 is configured to input the first condition information and the first potential space noise of the underground cavern group into the generator and output the first parameter evolution information; the first condition information is the condition information of the underground cavern group in the first time series, and the first parameter evolution information includes the first mechanical parameter evolution information and the first fracture evolution information, which respectively represent the mechanical parameters and fracture parameters of the surrounding rock of the underground cavern group in the first time series.

[0179] The temporal displacement estimation module 720 is configured to perform temporal displacement estimation of the surrounding rock based on the evolution information of the first mechanical parameters, and obtain the first displacement prediction data;

[0180] The discrimination processing module 730 is configured to input the first displacement prediction data, the first displacement observation data in the first time series, and the first parameter evolution information into the discriminator, and output comprehensive discrimination information.

[0181] The model update module 740 is configured to determine the comprehensive loss function value based on the comprehensive discrimination information and update the generator using the comprehensive loss function value.

[0182] The model application module 750 is configured to input the second condition information of the underground cavern group and the second potential space noise into the updated generator and output the second parameter evolution information.

[0183] In one embodiment, the first condition information includes geological background information, geostress field information, excavation sequence information, and support information; the first potential space noise is randomly generated Gaussian noise.

[0184] In one embodiment, the generator includes an encoder, a timing processing unit, and a decoder; the step of inputting the first condition information and the first latent space noise of the underground cavern group into the generator and outputting the first parameter evolution information includes: inputting the first condition information and the first latent space noise into the generator; extracting a first intermediate feature from the first condition information and the first latent space noise through the encoder; performing serialization reasoning on the first intermediate feature through the timing processing unit to obtain parameter field features; and mapping the parameter field features to the first parameter evolution information through the decoder.

[0185] In one embodiment, the mechanical parameters include elastic modulus, cohesion, internal friction angle, and permeability coefficient, and the fracture parameters include fracture probability; the mechanical parameters and the fracture parameters satisfy the following relationship:

[0186] ;

[0187] ;

[0188] ;

[0189] ;

[0190] in, E ( x , t ), c ( x , t ), ( x , t ), k ( x , t ) respectively represent the time series within the first time series t Time and space x The elastic modulus, cohesion, internal friction angle, and permeability coefficient of the rock mass at that location; F ( x , t ) indicates in t Time and space x The probability of fractures in the rock mass at that location; E 0、 c 0、 , k 0 represents the rock mass's elastic modulus, cohesion, internal friction angle, and permeability coefficient at the initial moment, respectively; or E , or c , , or k These represent the correlation coefficients between elastic modulus, cohesion, internal friction angle, permeability coefficient, and crack probability, respectively.

[0191] In one embodiment, the device is further configured to: construct a fracture network graph for multiple moments within the first time series based on the first fracture evolution information, wherein the nodes of the fracture network graph are rock mass units of an underground cavern group, and there are edges between two nodes corresponding to adjacent rock mass units, with edge weights characterizing the mechanical correlation strength between the two nodes; extract node aggregation features from the fracture network graph using a graph convolutional network; and input the first displacement prediction data, the first displacement observation data within the first time series, and the first parameter evolution information into a discriminator, and output comprehensive discriminant information, including: inputting the first displacement prediction data, The first displacement observation data, the first parameter evolution information, and the node aggregation features are input into the discriminator; based on the first displacement prediction data and the first displacement observation data, a first discrimination information based on data matching is determined; based on the first parameter evolution information and the node aggregation features, a second discrimination information based on energy dissipation is determined; based on the irreversibility, temporal smoothness, and coordination between the crack propagation direction and the principal stress direction of the first crack evolution information, a third discrimination information based on the rationality of crack evolution is determined; based on the first discrimination information, the second discrimination information, and the third discrimination information, the comprehensive discrimination information is output.

[0192] In one implementation, the edge weight is calculated using the following formula:

[0193] ;

[0194] in, w ij ( t ) represents a node i and nodes j Edge weights between them For indicator functions, F i ( t ), F j ( t ) are nodes i、j At any moment t The probability of a crack; Δ x ij For nodes i and nodes j distance, α 1. α 2. α 3 represents the weighting coefficient.

[0195] In one implementation, outputting the comprehensive discrimination information based on the first discrimination information, the second discrimination information, and the third discrimination information includes: fusing the first discrimination information, the second discrimination information, and the third discrimination information using the following formula to obtain the comprehensive discrimination information:

[0196] ;

[0197] in, Indicates the first discriminant information, Indicates the second discriminant information, This indicates the third discriminant information; c 1. c 2. c 3 represents the weighting coefficient.

[0198] In one embodiment, the step of calculating the temporal displacement of the surrounding rock based on the evolution information of the first mechanical parameters to obtain the first displacement prediction data includes: calculating the temporal displacement of the surrounding rock based on the evolution information of the first mechanical parameters using the following formula to obtain the first displacement prediction data:

[0199] ;

[0200] in, This indicates the evolution information of the first mechanical parameter. t i For the first time series, the first i Monitoring time, M This represents the total number of monitoring moments within the first time series. for t i Predicted first displacement data at time step; , For CNN spatial encoder functions, This is a Transformer timing encoder function.

[0201] In one implementation, determining the comprehensive loss function value based on the comprehensive discrimination information includes: determining a generative adversarial loss based on the comprehensive discrimination information; determining a data matching loss based on the first displacement prediction data and the first displacement observation data; determining a physical constraint loss based on the first parameter evolution information; determining a latent space regularization loss based on the first latent space noise; and determining the comprehensive loss function value based on the generative adversarial loss, the data matching loss, the physical constraint loss, and the latent space regularization loss.

[0202] The specific details of each part of the above-mentioned device have been described in detail in the method section of the implementation plan. For any undisclosed details, please refer to the implementation plan of the method section, and therefore will not be repeated here.

[0203] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to exemplary embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0204] This disclosure also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the method steps of various exemplary embodiments of this disclosure.

[0205] In one implementation, the computer program product can be a tangible product, such as a computer-readable storage medium storing a computer program. The readable storage medium can be based on electrical, magnetic, optical, electromagnetic, infrared, or other signals, and includes, but is not limited to: Random Access Memory (RAM), Read-Only Memory (ROM), magnetic tape, floppy disk, flash memory, Hard Disk Drive (HDD), Solid State Disk (SSD), etc. For example, the computer program product can be a non-volatile storage medium storing a computer program, such as read-only memory, NAND flash memory, etc.

[0206] In one implementation, the computer program product can be an intangible product. For example, the computer program product can be a virtual digital product, such as an executable file or installation package containing a computer program.

[0207] Computer program code can be written in one or more programming languages. Examples of programming languages ​​include C, Java, and C++. Program code can execute entirely on the user's computing device, partially on the user's computing device, or as a standalone software package. It can also execute partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, such as a Local Area Network (LAN) or a Wide Area Network (WAN), or it can be connected to an external computing device (e.g., via an internet connection provided by a mobile network operator).

[0208] Computer programs can be carried or transmitted via signals such as electrical, magnetic, optical, electromagnetic, and infrared rays. Electronic devices can convert the signals carrying computer programs into digital signals, thereby running the computer programs. When a computer program runs on an electronic device, its code is used to cause the electronic device to execute (more specifically, to be executed by the processor of the electronic device) the method steps of various embodiments of this disclosure, such as... Figure 1 The method and steps.

[0209] Implementing the above method steps through a computer program achieves the following technical effects: Firstly, by adopting a generative adversarial network architecture, a generator produces parameter evolution information containing mechanical parameter evolution information and fracture evolution information. A discriminator directly or indirectly discriminates the generated parameter evolution information, thereby updating and optimizing the generator. This ensures that the generator's output can completely and accurately depict the dynamic evolution process of surrounding rock parameters in underground cavern groups, breaking the static parameter field premise in related technologies and solving the problem of information loss in the surrounding rock deterioration process caused by static inversion. It also avoids the defects of related technologies that only use displacement data at the current moment to infer the final state parameters and ignore the evolution process itself. This allows the inversion results to accurately reflect the time-varying characteristics of surrounding rock parameters, enabling precise prediction of the surrounding rock response in the subsequent excavation stage and meeting the needs of dynamic engineering design and risk warning. Secondly, by introducing latent space noise, the generalization of the inversion results is improved, covering different possible surrounding rock evolution scenarios, providing reliable technical support for the safety design and risk management of deep underground engineering.

[0210] This disclosure also provides an electronic device. The electronic device includes a processor and a memory. The memory stores executable instructions for the processor, such as computer programs. The processor executes the executable instructions to perform the method steps of various exemplary embodiments of this disclosure.

[0211] The following is for reference. Figure 8The electronic device is illustrated by way of a general-purpose computing device. It should be understood that... Figure 8 The electronic device 800 shown is merely an example and should not be construed as limiting the functionality or scope of this disclosure.

[0212] like Figure 8 As shown, the electronic device 800 may include: a processor 810, a memory 820, a bus 830, an I / O (input / output) interface 840, and a network adapter 850.

[0213] The memory 820 may include volatile memory, such as RAM 821 and cache unit 822, and may also include non-volatile memory, such as ROM 823. The memory 820 may also include one or more program modules 824, including but not limited to: an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. For example, program module 824 may include the modules described above.

[0214] The processor 810 may include one or more processing units, such as an AP (Application Processor), a modem processor, a GPU (Graphics Processing Unit), an ISP (Image Signal Processor), a controller, an encoder, a decoder, a DSP (Digital Signal Processor), a baseband processor, and / or an NPU (Neural-Network Processing Unit).

[0215] The processor 810 can be used to execute executable instructions stored in the memory 820 to perform method steps of various embodiments of this disclosure, such as... Figure 1 The method and steps.

[0216] By executing the above method steps through processor 810, the following technical effects are achieved: Firstly, by adopting a generative adversarial network architecture, a generator produces parameter evolution information containing mechanical parameter evolution information and fracture evolution information. A discriminator directly or indirectly discriminates the generated parameter evolution information, thereby updating and optimizing the generator. This ensures that the generator's output can completely and accurately depict the dynamic evolution process of the surrounding rock parameters of the underground cavern group, breaking the static parameter field premise in related technologies, solving the problem of information loss in the surrounding rock deterioration process caused by static inversion, and avoiding the defects of related technologies that only use the displacement data at the current moment to infer the final state parameters and ignore the evolution process itself. This makes the inversion results accurately reflect the time-varying characteristics of the surrounding rock parameters, achieving accurate prediction of the surrounding rock response in the subsequent excavation stage, and meeting the needs of dynamic engineering design and risk warning. Secondly, by introducing latent space noise, the generalization of the inversion results is improved, which can cover different possible surrounding rock evolution scenarios, providing reliable technical support for the safety design and risk management of deep underground engineering.

[0217] Bus 830 is used to connect different components of electronic device 800 and may include data bus, address bus and control bus.

[0218] Electronic device 800 can communicate with one or more external devices 900 (such as keyboard, mouse, external controller, etc.) through I / O interface 840.

[0219] Electronic device 800 can communicate with one or more networks via network adapter 850. For example, network adapter 850 can provide mobile communication solutions such as 3G / 4G / 5G, or wireless communication solutions such as wireless LAN, Bluetooth, and near-field communication. Network adapter 850 can communicate with other modules of electronic device 800 via bus 830.

[0220] In one embodiment, the electronic device 800 further includes a display for displaying a graphical user interface.

[0221] although Figure 8 As not shown in the diagram, other hardware and / or software modules may also be configured in the electronic device 800, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (Redundant Arrays of Independent Disks) systems, tape drives, and data backup storage systems.

[0222] As can be seen from the above, the technical solutions disclosed herein can be implemented as methods, apparatus, systems, computer program products, storage media, electronic devices, etc. Those skilled in the art will understand that various aspects of this disclosure can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or an implementation combining hardware and software aspects. Exemplarily, these three forms can be referred to as "circuit," "module," and "system," respectively.

[0223] It should be understood that this disclosure is not limited to the specific methods, steps, or structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. Those skilled in the art will readily conceive of other embodiments based on the specific implementations provided in this disclosure. Therefore, the specific implementations provided in this disclosure are merely exemplary, and the scope and spirit of this disclosure are indicated by the claims, and should cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary technical means in the art not disclosed in this disclosure.

Claims

1. A method for inverting surrounding rock parameters of underground cavern groups by integrating fracture evolution, characterized in that, The method includes: The generator inputs the first condition information of the underground cavern group and the first potential space noise, and outputs the first parameter evolution information. The first condition information is the condition information of the underground cavern group in the first time series. The first parameter evolution information includes the first mechanical parameter evolution information and the first fracture evolution information, which respectively represent the mechanical parameters and fracture parameters of the surrounding rock of the underground cavern group in the first time series. Based on the evolution information of the first mechanical parameters, the temporal displacement of the surrounding rock is calculated to obtain the first displacement prediction data; The first displacement prediction data, the first displacement observation data within the first time series, and the first parameter evolution information are input into the discriminator, and comprehensive discrimination information is output. The comprehensive loss function value is determined based on the comprehensive discrimination information, and the generator is updated using the comprehensive loss function value; The second condition information of the underground cavern group and the second potential space noise are input into the updated generator, and the second parameter evolution information is output. The method further includes: constructing a fracture network graph for multiple moments within the first time series based on the first fracture evolution information, wherein the nodes of the fracture network graph are rock mass units of the underground cavern group, and there is an edge between two nodes corresponding to adjacent rock mass units, and the edge weight characterizes the mechanical correlation strength between the two nodes; and extracting node aggregation features from the fracture network graph through a graph convolutional network. The step of inputting the first displacement prediction data, the first displacement observation data within the first time series, and the first parameter evolution information into the discriminator and outputting comprehensive discrimination information includes: inputting the first displacement prediction data, the first displacement observation data, the first parameter evolution information, and the node aggregation features into the discriminator; determining first discrimination information based on data matching based on the first displacement prediction data and the first displacement observation data; determining second discrimination information based on energy dissipation based on the first parameter evolution information and the node aggregation features; determining third discrimination information based on the rationality of fracture evolution based on the irreversibility, temporal smoothness, and coordination between the fracture propagation direction and the principal stress direction of the first fracture evolution information; and outputting the comprehensive discrimination information based on the first discrimination information, the second discrimination information, and the third discrimination information.

2. The method according to claim 1, characterized in that, The first condition information includes geological background information, geostress field information, excavation sequence information, and support information; the first potential space noise is randomly generated Gaussian noise.

3. The method according to claim 1, characterized in that, The generator includes an encoder, a timing processing unit, and a decoder; The process of inputting the first condition information of the underground cavern group and the first potential space noise into the generator, and outputting the first parameter evolution information, includes: The first condition information and the first latent space noise are input into the generator; The encoder extracts a first intermediate feature from the first conditional information and the first latent space noise. The first intermediate feature is serialized and inferred by the time-series processing unit to obtain the parameter field feature; The decoder maps the parameter field features to the first parameter evolution information.

4. The method according to claim 1, characterized in that, The mechanical parameters include elastic modulus, cohesion, internal friction angle, and permeability coefficient; the fracture parameters include fracture probability; the mechanical parameters and the fracture parameters satisfy the following relationship: ； ； ； ； in, E ( x , t ), c ( x , t ), ( x , t ), k ( x , t ) respectively represent the time series within the first time series t Time and space x The elastic modulus, cohesion, internal friction angle, and permeability coefficient of the rock mass at that location; F ( x , t ) indicates in t Time and space x The probability of fractures in the rock mass at that location; E 0、 c 0、 , k 0 represents the rock mass's elastic modulus, cohesion, internal friction angle, and permeability coefficient at the initial moment, respectively; η E , η c , , η k These represent the correlation coefficients between elastic modulus, cohesion, internal friction angle, permeability coefficient, and crack probability, respectively.

5. The method according to claim 1, characterized in that, The formula for calculating the edge weight is as follows: ； in, w ij ( t ) represents a node i and nodes j Edge weights between them For indicator functions, F i ( t ), F j ( t ) are nodes i, j At any moment t The probability of a crack; Δ x ij For nodes i and nodes j distance, α 1. α 2. α 3 represents the weighting coefficient.

6. The method according to claim 1, characterized in that, The step of outputting the comprehensive discrimination information based on the first discrimination information, the second discrimination information, and the third discrimination information includes: The comprehensive discrimination information is obtained by fusing the first discrimination information, the second discrimination information, and the third discrimination information using the following formula: ； in, Indicates the first discriminant information, Indicates the second discriminant information, This indicates the third discriminant information; γ 1. γ 2. γ 3 represents the weighting coefficient.

7. The method according to any one of claims 1 to 6, characterized in that, The step of calculating the temporal displacement of the surrounding rock based on the evolution information of the first mechanical parameters to obtain the first displacement prediction data includes: The first displacement prediction data is obtained by extrapolating the temporal displacement of the surrounding rock based on the evolution information of the first mechanical parameter using the following formula: ； in, This indicates the evolution information of the first mechanical parameter. t i For the first time series, the first i Monitoring time, M This represents the total number of monitoring moments within the first time series. for t i Predicted first displacement data at time step; , For CNN spatial encoder functions, This is a Transformer timing encoder function.

8. The method according to any one of claims 1 to 6, characterized in that, Determining the comprehensive loss function value based on the comprehensive discriminant information includes: The generated adversarial loss is determined based on the comprehensive discrimination information; The data matching loss is determined based on the first displacement prediction data and the first displacement observation data; The physical constraint loss is determined based on the evolution information of the first parameter; The latent space regularization loss is determined based on the first latent space noise. The comprehensive loss function value is determined based on the generative adversarial loss, the data matching loss, the physical constraint loss, and the latent space regularization loss.

9. A device for inverting surrounding rock parameters of underground cavern groups by integrating fracture evolution, characterized in that, The device includes: The generation and processing module is configured to input the first condition information and the first potential space noise of the underground cavern group into the generator and output the first parameter evolution information; the first condition information is the condition information of the underground cavern group in the first time series, and the first parameter evolution information includes the first mechanical parameter evolution information and the first fracture evolution information, which respectively represent the mechanical parameters and fracture parameters of the surrounding rock of the underground cavern group in the first time series; The temporal displacement estimation module is configured to perform temporal displacement estimation of the surrounding rock based on the evolution information of the first mechanical parameters, and obtain the first displacement prediction data; The discrimination processing module is configured to input the first displacement prediction data, the first displacement observation data in the first time series, and the first parameter evolution information into the discriminator, and output comprehensive discrimination information. The model update module is configured to determine the comprehensive loss function value based on the comprehensive discrimination information, and update the generator using the comprehensive loss function value; The model application module is configured to input the second condition information of the underground cavern group and the second potential space noise into the updated generator, and output the second parameter evolution information. The device is further configured to: construct a fracture network graph for multiple moments within the first time series based on the first fracture evolution information, wherein the nodes of the fracture network graph are rock mass units of the underground cavern group, and there are edges between two nodes corresponding to adjacent rock mass units, with edge weights characterizing the mechanical correlation strength between the two nodes; and extract node aggregation features from the fracture network graph through a graph convolutional network. The step of inputting the first displacement prediction data, the first displacement observation data within the first time series, and the first parameter evolution information into the discriminator and outputting comprehensive discrimination information includes: inputting the first displacement prediction data, the first displacement observation data, the first parameter evolution information, and the node aggregation features into the discriminator; determining first discrimination information based on data matching based on the first displacement prediction data and the first displacement observation data; determining second discrimination information based on energy dissipation based on the first parameter evolution information and the node aggregation features; determining third discrimination information based on the rationality of fracture evolution based on the irreversibility, temporal smoothness, and coordination between the fracture propagation direction and the principal stress direction of the first fracture evolution information; and outputting the comprehensive discrimination information based on the first discrimination information, the second discrimination information, and the third discrimination information.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 8.

11. An electronic device, characterized in that, include: Processor and memory; The memory is used to store executable instructions of the processor; the processor is configured to implement the method of any one of claims 1 to 8 by executing the executable instructions.

Images