A method and system for active control of cracks in tunnel lining based on SHCC

Abstract

The application relates to the technical field of tunnel lining protection, and discloses a tunnel lining crack active control method and system based on SHCC. The method comprises the following steps: continuously collecting dynamic image data of a lining surface crack through a sensing array arranged in a tunnel; inputting the data into a crack semantic analysis module, which calls a crack historical archive and a tunnel structure database to jointly infer and outputs a crack control target vector; constructing a geometric topology network with lining segments as vertices and mechanical dependence between the segments as edges; after mapping the control target vector to the network, calculating the influence scores of all the vertices through a similarity diffusion algorithm, and screening a key control vertex set; running a multi-index decision model for the set to generate a candidate control instruction list, executing the instruction to complete high-precision injection of SHCC material, and realizing active control of the tunnel lining crack and adaptation to the control requirements of dynamic changes of the crack in the tunnel operation process.

Description

Technical Field

[0001] This invention relates to the field of tunnel lining protection technology, specifically to a method and system for active control of tunnel lining cracks based on SHCC. Background Technology

[0002] As a crucial infrastructure in transportation, water conservancy, and other fields, the lining structure of tunnels directly affects their overall stability and operational safety. During long-term service, tunnels are susceptible to cracking due to various factors such as geological fluctuations, repeated vehicle loads, concrete aging, and temperature stress changes. The continuous propagation of these cracks can compromise the integrity of the lining structure, reduce its load-bearing capacity and impermeability, and even trigger safety risks such as spalling and collapse, threatening the safe operation of the tunnel.

[0003] Tunnel lining crack control primarily relies on passive repair, with common techniques including traditional grouting and bonding reinforcement materials. Traditional grouting depends on periodic manual inspection of crack location and size, making it difficult to capture the dynamic development of cracks in real time. Often, repairs are delayed, leading to irreversible crack expansion. Furthermore, the selection and injection volume of grouting materials lack precise guidance, easily resulting in insufficient injection or excessive accumulation, failing to achieve targeted repair. Bonding reinforcement materials only acts on the lining surface, failing to address the root cause of internal mechanical imbalances in cracks. The reinforcement effect is short-lived and cannot adapt to the dynamic changes in cracks.

[0004] Existing active control methods often rely on isolated real-time sensor data without considering the historical development patterns of cracks and the overall tunnel structure. This leads to a one-sided assessment of crack conditions and a lack of comprehensive basis for setting control targets. When selecting control areas, the mechanical transmission relationship between different lining segments is not fully considered, and only local crack areas are treated, ignoring the influence of adjacent segments on crack propagation, resulting in poor control effects. At the same time, the generation of control commands does not take into account multiple factors such as structural safety and material compatibility. The injection precision of SHCC (strain hardening cement-based composite material) is insufficient, making it difficult to fully utilize its advantages of high toughness and crack resistance, and thus failing to achieve effective crack control at its root. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for active control of tunnel lining cracks based on SHCC, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for active control of tunnel lining cracks based on SHCC, the method comprising:

[0007] Dynamic image data of cracks on the lining surface are continuously collected by a sensor array deployed inside the tunnel.

[0008] The dynamic image data is input into the crack semantic analysis module, which calls the crack history archive and tunnel structure database to perform joint reasoning and outputs the crack control target vector.

[0009] Construct a geometric topology network for the tunnel lining, where network vertices correspond to lining segments and network edges represent the mechanical dependencies between segments;

[0010] The crack control target vector is mapped to a geometric topology network, and the influence score of each vertex is calculated by the similarity diffusion algorithm to select the set of key control vertices.

[0011] For the set of key control vertices, a multi-index decision model is run to generate a list of candidate control instructions; the instructions in the list of candidate control instructions are executed to complete the high-precision injection of SHCC materials.

[0012] Preferably, dynamic image data of cracks on the lining surface are continuously collected by a sensor array deployed within the tunnel; the dynamic image data is input into a crack semantic analysis module, which uses a crack history archive and a tunnel structure database for joint reasoning to output a crack control target vector, including:

[0013] Image enhancement algorithms are used to denoise and strengthen edges of dynamic image data to generate standardized crack images; a deep convolutional network is used to extract crack morphology feature vectors from the standardized crack images.

[0014] Retrieve similar crack development patterns from the crack history archive and encode them as historical feature vectors;

[0015] The lining material parameters and stress conditions are obtained from the tunnel structure database to form an engineering feature vector;

[0016] By fusing crack morphological feature vectors, historical feature vectors, and engineering feature vectors through an attention mechanism, a crack control target vector is generated.

[0017] Preferably, a geometric topology network for the tunnel lining is constructed, wherein network vertices correspond to lining segments, and network edges represent the mechanical dependencies between segments, including:

[0018] A lining structure model was built using finite element analysis software, and the lining was divided into multiple segments, each of which was assigned material properties and stress state.

[0019] Initialize the geometric topology network with segments as vertices and contact relationships or possible crack propagation paths between segments as edges; set the initial weights of the edges based on the distance between segments and material similarity.

[0020] When new monitoring data arrives, the weights of the affected edges are adjusted according to the crack expansion, and vertices are added or removed.

[0021] Preferably, the crack control target vector is mapped to a geometric topology network, and the influence score of each vertex is calculated using a similarity diffusion algorithm to select a set of key control vertices, including:

[0022] Graph embedding technique is used to represent each vertex in the geometric topology network as a low-dimensional vector; the crack control target vector is projected onto the same low-dimensional space; the cosine similarity between the crack control target vector and each vertex vector is calculated as the initial influence score;

[0023] Starting from the vertex with the highest initial influence score, the influence score is iteratively propagated to neighboring vertices, with the score decaying during the propagation process; all vertices with scores exceeding the threshold are collected to form a set of key control vertices.

[0024] Preferably, for the set of key control vertices, a multi-indicator decision model is run to generate a list of candidate control instructions, including:

[0025] For each vertex in the set of key control vertices, structural health indicators, crack urgency indicators, and environmental factor indicators are extracted; the entropy weight method is used to assign weights to each indicator, and the comprehensive control urgency score of the vertex is calculated.

[0026] The vertices are sorted according to their scores, and a candidate control instruction list is generated. Each instruction in the list corresponds to a control operation for a vertex.

[0027] Preferably, an image enhancement algorithm is used to denoise and strengthen the edges of the dynamic image data to generate a standardized crack image, including:

[0028] Gaussian filtering is applied to smooth dynamic image data and reduce random noise; the Laplacian operator is used to enhance the contrast of crack edges; the enhanced image is binarized to highlight the crack area; and the image size and resolution are adjusted to a uniform standard.

[0029] Preferably, the lining structure model is processed using finite element analysis software, dividing the lining into multiple segments, each segment being assigned material properties and stress states, including:

[0030] Import the three-dimensional geometric model of the tunnel lining; set the mesh generation parameters to generate a finite element mesh; assign the concrete elastic modulus and Poisson's ratio to each mesh element; apply boundary conditions and loads, and calculate the stress distribution results.

[0031] Preferably, graph embedding techniques are used to represent each vertex in the geometric topology network as a low-dimensional vector, including:

[0032] A random walk strategy is used to generate vertex sequences; a skip character model is used to train the vertex sequences to obtain the initial embedding vector for each vertex; and a graph autoencoder is used to optimize the embedding vectors while preserving the network structure information.

[0033] Preferably, structural health indicators, crack severity indicators, and environmental factor indicators are extracted for each vertex in the set of key control vertices, including:

[0034] The maximum principal stress value at the vertices is read from the finite element analysis results as a structural health indicator; the crack width and length are measured from dynamic image data as a crack hazard indicator; and humidity and temperature data are obtained from environmental sensors as environmental factor indicators.

[0035] Preferably, the present invention also includes an active control system for tunnel lining cracks based on SHCC, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the active control method for tunnel lining cracks based on SHCC as described above.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] By relying on a sensor array to continuously collect dynamic image data of cracks on the lining surface, this method overcomes the limitations of discontinuous and incomplete data acquisition in traditional monitoring. It can capture the dynamic changes of cracks from initiation to expansion in real time, making the perception of crack status more timely and complete. Through the crack semantic analysis module, it calls the crack history archive and tunnel structure database for joint reasoning, realizing the deep fusion of multi-source data. This avoids the one-sided judgment caused by single data processing, so that the output crack control target vector can fully reflect the characteristics of the crack itself and the background information of the tunnel structure, making the control direction more targeted.

[0038] The construction of the geometric topology network, using lining segments as network vertices and the mechanical dependencies between segments as network edges, clearly presents the mechanical transmission path of the tunnel lining structure. This network model can accurately characterize the mutual influence of each segment in the structural stress, providing an analytical basis consistent with the structural mechanics characteristics for subsequent control decisions. After mapping the crack control target vector to this geometric topology network, the influence score of each vertex is calculated using a similarity diffusion algorithm. This allows for the rapid identification of segments that play a key role in crack control, screening out the set of key control vertices, avoiding the problem of blindly selecting control locations in traditional control methods, and concentrating control resources on the core area, thus improving control efficiency.

[0039] The application of a multi-index decision-making model changes the traditional reliance on a single index for generating control commands. By comprehensively considering multiple factors such as crack location, propagation trend, and structural mechanical state, it produces a scientifically sound list of candidate control commands, ensuring a high degree of compatibility between the control commands and the actual crack conditions and structural requirements. Based on this command list, high-precision injection of SHCC material is executed, achieving a precise match between material application and control decisions. This fully leverages the excellent toughness and crack control capabilities of SHCC material, making the material's role more targeted.

[0040] This method forms a complete closed loop, encompassing dynamic crack monitoring, multi-source data fusion and inference, key area screening, scientific command generation, and precise material injection, enabling proactive control of tunnel lining cracks. Moving beyond passive repair after cracks appear, this method intervenes and regulates cracks in their early stages, effectively curbing further expansion and maintaining the integrity and load-bearing capacity of the lining structure. Simultaneously, by precisely screening key control vertices and scientifically planning injection schemes, unnecessary material consumption is reduced, improving resource utilization efficiency. Adaptable to the complex environment and dynamic stress characteristics of tunnels during service, this method can adjust control strategies in real time based on crack dynamic changes, making crack control more flexible and adaptable. This contributes to the long-term stable service of tunnel structures and reduces the cost of disease treatment and safety risks. Attached Figure Description

[0041] Figure 1 This is a flowchart of the active control method for tunnel lining cracks based on SHCC as described in this invention;

[0042] Figure 2 A flowchart for the crack semantic parsing module;

[0043] Figure 3 A flowchart for constructing the geometric topology network of the tunnel lining;

[0044] Figure 4 A trend chart showing the iterative diffusion of the influence score at the vertex.

[0045] Figure 5 Histogram of stress range distribution for finite element mesh elements in tunnel lining. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figure 1This invention provides a method for active control of tunnel lining cracks based on SHCC, the overall implementation scheme of which is as follows:

[0048] An array of sensors, consisting of high-resolution cameras and image sensors, is deployed inside the tunnel to continuously acquire dynamic image data of the lining surface at fixed time intervals, ensuring coverage of areas where cracks may occur. The acquired dynamic image data is transmitted in real time to a crack semantic analysis module embedded in the central processing unit. This module can access a pre-stored crack history archive and tunnel structure database for joint reasoning analysis. The crack history archive stores records of past crack development and repair data, while the tunnel structure database contains information on the lining's material properties, geometric dimensions, and stress conditions. The crack semantic analysis module processes the dynamic image data using internal algorithms, combining historical and structural data to output a crack control target vector. This vector quantifies the current state of the crack and the expected control direction.

[0049] The system constructs a geometric topology network for the tunnel lining, representing the mechanical relationships of the lining in a graph structure. Vertices in the network correspond to segmented units of the lining, each segment representing a physical part of the tunnel lining. Network edges characterize the mechanical dependencies between segments, such as stress transfer or crack propagation paths. The construction process is based on finite element analysis results, ensuring that the network reflects actual structural behavior. Crack control target vectors are mapped to this geometric topology network. A similarity diffusion algorithm is used to calculate the influence score of each vertex, simulating the propagation process of influence within the network. Based on the score, a set of key control vertices is selected; these vertices represent areas requiring priority intervention.

[0050] For the set of key control vertices, the system runs a multi-index decision model. This model integrates multiple indicators such as structural health, crack urgency, and environmental factors, and uses the entropy weight method to dynamically allocate weights, calculating a comprehensive control urgency score for each vertex. Based on the score ranking, the model generates a candidate control instruction list. Each instruction in the list specifies the control operation for a specific vertex, such as the injection location and dosage of SHCC material. The system executes the instructions in the candidate control instruction list, injecting SHCC material into the designated lining segment using a high-precision injection device, achieving active crack control. The entire process operates in a closed-loop manner, with a sensor array continuously monitoring crack changes. Feedback data is used to update inference and decision-making, ensuring the real-time nature and adaptability of the control.

[0051] Example 1: See Figure 2The sensor array deployed inside the tunnel adopts a ring arrangement, with cameras installed at equal intervals on the lining surface. When acquiring dynamic image data, the sampling frequency is set to one frame per second to ensure that subtle changes in the cracks are captured. The dynamic image data is first processed by an image enhancement algorithm, using Gaussian filtering for smoothing. The filter kernel size is set to 3x3, and the standard deviation is 1.0 to effectively reduce random noise in the image. Then, the Laplacian operator is used for edge enhancement. The operator template is in the form of second-order differential to enhance the contrast of the crack area and make the crack outline clearer. The enhanced image is then binarized, and the Otsu thresholding method is used to automatically determine the segmentation threshold, converting the image into a black and white binary image to highlight the crack area. Finally, the image size and resolution are adjusted and standardized to 1024x768 pixels to generate a standardized crack image for subsequent processing. Standardized crack images are input into a deep convolutional network for feature extraction. The deep convolutional network adopts a pre-trained ResNet architecture, removing fully connected layers and retaining convolutional and pooling layers to extract multi-scale features from the image. The convolution operation uses the ReLU activation function, and downsampling is performed stepwise through multiple convolutional layers and max pooling layers, finally outputting a 512-dimensional crack morphology feature vector. The crack morphology feature vector encodes information such as crack length, width, direction, and branching pattern. Simultaneously, the crack semantic parsing module accesses a crack history archive, which stores historical crack data in time-series format. During retrieval, similarity matching is performed based on the current crack morphology features, using the k-nearest neighbor algorithm to find similar crack records, which are encoded into historical feature vectors. The dimensions of the historical feature vectors are consistent with the crack morphology features, containing data on crack development rate and repair effectiveness.

[0052] In some embodiments, the tunnel structure database is integrated into the system backend, storing lining material parameters such as concrete strength and elastic modulus, as well as stress conditions such as ground stress distribution and load history. Database queries are based on lining segment identifiers, and after obtaining relevant parameters, an engineering feature vector is formed. The elements of the engineering feature vector include material property values ​​and stress state indices. The crack semantic parsing module uses an attention mechanism to fuse crack morphological feature vectors, historical feature vectors, and engineering feature vectors. The attention weights are calculated using the softmax function, dynamically adjusting the contribution of each vector to ensure that key information is given priority. After fusion, a crack control target vector is generated. The crack control target vector is a multi-dimensional numerical sequence used to guide subsequent control decisions. The Gaussian filtering in the image enhancement algorithm is implemented in the frequency domain. The image is converted to the frequency domain through a fast Fourier transform, and then inversely transformed back to the spatial domain after applying a Gaussian low-pass filter to ensure uniform noise reduction. Laplacian edge enhancement is performed by direct convolution in the spatial domain, with the enhancement factor set to 2.0 to balance noise amplification. The binarized image undergoes morphological operations, such as opening and closing operations, to eliminate small noise points and connect fractured crack segments.

[0053] Optionally, the deep convolutional network is trained using a transfer learning strategy, fine-tuned based on ImageNet pre-trained weights. Training data comes from a real set of tunnel crack images to ensure the accuracy of feature extraction. The retrieval process of the crack history archive uses Euclidean distance to measure similarity, and historical feature vectors are normalized to avoid scale differences. The query, key, and value vectors in the attention mechanism all come from the input features, and attention scores are calculated through a multilayer perceptron to achieve end-to-end fusion learning. In specific implementation, the circular arrangement of the sensor array ensures full coverage of the tunnel lining, and the camera resolution is set to above 5 megapixels to adapt to different lighting conditions. Dynamic image data is transmitted to the processing unit in real time after acquisition. The image enhancement algorithm is implemented using the OpenCV library, and the frequency domain processing of Gaussian filtering uses the FFT algorithm to optimize the computation speed. After Laplacian edge enhancement, the image contrast is adjusted using a histogram equalization method to further optimize crack visibility. The Otsu thresholding method in the binarization process automatically adapts to the image grayscale distribution, avoiding the deviation of manual threshold setting. The standardization steps include color space conversion to unify the image into grayscale format and reduce data redundancy.

[0054] It is understandable that the ResNet architecture of deep convolutional networks includes a residual block structure to avoid the gradient vanishing problem. During feature extraction, the number of filters in the convolutional layers increases layer by layer, capturing feature representations from low to high levels. The dimension of the crack morphology feature vector is fixed through a global average pooling layer to ensure output consistency. In the k-nearest neighbor algorithm for the crack history archive, the value of k is set to 10, determined based on cross-validation. The historical feature vector encoding includes the crack occurrence time, propagation direction, and repair material type. When extracting engineering feature vectors from the tunnel structure database, the material parameters are based on laboratory test data, and the stress conditions are updated based on finite element analysis results. The multilayer perceptron with attention mechanism contains two hidden layers with 256 and 128 neurons respectively. Dropout regularization is used to prevent overfitting. After the crack regulation target vector is generated, it is normalized by L2 to stabilize the numerical range.

[0055] In practical implementation, the Gaussian filter kernel size of the image enhancement algorithm is adaptively adjusted based on image resolution. For high-resolution images, the kernel size can be increased to 5x5, with corresponding adjustments to the standard deviation. After Laplacian operator enhancement, a noise suppression step is added, and median filtering is used to further smooth non-crack areas. In the morphological operations after binarization, opening operations use 3x3 structuring elements to eliminate isolated points, and closing operations use the same structuring element to connect breakpoints. Standardized image size adjustment uses bilinear interpolation to maintain the image aspect ratio. During the training phase of the deep convolutional network, the learning rate is set to 0.001, the Adam optimizer is used, and the training cycle is 100 epochs. Cosine similarity is used as an auxiliary metric for similarity matching in the crack history archive to improve retrieval accuracy. When forming engineering feature vectors, stress state indices include maximum and minimum principal stresses, and material property values ​​consider the influence of concrete age and are dynamically updated from the database.

[0056] In some embodiments, the fusion process of the attention mechanism includes a feature vector concatenation step, which connects the crack morphology feature vector, historical feature vector, and engineering feature vector into a long vector, which is then input into the attention layer. The attention score is calculated using a scaled dot product attention mechanism, with the scaling factor being the square root of the vector dimension. The output dimension of the crack regulation target vector matches the subsequent network input, typically set to 1024 dimensions to ensure information capacity. In the implementation of the image enhancement algorithm, the frequency domain processing of Gaussian filtering is optimized through a separation filter, applying filtering first to rows and then to columns to reduce computational complexity. The convolution operation of the Laplacian operator uses zero padding to preserve the image size. After feature extraction by the deep convolutional network, the feature vector passes through a batch normalization layer to accelerate training convergence. The data storage of the crack history archive adopts a time-series database structure, supporting fast range queries.

[0057] Optionally, the data acquisition triggering methods for the sensor array include timed triggering and event triggering. Event triggering is based on a motion detection algorithm, initiating acquisition when crack changes exceed a threshold. The parameter settings for the image enhancement algorithm are managed through a configuration file, allowing for on-site adjustments. The inference phase of the deep convolutional network uses TensorRT for acceleration, ensuring real-time processing. The crack history archive update mechanism includes periodic backups and incremental synchronization to prevent data loss. The attention mechanism's training data comes from historical crack cases, and the loss function uses mean squared error to monitor the fusion effect. It can be understood that the noise model in the image enhancement algorithm is based on the Gaussian-Poisson mixture hypothesis, adapting to different sensor characteristics. The threshold adjustment for binarization processing is based on local image characteristics, improved using an adaptive thresholding method. Input image preprocessing for the deep convolutional network includes color correction and illumination normalization to reduce the impact of environmental variations. The index structure of the crack history archive uses a B-tree to optimize query speed. The formation of engineering feature vectors integrates real-time sensor data, such as strain gauge readings.

[0058] In practical implementation, the overall process of the image enhancement algorithm is encapsulated as an independent module. The input is the original dynamic image data, and the output is a standardized crack image. The module execution time is constrained to within 100 milliseconds to meet real-time requirements. The frequency domain implementation of Gaussian filtering is accelerated using CUDA, making it suitable for GPU environments. After Laplacian edge enhancement, image quality is evaluated based on the peak signal-to-noise ratio (PSNR) to ensure enhancement effectiveness. The morphological operation parameters for binarization are determined through grid search, with the optimal structuring element size being 5x5. The ResNet architecture version for the deep convolutional network is ResNet50, balancing accuracy and speed. After feature vector extraction, principal component analysis is applied for dimensionality reduction to visualize the feature distribution. Data augmentation techniques for the crack history archive include crack image rotation and scaling to expand the training set. The normalization of the engineering feature vectors uses min-max scaling to maintain numerical stability.

[0059] It is understandable that the query vector in the attention mechanism comes from the crack morphology feature vector, the key vector comes from the concatenation of historical feature vectors and engineering feature vectors, the value vector is the same as the key vector, and the attention output retains the original information through residual connections. Application scenarios for crack control target vectors include crack classification and regression prediction, supporting multi-task learning. The denoising effect evaluation of the image enhancement algorithm is based on the structural similarity index. The convolution kernel design of the Laplacian operator considers orientation sensitivity, using multi-directional templates to capture cracks with different orientations. The training data annotation for the deep convolutional network is completed by domain experts, and the annotation information includes crack pixel-level masks. The similarity calculation of the crack history archive introduces a weight factor, prioritizing recent data. The update frequency of the engineering feature vector is synchronized with the monitoring data, refreshed hourly to ensure timeliness.

[0060] Example 2: See Figure 3The construction of the geometric topology network for tunnel lining begins with the application of finite element analysis software, such as ANSYS or ABAQUS. This software is used to build the lining structure model. First, the three-dimensional geometric model of the tunnel lining is imported. This model is created based on design drawings or laser scanning data to ensure geometric accuracy. In the finite element analysis software, meshing parameters are set, selecting tetrahedral or hexahedral element types. The mesh size is adaptively determined based on the lining thickness and curvature, controlled between 0.1 meters and 0.5 meters to balance computational accuracy and efficiency. After meshing, each element is considered a lining segment, serving as a vertex of the geometric topology network. Vertex attributes include material properties and stress state. Material properties such as concrete elastic modulus, Poisson's ratio, and tensile strength are directly assigned from the tunnel structure database, while the stress state is calculated through finite element analysis. Finite element analysis applies boundary conditions to simulate the actual stress environment of the tunnel. Boundary conditions include fixed constraints on the contact surface of the surrounding rock, as well as applied in-situ stress and external loads. Load conditions are set according to the tunnel burial depth and traffic load. After calculation, the stress distribution results of each mesh element are obtained, such as the maximum principal stress and strain value. With segments as vertices, the edge initialization of the geometric topology network is based on the contact relationship between segments or the possible crack propagation path. The contact relationship is determined by the node connection in the finite element model, and the crack propagation path is inferred based on the stress concentration area. The initial weight of the edge is calculated based on the Euclidean distance between segments and the material similarity. The closer the distance and the more similar the material properties, the higher the weight value, indicating a stronger mechanical dependence.

[0061] In some embodiments, when new monitoring data arrives from the sensor array, the system dynamically adjusts the geometric topology network. The adjustment process first detects crack propagation, identifies new cracks or the extension of existing cracks by comparing continuous image data, and marks the affected segment vertices. Subsequently, the weights of adjacent edges are updated according to the crack propagation direction. The weight adjustment factor is calculated based on the crack length change rate. If a crack crosses a segment, a new edge is added to represent a new mechanical path. Simultaneously, if a segment is severely damaged by cracks, new vertices may be added to represent subdivided regions, or failed vertices may be deleted. Network maintenance employs a real-time update mechanism to ensure that the topology always reflects the current state of the lining. Mesh generation parameter optimization in finite element analysis is achieved through sensitivity analysis. The mesh size is iteratively adjusted until the stress results converge. The aging effect of concrete is considered when assigning material properties. Parameters are dynamically updated from the tunnel structure database. The initial calculation of edge weights uses the inverse distance weighting method, with the distance weight factor set to 2.0. Material similarity is based on comparing parameters such as elastic modulus using Euclidean distance.

[0062] Optionally, when processing new monitoring data, crack propagation detection uses an image difference algorithm, comparing changes between previous and subsequent frames, and updating the weight decay factor of network edges to 0.9 for smooth transition. Vertex addition or deletion is based on crack area thresholds, and subdivision is triggered when the crack coverage area exceeds 5%. In specific implementation, the finite element analysis software automates operations through a script interface. After importing the 3D geometric model, geometric cleanup is performed to repair gaps and overlaps between facets. In the mesh generation parameter settings, the mesh is refined to 0.1 meters in areas with large curvature, such as arch corners, and widened to 0.5 meters in straight areas. Hexahedral elements are preferred as the element type due to their better numerical stability. Material property assignment is achieved through mapping functions, associating material parameters from the database with mesh elements according to their coordinate positions. Stress state calculation uses static linear analysis, and a direct solver is selected to improve calculation speed. When applying boundary conditions, the surrounding rock contact surface is simulated using surface constraints. Ground stress is uniformly applied to the outer surface of the model in the form of pressure. External live load is simulated as dynamic pressure based on traffic volume. The stress distribution results are output as nodal stress values, and the stress value at the element center point is obtained through interpolation.

[0063] It is understandable that the edge initialization of the geometric topology network is based on the shared surface relationship of the elements in the finite element model. If two elements share a surface or edge, a connecting edge is created. Crack propagation path inference is based on the direction of the maximum principal stress. When the stress direction crosses the element boundary, an edge is created. The initial weight of the edge is calculated by dividing the weight by the material similarity score, which is obtained by comparing the elastic modulus and tensile strength. During the dynamic adjustment of the network, the image difference algorithm for crack propagation detection is assisted by optical flow to improve the recognition accuracy of moving cracks. The weight adjustment factor is proportional to the crack propagation speed; the faster the propagation speed, the larger the weight update amplitude. The addition of new edges must meet the stress continuity condition, i.e., the stress difference between adjacent elements is less than a threshold. Vertex deletion is only performed when the element stress is below the material fatigue limit.

[0064] In practical implementation, the finite element analysis software uses a leading-edge propagation algorithm for mesh generation to produce high-quality meshes. Mesh quality checks are based on the Jacobian matrix to eliminate distorted elements. Anisotropy is considered in material properties. When the tunnel lining is fiber-reinforced, the elastic modulus is assigned along the fiber direction. The rationality of the stress state calculation results is verified, such as checking whether the stress exceeds the material strength limit. The vertex storage structure of the geometric topology network uses an adjacency list for easy and fast traversal of neighboring vertices. Edge weights are initialized and then normalized to ensure that the weights sum to one. When new monitoring data arrives, crack propagation detection and finite element analysis results are cross-validated to ensure that the crack path is consistent with the stress concentration area. Network update operations are logged for backtracking analysis.

[0065] In some embodiments, the fixed constraints applied by the boundary conditions use displacement boundary conditions, setting the displacement of the nodes at the contact surface of the surrounding rock to zero. The in-situ stress loading is applied in stages, first applying a gravity field, then superimposing tectonic stress. External live loads are input in time-series form to simulate vehicle traffic effects. In the initial weight calculation of edges, distance is calculated based on the coordinates of the element center point, and the material similarity score is calculated using cosine similarity; a higher score indicates greater material similarity. The real-time performance of the network dynamic adjustment is ensured through multi-threading technology, with crack detection and network updates executed in parallel. The weight decay factor is dynamically adjusted according to the network size; a smaller decay factor is used for large-scale networks to avoid over-propagation.

[0066] Optionally, the subdivision operation when adding vertices employs a mesh re-division strategy, subdividing the original element into multiple sub-elements. New vertices inherit the material properties of the original vertices, and stress states are obtained through interpolation. After a vertex is deleted, adjacent elements are merged, and the stress of the merged element is recalculated. This can be understood as follows: the load conditions in finite element analysis consider groundwater pressure, simulated using pore pressure elements; material property updates are based on health monitoring data, such as a concrete strength variation model over time; the edge directionality of the geometric topology network can be set to directed edges, representing the stress transmission direction; and the calculation of the weight adjustment factor incorporates a crack width factor, with larger crack widths resulting in higher factor values. The network maintenance mechanism includes periodic global optimization, such as recalculating all edge weights every 24 hours to eliminate accumulated errors.

[0067] In practical implementation, the finite element analysis software exports the calculation results in XML format, including element numbers, coordinates, and stress values. The geometric topology network construction program reads the XML file, automatically creating vertices and edges. Edge initialization rules are configurable, supporting distance thresholds or stress gradient thresholds. The network is stored in a graph database, supporting complex queries. New monitoring data preprocessing includes timestamp alignment to ensure image data corresponds to the finite element analysis time points. Crack propagation detection results are converted into binary masks and mapped to mesh element positions. The weight adjustment factor is calculated using a lookup table method, with a predefined correspondence between crack length change rate and the factor. Vertex addition triggers local remeshing of the finite element model, remeshing only the affected area to reduce computation. After vertex deletion, adjacency relationships are updated, and relevant edge weights are recalculated.

[0068] It is understandable that when comparing material similarity, if the material parameters are vectors, such as anisotropic elastic matrices, then matrix similarity is used as a metric. Edge weight updates consider crack history, assigning higher weights to edges where cracks have previously appeared. The network dynamic adjustment algorithm includes consistency checks to prevent network oscillations caused by noisy data. The selected finite element analysis software supports secondary development, calling mesh generation and solution functions via API interfaces. The construction of the geometric topology network and finite element analysis are iteratively performed, with each new stress result used to optimize the network structure. New monitoring data integrates multi-source information, such as acoustic emission data, to assist in crack propagation judgment.

[0069] Example 3: The process of mapping the crack control target vector to the geometric topology network adopts graph embedding technology to realize network representation learning. The graph embedding technology first uses a random walk strategy to generate a vertex sequence. The walk length of the random walk strategy is set to 100 steps. Starting from each vertex, multiple walks are performed to generate a large number of sequence samples. The random walk parameters include return probability and entry probability. The return probability is set to 0.5 and the entry probability is set to 1.0 to balance local and global structural information. The generated vertex sequence is input into the jump word model for training. The jump word model window size is set to 5, which means that the context of the previous and next 5 vertices is considered. The training efficiency is optimized by negative sampling. The initial embedding vector of each vertex is output. The dimension of the initial embedding vector is set to 128 to capture the adjacency relationship in the network. The initial embedding vector is further optimized through a graph autoencoder. The graph autoencoder structure includes an encoder and a decoder. The encoder is a multi-layer graph convolutional network. Each layer applies graph convolution operations to aggregate neighbor information. The decoder attempts to reconstruct the network's adjacency matrix. The training objective is to minimize the reconstruction error while preserving network structure information such as vertex degree and community structure. The optimized embedding vector serves as a low-dimensional representation of the vertices. At the same time, the crack modulation target vector is projected onto the same low-dimensional space through a linear projection layer. The projection weights are learned through training to ensure compatibility.

[0070] In a low-dimensional space, the cosine similarity between the crack control target vector and each vertex vector is calculated as the initial influence score. The cosine similarity calculation is normalized, and the result ranges from -1 to 1. The absolute value is taken as the score benchmark. The similarity diffusion algorithm iteratively propagates from the vertex with the highest initial influence score. The propagation process adopts a queue structure, propagating the score from the current vertex to all neighboring vertices each time. The propagation attenuation factor is set based on the edge weight, and the attenuation factor is between 0.7 and 0.9 to ensure that the influence of distant vertices is weakened. Iteration continues until the score change is less than a threshold or the maximum number of iterations is reached. All vertices with scores exceeding a preset threshold are collected. The threshold is dynamically adjusted according to historical data to form a set of key control vertices. The random walk strategy in graph embedding employs the node 2vec algorithm, combining breadth-first and depth-first search to balance homogeneity and structural equivalence. The skip-gram model is trained using a Skip-gram architecture with a learning rate of 0.01 and 100 training epochs. The graph autoencoder has 3 encoder layers, with each layer decreasing the output dimension. The final embedding vectors are validated for dimensionality reduction using principal component analysis. The maximum number of iterations for similarity propagation is 100, and the threshold is set based on the quantiles of the score distribution, typically selecting the top 20% of vertices as candidates. The entire mapping process is parallelized and accelerated using GPUs to ensure real-time performance.

[0071] In some embodiments, the random walk strategy is implemented by initializing the starting point for each vertex, fixing the step size, selecting the direction based on the vertex degree probability distribution, controlling the likelihood of returning to the previous vertex, controlling the tendency to explore new vertices, and storing the generated sequence in text format for the skip-word model to read. During the training phase of the skip-word model, the input sequence is processed through a sliding window, with the center word being the current vertex and the context words being other vertices within the window. The negative sampling quantity is set to 5, and negative samples are randomly selected from the vertex set. The loss function is negative log-likelihood, the optimizer is stochastic gradient descent, and L2 normalization is performed after the initial embedding vectors are trained to make the vector magnitude 1. The encoder part of the graph autoencoder uses graph convolutional network layers. Each graph convolution operation is defined as aggregating neighboring vertex features, the activation function is ReLU, the decoder uses inner product operations to reconstruct the adjacency matrix, the reconstruction error is calculated using cross-entropy loss, the training batch size is set to 256, and the training cycle is 50 rounds.

[0072] The projection of the crack control target vector is achieved through a fully connected layer. The input dimension of the fully connected layer matches the original dimension of the crack control target vector, and the output dimension is 128-dimensional, aligned with the vertex embedding vector. The projection weight matrix is ​​learned through backpropagation, and the training data comes from historical crack cases. The loss function is the mean squared error between the projected vector and the true embedding vector. When calculating cosine similarity, the dot product of the two vectors is divided by the product of their magnitudes, and the absolute value of the result is scaled to the range of 0-1 as the initial influence score. The initial influence scores are stored in an array, sorted by vertex index. The similarity diffusion algorithm uses a first-in, first-out (FIFO) queue structure. During initialization, the vertex with the highest initial influence score is enqueued. During propagation, for the current vertex, the edge weights between it and its neighboring vertices are calculated. After weight normalization, these weights are used as the propagation ratio. The new score of a neighboring vertex is the current vertex's score multiplied by the propagation ratio and then by a decay factor, plus the neighboring vertex's original score. This iteration continues until the queue is empty or the score change is less than the threshold of 0.001.

[0073] In some embodiments, the threshold adjustment mechanism is dynamically optimized based on the size of the historical key control vertex set. If the number of vertices in the historical set is too large, the threshold is increased to reduce the number of candidates; conversely, the threshold is decreased. The threshold calculation uses the percentile method, taking the 80th percentile value of the score vector as the benchmark. The propagation attenuation factor is set considering network density. For dense networks, the attenuation factor is smaller, such as 0.7, while for sparse networks, the attenuation factor is larger, such as 0.9. The factor value is determined through grid search. Training data preparation for graph embedding technology includes network topology and vertex attributes, such as vertex degree and clustering coefficient, used to enhance the embedding representation. The word vector dimension of the skip character model is adjustable, and is optimized to 128 dimensions through the validation set. The parallelization of the similarity diffusion algorithm is implemented through OpenMP, with multi-threaded processing of vertex propagation. Iteration count monitoring is set with timeout protection to prevent infinite loops.

[0074] Optionally, repeated visits to vertices are allowed during vertex sequence generation, but the frequency of visits is recorded to avoid excessive skewness. The window size of the skip character model can be dynamically adjusted based on the vertex average degree setting. The graph autoencoder uses the Adam optimizer for optimization, with the learning rate halved every 10 rounds. A non-linear activation function such as tanh can be added to the projection layer of the crack control target vector to improve representational power. Cosine similarity calculation is optimized using vectorization to accelerate computation. The similarity diffusion algorithm updates scores asynchronously, with neighbor vertex scores updated immediately to reduce synchronization overhead. It can be understood that the return and exit probabilities of the random walk strategy are optimized through a validation set to balance network exploration and exploitation. The negative sampling distribution of the skip character model is based on the 3 / 4 power of the vertex frequency to alleviate the dominance of high-frequency vertices. The graph autoencoder uses an early stopping strategy to monitor reconstruction errors and prevent overfitting.

[0075] In practical implementation, the overall process of graph embedding technology is encapsulated as an independent module. The input is the adjacency matrix and vertex attribute matrix of the geometric topology network, and the output is the vertex embedding vector matrix. The module execution time is constrained to within a few minutes, meeting the real-time requirements of engineering projects. The number of walks in the random walk strategy is set according to the network size; for large-scale networks, the number of walks increases to 100 per vertex to ensure sequence coverage. The skip word model training is implemented using the Word2Vec library, with word vectors initialized to random values. After training, the model file is saved. The graph autoencoder is implemented based on deep learning frameworks such as PyTorch. The graph convolutional network layer uses a variant of ChebNet, approximating graph convolution with Chebyshev multinomials to reduce computational complexity. The projection training of the crack control target vector is performed jointly with graph embedding, sharing the underlying representation for multi-task learning. Gradient clipping of the projection layer avoids gradient explosion.

[0076] Vector normalization before cosine similarity calculation ensures numerical stability. The threshold setting interface for the similarity diffusion algorithm provides a slider adjustment, allowing operators to fine-tune it based on experience. The number of iterations in the propagation process is recorded for performance analysis. Visualization of vertex embedding vectors uses the t-SNE method to reduce dimensionality to 2D space to examine clustering effects. The output format of the key control vertex set is a list of vertex IDs with score sorting. In some embodiments, parameter tuning of the graph embedding technique uses Bayesian optimization, with the objective function being embedding quality metrics such as link prediction accuracy. Path sampling for the random walk strategy is accelerated using an aliasing method, and training data augmentation for the skip-word model includes sequence perturbation and noise injection. Projection error analysis of the crack control target vector monitors projection quality through residual calculation. The propagation rule of the similarity diffusion algorithm can be extended to multi-source propagation, simultaneously propagating from multiple high-partition vertices.

[0077] It can be understood that the propagation formula affecting the score in the similarity diffusion algorithm can be expressed as:

[0078]

[0079] in, Represents vertices In the The impact score after the next iteration Specifically refers to the vertex In the The impact score is updated after the next iteration. It is the propagation attenuation factor, with a value ranging from 0.7 to 0.9. It is the vertex The set of neighboring vertices, It is the vertex and vertex Edge weights between them Specifically, it refers to the influence score of a neighboring vertex u of vertex v after the t-th iteration. This is a normalization factor that ensures the propagation weights sum to 1. The formula implements iterative updates of scores; in each iteration, the vertex score retains part of its original value and absorbs part of its neighbor's score, while the attenuation factor controls the propagation strength. The formula symbols are clearly defined. Represents the score. Represents the number of iterations. As the attenuation factor, and Vertex identifier, For edge weights, For the neighbor set, these symbols are uniquely defined in this embodiment and are not repeated in other embodiments.

[0080] In practice, the formula is implemented through iterative iteration. The score array is initialized with initial influence scores, the iteration counter is reset to zero, and each iteration iterates through all vertices, applying the formula to update the scores. Difference calculations check for convergence, and vertices with scores exceeding a threshold are filtered out after the iteration ends. Edge weights in the formula are read from the geometric topology network, and normalization factors are pre-calculated and stored to improve efficiency, while attenuation factors are propagated. It is configurable and can be set via a configuration file. The numerical stability of the formula is maintained through score pruning to prevent overflow, and the iteration process is logged for debugging. The entire process of the similarity diffusion algorithm is integrated into the system as a core component for key control vertex selection.

[0081] See Figure 4 This diagram is the core visualization result of the process of mapping the crack control target vector to the geometric topology network and calculating the vertex influence score through the similarity diffusion algorithm. The diagram uses the iteration number as the horizontal axis and the influence score as the vertical axis, with two curves showing the dynamic changes of the highest and average influence scores: reflecting the reasonable weakening of the influence of core vertices with each propagation iteration, and reflecting the diffusion and convergence process of influence from core vertices to neighboring vertices in the topology network. This change pattern verifies the effectiveness of the similarity diffusion algorithm, ensuring both accurate quantification of the influence of core control vertices and reasonable propagation of influence in the network, providing a quantitative basis for the selection of key control vertex sets. The data originates from the low-dimensional vector representation of topology network vertices using graph embedding technology, the spatial projection of the crack control target vector, and iterative propagation calculation based on edge weights, fully restoring the dynamic evolution logic of influence scores in the tunnel lining geometric topology network. This is a key analytical step in achieving precise positioning of the control area in active crack control.

[0082] Example 4: For the set of key control vertices, a multi-index decision model is run to generate a list of candidate control instructions. The multi-index decision model first extracts three types of indicators for each vertex in the set of key control vertices: structural health indicators, crack hazard indicators, and environmental factor indicators. The structural health indicators are read from the finite element analysis results, including the maximum principal stress value, strain energy, and safety factor of the vertex. The maximum principal stress value directly reflects the stress state of the material. The strain energy integral calculates the energy distribution. The safety factor is derived based on the ratio of material strength to stress. The crack hazard indicators are measured from dynamic image data. Image processing algorithms are used to quantify the crack width and length. The width is converted to physical units through pixel calibration. The length is calculated based on the centerline extracted by the skeletonization algorithm, while also considering the number and orientation of crack branches. The environmental factor indicators are obtained from environmental sensors deployed in the tunnel. The sensors monitor humidity and temperature data. Humidity is recorded as a relative percentage, and temperature is collected in degrees Celsius. The data is stored in time series form. The recent average value and fluctuation variance are calculated during extraction. After the indicators are extracted, they are standardized to eliminate dimensional differences. The standardization method uses min-max scaling to map each indicator value to the 0-1 range. The multi-indicator decision model uses the entropy weight method to assign weights to each indicator. The entropy weight method calculates weights based on the dispersion of indicator values; the higher the dispersion, the greater the weight. In the specific calculation, an indicator matrix is ​​first constructed, with rows corresponding to vertices and columns corresponding to indicator values. The entropy value of each indicator is calculated. The smaller the entropy value, the greater the information content of the indicator and the higher the weight. After the weights are assigned, the comprehensive control urgency score of each vertex is calculated. The scoring formula is a weighted sum. The higher the score, the stronger the control urgency. After the score calculation is completed, the vertices in the key control vertex set are sorted in descending order of score to generate a candidate control instruction list. Each instruction in the list corresponds to a vertex, specifying the control operation type such as SHCC injection amount, injection pressure, and timing.

[0083] The structural health indicators extracted are updated synchronously with the finite element analysis, typically once per hour. The image measurement of the crack critical indicator uses a sub-pixel precision algorithm, with the width measurement error controlled within 0.1 mm. Environmental factor indicators are integrated with real-time data streams, with humidity sensor accuracy at ±2% and temperature sensor accuracy at ±0.5°C. When calculating using the entropy weight method, the indicator matrix is ​​preprocessed to fill missing values, and the entropy value is calculated using the natural logarithm. Weight normalization ensures that the sum is 1. After calculating the comprehensive score, the sorting algorithm uses quicksort. The instruction list is output in JSON format, containing vertex ID, score, and control parameters. The multi-indicator decision model has a configurable running cycle to adapt to different monitoring needs. In practice, structural health indicators are extracted using the post-processing function of finite element analysis software. The stress cloud map of the mesh element corresponding to each vertex is read, the maximum principal stress value is taken as the maximum value of the element, the strain energy is calculated by the integral stress-strain curve, the safety factor is determined based on the ratio of concrete tensile strength to the maximum principal stress, the image processing of crack criticality indicators uses OpenCV library functions, the crack width is measured by sampling multiple points along the crack normal direction and averaging, the length is measured by tracking the crack outline through the chain code algorithm, the number of branches is counted to connect components, the sensor data of environmental factor indicators are collected through the Modbus protocol, the average value is calculated using a sliding window, the window size is set to the data of the most recent 24 hours, and the variance is calculated to reflect the degree of environmental fluctuation.

[0084] Optionally, the minimum-maximum scaling formula for standardized processing is the index value minus the minimum value divided by the difference between the maximum and minimum values. After scaling, the index distribution is uniform. In the entropy weight method, the index matrix rows have a vertex number sequence representing the number of indicators. Before calculating the entropy value, the index values ​​are normalized so that their sum is 1. The weights are equal to 1 minus the entropy value and then normalized. The comprehensive control urgency score is the sum of each index value multiplied by its weight, with a score range between 0 and 1. The generation of the candidate control instruction list also includes instruction priority marking: instructions with a score higher than 0.8 are marked as urgent, those with a score between 0.5 and 0.8 are marked as important, and those with a score lower than 0.5 are marked as normal. After the list is output, it is pushed to the scheduling system for execution. The parameters of the multi-indicator decision model are adjustable, such as the weight calculation window size and score threshold, and are managed through configuration files.

[0085] It is understandable that the strain energy calculation in structural health indicators is based on the stress and strain tensors of finite element elements, multiplied and integrated. The safety factor considers the long-term load effect and introduces a reduction factor. The orientation analysis of the crack urgency index calculates the direction of the crack principal axis, and the correlation analysis with environmental factor indicators can help determine the cause of cracks. The dispersion calculation of the entropy weight method uses standard deviation for verification, and the weight allocation needs to pass a consistency check to ensure logical rationality. The weighted sum calculation of the urgency score of comprehensive control can introduce nonlinear transformations, such as logarithmic processing of extreme values, and the stability of the sorting algorithm is verified by comparison sorting. The control parameters of the instruction list are determined based on linear interpolation of the score values, such as the injection volume score multiplied by the maximum allowable injection volume.

[0086] In practical implementation, the multi-index decision-making model is deployed on a server cluster. The index extraction module communicates with the finite element analysis software and image processing system via API interfaces, and the data format is uniformly JSON or XML to ensure compatibility. The finite element result file parsing for structural health indicators uses a dedicated library such as VTK to extract and average the nodal stress values ​​of the elements corresponding to the vertices. Strain energy calculation is performed at the element level, and the safety factor reduction factor is read from the database based on the concrete age. The image measurement process for crack criticality indicators is automated. After image preprocessing, crack segmentation is performed, followed by geometric feature extraction. The branch count uses a depth-first search algorithm. Database queries for environmental factor indicators optimize the index structure, supporting fast range queries. Incremental algorithms are used for average value calculation to reduce memory usage.

[0087] The entropy weighting method is implemented as an independent module. The input is an index matrix, and the output is a weight vector. Missing values ​​in the matrix are filled using column mean values. Entropy calculation avoids division by zero errors, and weight vector caching improves efficiency. The comprehensive score calculation module processes each vertex in parallel. Weighted sum operations are accelerated using the BLAS library, and sorting is implemented using the standard library's quicksort. The generated instruction list is persistently stored in the database, and version management supports rollback operations. The monitoring log of the multi-index decision model records index values, weights, and scores for later analysis and optimization. Refer to Table 1, the index data table, which shows the index values, weight allocations, and comprehensive score calculation results for some vertices in the key control vertex set.

[0088] Table 1. Vertex Regulation Indicators and Weight Allocation

[0089] Vertex Number Maximum principal stress (MPa) Crack width (mm) Ambient humidity (%) Stress weight Width weight Humidity weight Overall Score V1024 12.5 0.85 78.2 0.35 0.45 0.20 0.78 V2057 8.9 1.20 82.5 0.33 0.48 0.19 0.85 V3098 10.2 0.65 75.8 0.36 0.42 0.22 0.69 V4011 15.1 1.05 80.1 0.38 0.46 0.16 0.82

[0090] In the table, vertex numbers correspond to vertex identifiers in the geometric topology network. Maximum principal stress, crack width, and ambient humidity are the original index values. Stress weight, width weight, and humidity weight are weights calculated using the entropy weighting method. The overall score is a weighted sum. Weight calculation is based on the dispersion of all vertex index values, and the overall score is used to sort and generate an instruction list.

[0091] In some embodiments, structural health indicators may also include displacement and deformation data, with nodal displacement values ​​read from finite element analysis. Crack severity indicators include crack propagation rate, calculated through time-series image differencing. Environmental factor indicators include vibration acceleration, obtained from accelerometers. The entropy weighting method's weight calculation can incorporate expert scoring adjustments, combining objective entropy values ​​with subjective importance. The comprehensive score calculation can be modified to use the TOPSIS method, comparing the closeness to the ideal solution. The control parameters in the instruction list can be refined into material proportions, injection rates, and curing schemes to adapt to different crack types.

[0092] Optionally, real-time performance of indicator extraction is ensured through stream processing frameworks, such as Apache Kafka for sensor data processing, incremental updates of finite element analysis results, and GPU acceleration for image measurements. The entropy weighting matrix construction supports dynamically adding new indicators, with online weight updates. A time decay factor is introduced into the comprehensive score calculation, giving higher weights to recent indicators. The instruction list output interface supports a RESTful API for easy access by other systems. Performance analysis of the multi-indicator decision model includes computational latency and resource consumption, as well as algorithm and resource allocation optimization.

[0093] In practical implementation, the width measurement of the crack criticality index uses edge detection combined with morphological operations to improve measurement robustness; length calculation avoids jagged effects through smoothing; humidity data correction for environmental factor indicators is based on dry-bulb and wet-bulb thermometer comparisons, and temperature data compensates for sensor drift. The entropy weight method avoids logarithmic zero values ​​in entropy calculation by adding a small offset; sensitivity is verified after weight allocation; and the comprehensive score calculation is standardized to prevent overflow. The generation of the instruction list considers resource constraints, such as simultaneously adjusting the upper limit of the number of vertices, and prioritizing resource allocation to vertices with higher scores. The multi-indicator decision model is validated through historical data playback, comparing instructions with actual effects, and continuously optimizing parameters.

[0094] Example 5: When processing the lining structure model using finite element analysis software, the three-dimensional geometric model of the tunnel lining is first imported. The three-dimensional geometric model format is STEP or IGES standard format to ensure compatibility with the finite element analysis software. The import operation is completed through the graphical user interface or script interface of the finite element analysis software. After import, a geometric integrity check is performed to repair any possible missing patches, overlapping edges, or gap errors, ensuring the watertightness of the model. Mesh generation parameters are set, with the mesh type selected as solid elements such as tetrahedrons or hexahedrons. The mesh size is determined based on the lining thickness and radius of curvature. In areas with a large lining thickness and gentle curvature, the mesh size is set to 0.3 meters. In areas with a small lining thickness or drastic curvature changes, such as arch corners and joints, the mesh size is refined to 0.1 meters. The mesh generation algorithm uses the leading edge method or Delaunay triangulation to generate uniform and high-quality mesh elements. Each grid cell serves as a segment of the lining and is assigned material properties. These properties are read from the tunnel structure database and include concrete elastic modulus, Poisson's ratio, density, and compressive strength. The property assignments are based on the spatial coordinates of the grid cell. A mapping function is used to associate the material parameters in the database with each cell. Considering the heterogeneity of the material, different sections of the tunnel may use different grades of concrete, and the material parameters are assigned separately according to the section identifier.

[0095] In practical implementation, the stress state is calculated by applying boundary conditions and loads. Boundary conditions are set with the surrounding rock contact surface as a fixed constraint to simulate the interaction between the tunnel and the surrounding rock mass. Load conditions include self-weight stress, in-situ stress, and external live load. Self-weight stress is achieved by applying gravitational acceleration. In-situ stress is set based on geological exploration report data, using a linear distribution assumption to simulate the original in-situ stress field. External live load is set according to the tunnel's purpose; vehicle loads are considered for highway tunnels, and track loads for railway tunnels. Loads are applied to the inner surface of the lining or corresponding locations in the form of pressure or concentrated force. Finite element analysis is performed using a static mechanics analysis module to calculate the stress distribution results for each mesh element. The output results include stress tensor, strain tensor, and displacement field. In the post-processing stage, the maximum principal stress value of each element is extracted as a key indicator for subsequent geometric topology network construction. Model verification is performed by comparing on-site measured displacement data or strain gauge readings to ensure the reliability of the finite element analysis results. The entire process is automated through script control, reducing manual intervention.

[0096] In some embodiments, the creation of the 3D geometric model is based on tunnel design drawings, drawn using CAD software or reconstructed from 3D laser scanning point cloud data. Unit consistency is checked during model format conversion to avoid scale errors. The element type selection in the mesh generation parameters considers a balance between computational efficiency and accuracy. Hexahedral elements offer high computational accuracy but result in complex mesh generation, while tetrahedral elements are suitable for complex geometries but require a larger number of elements. Mesh quality checks include element Jacobian ratio and warpage; unqualified elements are discarded. When assigning material properties, the elastic modulus and Poisson's ratio are obtained from laboratory material test reports, density is calculated based on the mix proportions, and compressive strength is determined by age, using the 28-day standard value. The assignment operation is completed in batches using the material module of the finite element analysis software.

[0097] Optionally, when applying boundary conditions, fixed constraints can be nodal displacement constraints or surface constraints. Spring elements are used to simulate elastic support at the surrounding rock contact surface, more realistically reflecting the interaction between the surrounding rock and the lining. In-situ stress loading adopts a step-by-step application strategy: first, the in-situ stress field is initialized; then gravity is applied; and finally, external live load is applied. The load values ​​are determined according to design specifications, such as 20 kPa for vehicle loads in highway tunnels. Stress calculation results are output as nodal stresses or element center stresses, saved as contour maps or data files. The maximum principal stress is extracted and stored in CSV format for easy reading by subsequent programs. The measured data for model verification are integrated into the finite element analysis software for back-analysis to correct model parameters.

[0098] In practice, commercial platforms such as ANSYS or ABAQUS are selected for finite element analysis software. Scripts are written using their parametric design languages ​​to automate the entire process from model import to result output. Importing the 3D geometric model utilizes the File>Import menu function or command flow. Geometric cleanup operations include stitching gaps, repairing damaged surfaces, and simplifying small features. Mesh generation parameters are set in the Mesh module, with the global mesh size set to the base value, and local mesh refinement controlled through the surface or volume size functions. Material properties are defined in the EngineeringData module, a material library is created, a concrete material model is created, and parameters such as the elastic modulus and Poisson's ratio are input. Values ​​are assigned by dragging and dropping materials onto the geometry or directly selecting elements.

[0099] Boundary conditions are applied in the Load module. Fixed constraints are selected at the surrounding rock contact surface, with displacement constraints set to zero. In-situ stress is applied as an initial stress field, achieved by defining stress tensor components. Self-weight stress is set in the inertial load module, specifying the direction and magnitude of gravitational acceleration. External live loads are selected as either pressure or force loads based on the load distribution. The solver settings are static analysis type, with either a direct solver or an iterative solver selected. After calculation, the stress contour plot is viewed in the Result module, and the stress data is exported. Model verification involves comparing the displacement calculated by the finite element method with the monitored displacement. If the error exceeds the allowable range, the model parameters are adjusted, and the calculation is repeated.

[0100] Understandably, the leading-edge mesh generation method is suitable for complex geometries, automatically generating high-quality meshes. Delaunay triangulation mathematics ensures good mesh shape, and smooth mesh size transitions avoid stress concentration. Material properties consider the nonlinearity of concrete, defining multilinear elastic models or damage models to more accurately simulate material behavior. The spring stiffness of boundary conditions is calculated using subgrade coefficients to simulate the effect of surrounding rock stiffness. The in-situ stress field is initialized using the equilibrium in-situ stress method to eliminate initial unbalanced forces. Load combinations consider both basic and accidental combinations, combining different load cases according to code requirements. Post-processing of stress results calculates equivalent stresses such as Von Mises stress to assess the material yield state.

[0101] Optionally, the 3D geometric model can be imported from the BIM model, preserving component information. Adaptive meshing technology is used for mesh generation, automatically refining high-gradient regions based on stress gradients. Material property assignment supports temperature-dependent variations, allowing for the definition of temperature-related material parameters. Symmetrical constraints can be applied to boundary conditions to simplify the model and reduce computational load. Loads are considered during the construction phase, simulating the construction process by activating or deactivating loads in stages. Stress results output is linearized with path settings, extracting stress distribution along the wall thickness. Model validation uses sensitivity analysis to assess the impact of parameter variations on the results.

[0102] In practice, the finite element analysis software's processing flow is encapsulated into a standardized workflow. Inputs include a 3D geometric model file and data parameters, while the output is a stress result file. Mesh generation parameters are optimized through parametric studies, comparing stress results under different mesh sizes to select a mesh that balances accuracy and efficiency. The material property database is linked to the tunnel structure database, automatically synchronizing updated material parameters to the finite element model. Boundary conditions and load settings are templated, with predefined load templates for common tunnel types. Stress calculation results are automatically transferred to the geometric topology network construction module, achieving seamless data transfer.

[0103] See Figure 5This figure visualizes the stress distribution results obtained by processing a tunnel lining structure model using finite element analysis (FEM) software. The figure uses stress range as the vertical axis and element count as the horizontal axis, clearly showing the distribution of finite element mesh elements within different stress ranges. This data originates from the FEM workflow: after importing the 3D geometric model of the tunnel lining, meshing is performed, material properties such as the elastic modulus of concrete are assigned to each mesh element, and then boundary conditions and loads are applied to calculate the stress distribution. The figure intuitively reflects the stress concentration in the tunnel lining structure. The 5-10 MPa range has the most elements, indicating that this stress range is the main stress-bearing range of the lining structure; elements above 20 MPa are fewer but present, suggesting that these areas may be high-stress risk zones where cracks are prone to initiation. This information provides crucial structural health indicators for assigning vertex stress states and selecting key control vertices in the subsequent geometric topology network construction, and is one of the core components in the quantitative analysis of structural mechanical state in the active control of tunnel lining cracks based on SHCC.

[0104] In some embodiments, complex tunnel structures such as intersections or shaft areas employ a partitioned meshing strategy, setting different mesh parameters for different regions, and using contact elements at the boundaries to handle interactions. Material properties consider the reinforcement effect of steel reinforcement, embedding steel reinforcement elements into concrete elements or using an equivalent material model. Boundary conditions simulate construction joints or expansion joints, setting contact pairs or connection elements. Load analysis considers seismic effects, performing dynamic time-history analysis. Stress results are assessed using strength theories, such as the maximum tensile stress theory, to determine the risk of concrete cracking.

[0105] Optionally, the finite element analysis software is integrated with Python scripts to automate batch simulations. Mesh generation utilizes advanced size functions to control the mesh growth rate. Material properties are incorporated into a damage-plasticity model to simulate the concrete cracking process. Boundary conditions consider seepage-stress coupling to simulate the influence of groundwater. Temperature loads are considered, and temperature stresses are calculated. Custom scripts are developed for post-processing stress results to automatically extract key indicators. Model validation employs the Monte Carlo method to statistically evaluate model uncertainty.

[0106] In practice, lightweight processing of the 3D geometric model removes detailed features that do not affect the results, improving mesh generation efficiency. After mesh generation, mesh independence verification is performed to ensure convergence of stress results. Material properties consider time effects, defining a creep model to calculate long-term stress. Boundary conditions support long-range displacement constraints to simulate far-field boundary effects. Dynamic load loading uses amplitude curves to simulate varying loads. Stress results are output in a standardized format, supporting various post-processing software. Model validation establishes error evaluation criteria to continuously improve model accuracy.

[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for active control of tunnel lining cracks based on SHCC, characterized in that, The method performs the following operations: Dynamic image data of cracks on the lining surface are continuously collected by a sensor array deployed inside the tunnel. The dynamic image data is input into the crack semantic analysis module, which calls the crack history archive and tunnel structure database to perform joint reasoning and outputs the crack control target vector. Construct a geometric topology network for the tunnel lining, where network vertices correspond to lining segments and network edges represent the mechanical dependencies between segments; The crack control target vector is mapped to a geometric topology network, and the influence score of each vertex is calculated by the similarity diffusion algorithm to select the set of key control vertices. For the set of key control vertices, a multi-index decision model is run to generate a list of candidate control instructions; the instructions in the list of candidate control instructions are executed to complete the high-precision injection of SHCC materials. The crack control target vector is mapped to a geometric topology network, and the influence score of each vertex is calculated using a similarity diffusion algorithm to select a set of key control vertices, including: Graph embedding technique is used to represent each vertex in the geometric topology network as a low-dimensional vector; the crack control target vector is projected onto the same low-dimensional space; the cosine similarity between the crack control target vector and each vertex vector is calculated as the initial influence score; Starting from the vertex with the highest initial influence score, the influence score is iteratively propagated to neighboring vertices, with the score decreasing during the propagation process; all vertices with scores exceeding the threshold are collected to form a set of key control vertices; The propagation formula affecting scores in the similarity diffusion algorithm can be expressed as: ； in, Represents vertices In the The impact score after the next iteration Specifically refers to the vertex In the The impact score is updated after the next iteration. It is the propagation attenuation factor, with a value ranging from 0.7 to 0.

9. It is the vertex The set of neighboring vertices, It is the vertex and vertex Edge weights between them Specifically, it refers to the influence score of a neighboring vertex u of vertex v after the t-th iteration. It is a normalization factor that ensures the propagation weights sum to 1.

2. The active control method for tunnel lining cracks based on SHCC according to claim 1, characterized in that, Dynamic image data of cracks on the lining surface are continuously collected by a sensor array deployed inside the tunnel; the dynamic image data is input into a crack semantic analysis module, which uses a crack history archive and a tunnel structure database for joint reasoning to output a crack control target vector, including: Image enhancement algorithms are used to denoise and strengthen edges of dynamic image data to generate standardized crack images; a deep convolutional network is used to extract crack morphology feature vectors from the standardized crack images. Retrieve similar crack development patterns from the crack history archive and encode them as historical feature vectors; The lining material parameters and stress conditions are obtained from the tunnel structure database to form an engineering feature vector; By fusing crack morphological feature vectors, historical feature vectors, and engineering feature vectors through an attention mechanism, a crack control target vector is generated.

3. The active control method for tunnel lining cracks based on SHCC according to claim 2, characterized in that, Construct a geometric topology network for the tunnel lining, where network vertices correspond to lining segments, and network edges represent the mechanical dependencies between segments, including: A lining structure model was built using finite element analysis software, and the lining was divided into multiple segments, each of which was assigned material properties and stress state. Initialize the geometric topology network with segments as vertices and contact relationships or possible crack propagation paths between segments as edges; set the initial weights of the edges based on the distance between segments and material similarity. When new monitoring data arrives, the weights of the affected edges are adjusted according to the crack expansion, and vertices are added or removed.

4. The active control method for tunnel lining cracks based on SHCC according to claim 3, characterized in that, For the set of key control vertices, a multi-indicator decision model is run to generate a list of candidate control instructions, including: For each vertex in the set of key control vertices, structural health indicators, crack urgency indicators, and environmental factor indicators are extracted; the entropy weight method is used to assign weights to each indicator, and the comprehensive control urgency score of the vertex is calculated. The vertices are sorted according to their scores, and a candidate control instruction list is generated. Each instruction in the list corresponds to a control operation for a vertex.

5. The active control method for tunnel lining cracks based on SHCC according to claim 4, characterized in that, Image enhancement algorithms are used to denoise and enhance edges of dynamic image data, generating standardized crack images, including: Gaussian filtering is applied to smooth dynamic image data and reduce random noise; the Laplacian operator is used to enhance the contrast of crack edges; the enhanced image is binarized to highlight the crack area; and the image size and resolution are adjusted to a uniform standard.

6. The active control method for tunnel lining cracks based on SHCC according to claim 5, characterized in that, The lining structure model was processed using finite element analysis software, dividing the lining into multiple segments. Each segment was assigned material properties and stress states, including: Import the three-dimensional geometric model of the tunnel lining; set the mesh generation parameters to generate a finite element mesh; assign the concrete elastic modulus and Poisson's ratio to each mesh element; apply boundary conditions and loads, and calculate the stress distribution results.

7. The active control method for tunnel lining cracks based on SHCC according to claim 6, characterized in that, Graph embedding techniques are used to represent each vertex in a geometric topology network as a low-dimensional vector, including: A random walk strategy is used to generate vertex sequences; a skip character model is used to train the vertex sequences to obtain the initial embedding vector for each vertex; and a graph autoencoder is used to optimize the embedding vectors while preserving the network structure information.

8. The active control method for tunnel lining cracks based on SHCC according to claim 7, characterized in that, For each vertex in the set of key control vertices, structural health indicators, crack severity indicators, and environmental factor indicators are extracted, including: The maximum principal stress value at the vertices is read from the finite element analysis results as a structural health indicator; the crack width and length are measured from dynamic image data as a crack hazard indicator; and humidity and temperature data are obtained from environmental sensors as environmental factor indicators.

9. A tunnel lining crack active control system based on SHCC, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the active control method for tunnel lining cracks based on SHCC as described in any one of claims 1 to 8.

Images