Computer vision-based surface damage detection method for key structural components of a shield machine

Abstract

The present application relates to a computer vision-based shield machine key structural part surface damage detection method. First, the high-definition surface image and the working condition label of the structural part are obtained, the semantic damage topology data carrying the crack, corrosion and wear morphology characteristics are generated combined with the field expert interactive labeling, and the vectorization and space normalization processing are carried out. Further, the standardized damage atlas containing multi-modal node descriptor is constructed, the node feature aggregation and edge weight adaptive learning are realized through a three-layer graph convolution network, and the graph embedding vector representing the global damage semantics is output. Through cross-graph comparison learning and joint loss optimization, the decoupling ability of the model to different damage type characteristics is improved, and efficient collaborative classification is realized. The method effectively improves the damage identification accuracy and generalization ability of the shield machine structural part in the complex stratum environment.

Description

Technical Field

[0001] This invention relates to the field of tunnel boring machine (TBM) structural health monitoring and computer vision intelligent detection technology, and in particular to a method for detecting surface damage of key structural components of TBMs based on computer vision. Background Technology

[0002] Currently, in the field of tunnel boring machine (TBM) structural health monitoring and surface damage identification, automated detection technology based on computer vision is widely used to reduce the labor costs of manual inspections and improve the visual management capabilities for the safe operation of structural components. The mainstream technical approach uses convolutional neural networks (CNNs) as the core, training models with large-scale surface image data to achieve automatic identification and classification of typical damages such as cracks, corrosion, and wear. Existing solutions generally adopt a two-stage or multi-stage architecture of target detection, segmentation, and classification, dividing the damage type identification process into "pixel-level mask segmentation—target region aggregation—category discrimination," and emphasizing the role of pixel-level supervisory data and the scale of deep neural network parameters in improving classification accuracy.

[0003] In recent years, academic and engineering research on damage detection of industrial structural components has seen three important development trends: First, collaborative identification of multiple damage types has become an industry necessity, especially given the diverse working conditions and complex damage morphologies encountered during the service of tunnel boring machines. The model's ability to distinguish between cracks, corrosion, and wear directly affects the accuracy and safety of equipment operation and maintenance. Second, small-sample and weakly supervised learning are gradually replacing the development paradigm that relies solely on large-scale data annotation. New methods attempt to improve the generalization robustness and applicability flexibility of models through mechanisms such as transfer learning, self-supervision, or contrastive learning. Third, the demand for edge deployment and energy efficiency optimization is increasing. Existing large-scale model architectures often fail to meet the practical requirements of real-time inference and rapid adaptation when deployed in the field due to limited computing resources.

[0004] Current representative technologies mainly employ segmentation-classification combinations based on fully convolutional networks (such as U-Net and DeepLab). These methods utilize pixel-level segmentation masks and manually defined regional feature extraction to automate the analysis of surface damage areas on structural components such as the main beam and cutterhead flanges of tunnel boring machines (TBMs). These methods are suitable for situations where the target damage is clear, the operating conditions are uniform, and the data acquisition environment is consistent, achieving high accuracy. However, under complex operating conditions, strong lighting variations, scarce damage samples, and overlapping damage types, existing models often suffer from insufficient generalization ability and high false positive and false negative rates. Furthermore, some studies introduce multi-level feature fusion, attention mechanisms, or multi-task learning structures, which can improve performance to some extent, but these significantly rely on pixel-level annotation and costly scene customization, limiting their application in the diverse physical forms of damage and unevenly distributed data environments of TBMs.

[0005] The existing methods have the following prominent technical shortcomings: First, there is a lack of effective physical and semantic relationship modeling for collaborative identification of multiple damage types. Most methods focus on modeling each type of damage individually, failing to utilize the inherent connections between damages at the spatial, textural, and mechanical levels, resulting in poor adaptability to unknown combinations of damage types or new working conditions.

[0006] Secondly, high-performance models often rely on a large amount of pixel-level labeled data. The preparation of large-scale, high-precision data in the actual operating environment of tunnel boring machines is extremely difficult and costly. The model performance deteriorates rapidly under small sample, weak supervision, or unlabeled conditions, making it difficult to support subsequent continuous self-learning and rapid adaptation.

[0007] Third, mainstream convolutional architectures have limited effectiveness in handling local occlusion, rust coverage, and complex backgrounds, making it difficult to flexibly express high-level semantic features such as the geometric connectivity of cracks, the spatial diffusion of corrosion, and the force transmission of wear.

[0008] Fourth, existing methods generally have a large number of parameters and high inference computing power consumption, which are not suitable for the low power consumption, lightweight and high real-time scenarios of edge computing equipment at shield tunneling sites, thus limiting their large-scale deployment in actual shield tunneling operation and maintenance safety monitoring systems.

[0009] Furthermore, current solutions mostly focus on fine-tuning the model architecture, data cleaning, and expanding the labeled content, failing to fundamentally reshape the classification generalization problem by utilizing damaged physical attribute ontology. Summary of the Invention

[0010] This application provides a computer vision-based method for detecting surface damage to key structural components of tunnel boring machines, aiming to solve one of the problems or issues of the existing technology mentioned in the background section.

[0011] The surface damage detection method for key structural components of tunnel boring machines based on computer vision provided in this application specifically includes: S1: Obtain high-resolution surface images and corresponding condition labels of key structural components of the tunnel boring machine under different service conditions, and generate original damage semantic annotation data based on the semantic connectivity regions drawn interactively by domain engineers. S2: The original damage semantic annotation data is abstracted and mapped to construct a set of multi-type damage subgraph structures; S3: Based on the set of multi-type damage subgraph structures, perform normalized scale mapping and spatial coordinate alignment operations to generate a standardized damage semantic topology map carrying multimodal local descriptors; S4: Using the standardized damage semantic topology map as input, a global graph embedding vector is generated based on a preset graph neural network model; S5: Based on the global graph embedding vector, perform cross-graph comparison learning, construct positive sample pairs between subgraphs of the same type of damage and negative sample pairs between subgraphs of different types of damage, and generate a multi-type damage collaborative classification model by jointly optimizing graph structure similarity constraints and node feature consistency constraints; S6: Extract the semantic topology map of the damage to be tested from the image of the structural component of the tunnel boring machine to be inspected, and input the semantic topology map of the damage to be tested into the multi-type damage collaborative classification model, and output the joint confidence distribution result containing multiple types of damage; S7: Based on the joint confidence distribution results, perform a graph structure integrity assessment. Based on the assessment results, generate a low confidence warning signal and mark areas recommended for manual review, or output the final damage classification judgment conclusion.

[0012] S8: Based on the final damage classification judgment and low confidence warning signal, dynamically update the damage evolution records in the shield machine structural health monitoring database, and fine-tune and optimize the adaptive edge weights of the multi-type damage collaborative classification model according to the newly entered working condition labels and damage true values, so as to achieve iterative improvement of the model's generalization ability under complex working conditions.

[0013] The surface damage detection method for key structural components of tunnel boring machines based on computer vision provided in this application has the following beneficial effects: (1) By constructing a semantic topological map based on physical damage morphology and using a lightweight graph convolutional network for structured understanding, this scheme effectively overcomes the high dependence of traditional visual detection methods on pixel-level segmentation annotation and large-scale supervision data. It achieves accurate identification and classification of three typical structural damages: cracks, corrosion, and wear, without the need for dense box selection or mask annotation. Compared with existing CNN-based target detection or semantic segmentation models, which are prone to false positives and false negatives in complex lighting, corrosion interference, and local occlusion scenarios, this invention starts from the geometric topology and physical evolution path of damage. It models cracks as a centerline graph structure with directional and curvature characteristics, represents corrosion as the spatial adjacency relationship of pore gradient connected domains, and abstracts wear as a pressure attenuation path distributed along the force trajectory. It fully preserves the essential morphological characteristics and spatial evolution law of various types of damage, significantly improves the generalization ability and discrimination robustness of the model under non-ideal imaging conditions, and can still maintain stable output, especially in harsh service environments such as water-rich soft soil and highly abrasive rock layers.

[0014] (2) The introduction of hierarchical graph convolution and cross-graph contrastive learning mechanism enables the model to autonomously mine semantic distinction boundaries and potential association patterns between multiple types of damage without pixel-level supervision signals, solving the optimization bias problem caused by label noise and class imbalance in traditional multi-task learning. By constructing positive sample pairs between subgraphs of the same type of damage and negative sample pairs between subgraphs of different types of damage, and jointly optimizing graph structure similarity and node feature consistency constraints, the model not only strengthens its ability to perceive the inherent consistency of similar damage, but also discovers the potential coupling relationship between corrosion and wear in the oxide film evolution process, thereby achieving a deeper understanding of the damage mechanism. Combining adaptive edge weight learning and sparse tensor computation architecture, the number of parameters in the entire inference process is controlled within 1.2 million, supporting hundreds of milliseconds of response on embedded GPU devices, meeting the stringent requirements of low latency, low power consumption and edge deployment for tunnel boring machine on-site inspection, while avoiding computational redundancy and training instability caused by the dual-branch network structure.

[0015] (3) An end-to-end closed-loop process of graph generation, inference, and calibration is designed, integrating a fast edge-guided topology skeleton extraction engine and a graph structure integrity assessment module, further enhancing the system's automation level and decision reliability. After inputting the original image, the system can automatically complete damage map construction, graph embedding learning, and joint confidence output, and identify low-quality map regions through a connectivity dynamic monitoring mechanism. When the connectivity of a certain type of damage sub-graph is lower than a preset threshold, an early warning prompt is automatically triggered, guiding manual review of key parts, effectively preventing high-risk misjudgments caused by image blurring or severe occlusion. This mechanism significantly reduces the trust cost of the model output results for maintenance personnel and improves the interpretability of the diagnostic system and the efficiency of human-machine collaboration. The overall solution eliminates the dependence on text prompt engineering and external knowledge bases, has good scalability and rapid iteration capabilities, and can flexibly adapt to newly emerging damage patterns or new structural component types, providing efficient, accurate, and sustainable technical support for the full life cycle health monitoring of key components of tunnel boring machines.

[0016] The synergistic effect of the above-mentioned technologies has formed an intelligent damage identification system with "physical map-driven, structural semantic understanding, lightweight real-time reasoning, and dynamic reliable calibration" as its core. This system not only significantly reduces the model's dependence on labeled data and high-performance hardware, but also achieves a systematic improvement in terms of adaptability to complex working conditions, system deployment flexibility, and the reliability of diagnostic results, effectively meeting the core needs of intelligent operation and maintenance of major equipment in fields such as rail transit and underground engineering. Attached Figure Description

[0017] Figure 1 This is the main flowchart of a computer vision-based method for detecting surface damage to key structural components of tunnel boring machines. Figure 2This is a sub-flowchart of a computer vision-based method for detecting surface damage to key structural components of tunnel boring machines. Figure 3 This is another sub-flowchart of a computer vision-based method for detecting surface damage to key structural components of tunnel boring machines. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0020] like Figure 1 As shown, this application provides a computer vision-based method for detecting surface damage to key structural components of tunnel boring machines, specifically including: S1: Obtain high-resolution surface images and corresponding condition labels of key structural components of the tunnel boring machine under different service conditions, and generate original damage semantic annotation data based on the semantic connected regions interactively drawn by domain engineers. This data includes crack centerline topological chains with directional and curvature attributes, corrosion connected domain boundaries characterizing porosity gradients, and wear gradient grayscale bands distributed along the force trajectory.

[0021] S2: The original damage semantic annotation data is abstracted and mapped to construct a set of multi-type damage subgraph structures with geometric continuity and texture evolution. Specifically, crack bifurcation points and endpoints are defined as nodes, crack segments are defined as edges with curvature weights, pit centers are defined as nodes, diffusion adjacency relationships are defined as edges, stress concentration points are defined as nodes, and wear path directions are defined as edges, thus obtaining the set of multi-type damage subgraph structures.

[0022] S3: Based on the set of multi-type damage subgraph structures, perform normalized scale mapping and spatial coordinate alignment operations to generate a standardized damage semantic topology map carrying multimodal local descriptors. Specifically, based on the set of multi-type damage subgraph structures, perform normalized scale mapping and spatial coordinate alignment operations, and extract the microstructure gradient histogram, edge direction entropy, local contrast variation coefficient, and surface roughness estimate of each node to obtain the standardized damage semantic topology map.

[0023] S4: Using the standardized damage semantic topology graph as input, construct a graph neural network model containing three hierarchical graph convolution modules. The first layer is used to perform the first-level graph convolution aggregation operation; the second layer is used to introduce an adaptive edge weight learning mechanism, and perform the second-level graph convolution propagation processing based on the edge weights generated by the mechanism to output the semantic similarity between different types of damage; the third layer is used to perform the third-level graph pooling operation to generate the global graph embedding vector.

[0024] S5: Based on the global graph embedding vector, perform cross-graph comparative learning to construct positive sample pairs between subgraphs of the same type of damage and negative sample pairs between subgraphs of different types of damage. By jointly optimizing graph structure similarity constraints and node feature consistency constraints, a multi-type damage collaborative classification model with cross-damage type feature decoupling capability is generated.

[0025] S6: Extract the semantic topological map of the damage to be tested from the image of the tunnel boring machine structural component to be inspected, and input it into the multi-type damage collaborative classification model to output the joint confidence distribution result containing multiple types of damage. Specifically, the multi-type damage includes three types of damage: cracks, corrosion, and wear. Input the image of the tunnel boring machine structural component to be inspected into the map generation engine, and use fast edge-guided topological skeleton extraction processing to reconstruct the semantic topological map of the damage to be tested in the current scene in real time.

[0026] S7: Based on the joint confidence distribution results, perform a graph structure integrity assessment. According to the assessment results, generate a low-confidence warning signal and mark areas recommended for manual review, or output the final damage classification conclusion. Specifically, calculate the connectivity of various damage subgraphs in the joint confidence distribution results; if the connectivity of a certain type of damage subgraph is determined to be lower than a preset threshold, generate the low-confidence warning signal and mark areas recommended for manual review; otherwise, output the final damage classification conclusion.

[0027] S8: Based on the final damage classification judgment and low confidence warning signal, dynamically update the damage evolution records in the shield machine structural health monitoring database, and fine-tune and optimize the adaptive edge weights of the multi-type damage collaborative classification model according to the newly entered working condition labels and damage true values, so as to achieve iterative improvement of the model's generalization ability under complex working conditions.

[0028] Step S1: Obtain high-resolution surface images and corresponding condition labels of key structural components of the tunnel boring machine under different service conditions. Based on the semantically connected regions interactively delineated by domain engineers, generate original damage semantic annotation data containing crack centerline topological chains with directional and curvature attributes, corrosion connected domain boundaries characterizing porosity gradients, and wear gradient grayscale bands distributed along the stress trajectory. Specifically, this includes: S1.1: High-definition surface images of key structural components such as the main beam, cutterhead flange, and propulsion cylinder support of the tunnel boring machine under different service cycles are acquired and processed, and working condition labels such as water-rich soft soil tunneling, soft upper and hard lower strata, and highly abrasive rock strata are simultaneously associated to generate a raw image dataset with environmental context information.

[0029] An active vision acquisition unit consisting of an industrial-grade high-resolution area array camera and a linear laser light source is deployed to perform multi-view scanning of the surfaces of key structural components such as the main beam of the tunnel boring machine, the cutterhead flange, and the propulsion cylinder support, and to acquire raw RGB image data streams.

[0030] The PLC real-time data interface of the tunnel boring machine's main control system is synchronously connected to extract the current tunneling ring number, propulsion speed, cutterhead torque and geological parameters of the strata, and construct an environmental context vector containing multi-dimensional working condition labels such as water-rich soft soil, soft upper and hard lower layers and highly abrasive rock strata.

[0031] The original image is subjected to illumination normalization processing based on histogram equalization to eliminate grayscale distortion caused by non-uniform illumination in the tunnel, and bilinear interpolation is used to resample the image resolution to a preset standard size to ensure spatial scale consistency.

[0032] A unique mapping index between image file identifiers and working environment context vectors is established, and each frame of high-definition surface image is strongly associated with its corresponding service cycle stage and geomechanical state through a timestamp alignment mechanism.

[0033] The bound image-condition pairing data is serialized and stored to generate a raw image dataset with complete environmental context information, providing a physically interpretable input basis for subsequent damage semantic annotation.

[0034] By using the above-mentioned multi-source data fusion and standardized preprocessing methods, discrete visual signals and operating parameters are transformed into structured raw image datasets, achieving spatiotemporal alignment between detection data and environmental variables, and significantly improving the model's generalization and adaptability under complex operating conditions.

[0035] S1.2: Based on the original image dataset, perform interactive annotation operations by domain experts, and use graphical annotation tools to delineate the directional centerline of the crack, the boundary of the connected domain with porosity gradient of the corrosion, and the gradient grayscale band distributed along the force trajectory of the wear, so as to generate an original semantic connected region set containing the geometric features of the three types of damage.

[0036] Load the original image dataset with working condition labels into the graphical annotation interface, and initialize the dedicated annotation layers and semantic attribute templates for the three types of damage: cracks, corrosion, and wear.

[0037] For crack damage, domain experts draw a directional centerline skeleton along the crack direction. The system captures the mouse trajectory coordinate sequence in real time and records the vector direction angle from the starting point to the ending point, forming linear annotation data with topological connectivity.

[0038] For corroded areas, experts delineate the outer closed boundary of the pit cluster. The annotation tool automatically calculates the gray-level distribution histogram of pixels inside the boundary, extracts the porosity gradient feature value and associates it with the connected domain object, generating region annotation data containing texture evolution information.

[0039] To detect wear marks, experts traced the central axis of a gradient grayscale band along the force trajectory. The system simultaneously sampled the grayscale change rate within a preset width on both sides of the axis, constructed a grayscale gradient field mapping relationship distributed along the path, and formed a band-shaped labeled data characterizing the evolution of mechanical wear.

[0040] The system performs spatial overlap detection and semantic conflict verification on three types of labeled data to ensure that labeled regions of different damage types are mutually exclusive at the pixel level or merged according to preset priorities, thereby eliminating label ambiguity.

[0041] The verified annotation data is encapsulated into a set of original semantically connected regions containing geometric coordinates, orientation attributes, texture features, and working condition context, and stored in structured JSON format.

[0042] Through the interactive annotation and feature extraction processing methods described above, the unstructured original image is transformed into a set of original semantic connected regions containing three types of damage geometric features, realizing the initial mapping of damage physical form to semantic topological data, and providing a high-fidelity input foundation for subsequent vectorization conversion.

[0043] S1.3: The original semantic connected region set is vectorized and transformed, the crack center line is abstracted into a sequence of line segments with curvature properties, the boundary of the corrosion connected region is abstracted into a closed polygonal outline, and the wear gradient gray band is abstracted into a band-shaped region with a gray gradient field, so as to generate a damage morphology vector description set with clear geometric definition.

[0044] Receive the original semantically connected region set generated by S1.2, which includes the crack centerline pixel coordinate sequence, the boundary pixel set of the corrosion region, and the gray-level distribution matrix of the wear region.

[0045] B-spline curve fitting is performed on the pixel sequence of the crack centerline to transform discrete pixels into a continuous parameterized curve. The second derivative of each sampling point on the curve is calculated to quantify the local curvature and generate a vector description of the crack line segment with curvature properties.

[0046] The Douglas-Puk algorithm is used to thin out key points in the crack vector description, retaining endpoints and high curvature inflection points to form a simplified crack topology skeleton sequence, ensuring that the geometric feature retention rate is higher than 95% while reducing data redundancy.

[0047] Polygon approximation is performed on the boundary pixel set of the eroded connected domain. Vertex detection is used to identify boundary corners and extreme points, a closed polygon contour vector is constructed, and the area-to-perimeter ratio of the polygon is calculated to characterize the compactness of the eroded region.

[0048] Gradient field calculations are performed on the wear gradient grayscale band region. The Sobel operator is applied to extract the grayscale change rate in the horizontal and vertical directions, and gradient amplitude map and direction map are synthesized to construct a two-dimensional grayscale gradient field vector description that reflects the wear depth.

[0049] Through the above vectorization transformation process, the pixel-level semantically connected regions of the previous step are transformed into a set of damage morphology vector descriptions with clear geometric definitions and physical properties, realizing the mapping from image space to geometric feature space and providing standardized input for subsequent topology network construction.

[0050] S1.4: Based on the damage morphology vector description set, perform a topology connection relationship construction operation, extract the crack bifurcation points and endpoints as connection nodes, the pit center as diffusion nodes, and the stress concentration points as wear path nodes, and establish the adjacency logic between nodes to generate a preliminary damage space topology connection network.

[0051] Receive the damage morphology vector description set generated by S1.3. This dataset contains the crack centerline segment sequence, the closed polygon of the corrosion connected domain, and the gray-scale gradient region of the wear. Perform topology node extraction processing on the crack centerline segment sequence, traverse the coordinates of all segment endpoints and intersections, use the curvature change rate threshold to filter out the curvature extrema points as bifurcation nodes, and mark the ends of unconnected segments as endpoint nodes to generate a crack topology node set.

[0052] Based on the set of crack topology nodes, an adjacency logic is constructed. The Euclidean distance between any two nodes is calculated. If the distance is less than a preset connectivity threshold and there is an original vector line segment connection between the two points, a directed edge connection is established to form a crack space topology subnet.

[0053] Perform centroid calculation on the closed polygon of the eroded connected domain, apply polygon centroid processing to calculate the geometric center coordinates of each independent eroded region, define them as erosion diffusion nodes, and generate a set of eroded region nodes.

[0054] Based on the set of nodes in the corrosion region, a diffusion adjacency determination is performed, the Euclidean distance between each corrosion node is calculated, and a maximum diffusion radius threshold is set according to the physical model of corrosion spread. Node pairs with a distance less than the threshold are connected by undirected edges to form a corrosion spread topology subnet.

[0055] Stress feature point mapping is performed on the wear gradient grayscale band region. Local gradient modulus maxima are tracked along the grayscale gradient direction. Wear start point, end point and intermediate stress concentration point are identified and defined as wear path nodes to generate a wear stress node set.

[0056] Based on the set of wear stress nodes, a path connection relationship is constructed. According to the force trajectory vector, adjacent stress nodes are connected sequentially in time or space order, and edge direction attributes are assigned to form a wear evolution topology subnet.

[0057] The node and edge information of the three subnets of crack, corrosion and wear are registered in a unified spatial coordinate system and merged to generate a preliminary damage spatial topology connection network.

[0058] By constructing the above-mentioned topological connection relationship, the vector description of the previous step is transformed into damage space topological connection network data with node-edge structure, realizing the transformation of damage geometry into graph structure data, and providing a structured input foundation for subsequent graph neural network modeling.

[0059] S1.5: Perform structured encapsulation processing on the damage space topology connection network, integrate working condition labels, damage type truth values ​​and topology connection information, and output the original damage semantic annotation data according to the preset data format, which includes the crack centerline topology chain with directionality and curvature attributes, the corrosion connected domain boundary characterizing the porosity gradient, and the wear gradient grayscale band distributed along the force trajectory, for use in subsequent abstract mapping processing.

[0060] Step S2: The original damage semantic annotation data is abstracted and mapped. Crack bifurcation points and endpoints are defined as nodes, crack segments are defined as edges with curvature weights, pit centers are defined as nodes, diffusion adjacency relationships are defined as edges, stress concentration points are defined as nodes, and wear path directions are defined as edges, thereby constructing a set of multi-type damage subgraph structures with geometric continuity and texture evolution. Specifically, this includes: S2.1: Extract skeleton nodes from the crack centerline topology chain in the original damage semantic annotation data, and use curvature extremum detection to identify the bifurcation points and endpoint coordinates on the line to generate a set of crack topology nodes containing spatial location information.

[0061] The system receives the crack centerline topology vector data generated in S1.3, which contains a sequence of discrete coordinate points arranged in pixel order. B-spline curve fitting is performed on the discrete coordinate point sequence, and the control point positions are optimized using the least squares method. This converts the original polyline annotations into smooth parametric curves with continuous first and second derivatives, eliminating high-frequency jitter noise introduced by manual drawing.

[0062] The curvature values ​​of each sampling point are calculated based on the fitted smooth curve. The rate of change of the tangent vector is solved using the discrete difference method to construct the curvature distribution function along the crack direction. The curvature distribution function is traversed through the sliding window extreme value search process to identify local curvature maxima. Maximal points exceeding a preset threshold are identified as crack bifurcation points or abrupt turning point feature points.

[0063] The topological connectivity of the curve endpoints is detected, and the first endpoint without adjacent predecessor nodes and the last endpoint without adjacent successor nodes are extracted and marked as crack termination or start feature points. The set of curvature extrema points and the set of endpoints are merged, and redundant and duplicate points with a spatial distance less than the pixel resolution are removed to generate a set of crack topological nodes with unique identifiers.

[0064] By using curvature extremum detection and endpoint extraction, the continuous vector lines from the previous step are transformed into discrete key geometric node data, realizing the mapping of crack morphology from linear description to graph structure node definition, laying the topological foundation for the subsequent construction of a graph neural network with weighted edges.

[0065] S2.2: Based on the set of crack topology nodes, perform an adjacency relationship construction operation, use a piecewise linear fitting method to connect adjacent nodes to form crack segments, and calculate the curvature parameter of each crack segment as curvature weight to generate a set of directed crack edges with weight attributes.

[0066] Based on the set of crack topology nodes generated by S2.1, spatial proximity search of adjacent nodes is performed, potential connection pairs between nodes are determined according to the principle of minimum Euclidean distance, and the initial adjacency matrix of the crack skeleton is constructed.

[0067] For each pair of connected nodes in the initial adjacency matrix, extract their discrete pixel sequence in the original image coordinate system, and use the least squares method to perform piecewise linear fitting to generate a smooth geometric path for the crack centerline, thus eliminating the jagged noise interference caused by manual annotation.

[0068] Along the fitted crack centerline path, the rate of change of the tangent vector direction angle is calculated using a fixed step size sampling method. The degree of curvature of each line segment element is quantified using the discrete curvature calculation formula, where the formula for calculating curvature K is:

[0069] in and The second derivatives of the coordinate points are respectively, (x i ,y i ) represents the x and y coordinates of the i-th sampling point on the crack centerline.

[0070] The calculated curvature values ​​are normalized to the [0,1] interval and used as edge weight attributes to characterize crack propagation resistance and stress concentration. The corresponding directed edges are then assigned to form a weighted crack topology edge set.

[0071] By using the above processing method, the discrete nodes of the previous step are transformed into a weighted directed graph structure containing the evolution law of geometric morphology, so as to achieve the expected technical effect of mapping crack damage from geometric description to topological semantics.

[0072] S2.3: Perform centroid localization processing on the boundaries of erosion connected domains in the original damage semantic annotation data, and use morphological centroid calculation to locate the geometric center of each independent pit region, generating a set of erosion region nodes that characterize the corrosion distribution density.

[0073] Read the original damage semantic annotation data containing the boundaries of eroded connected components from the output of S1.5, and extract the set of pixel coordinates and gray-level distribution matrix within the closed polygon contour.

[0074] A binary mask is constructed by performing a binary mask on the extracted set of pixels in the eroded region, assigning a value of 1 to the pixels inside the contour and a value of 0 to the pixels outside the contour, thus forming a spatial mask image that represents the existence of erosion.

[0075] Based on the spatial mask image, morphological centroid calculation processing is applied to traverse all pixels marked as 1, accumulate their horizontal and vertical coordinate values ​​to calculate the first moment, and at the same time, the total number of effective pixels is counted as the zero moment.

[0076] The geometric centroid coordinates of the corroded region are calculated using the following formula:

[0077]

[0078] Where, x c With y c x is the x and y coordinates of the centroid. i With y i Let be the coordinates of the i-th valid pixel, and N be the total number of valid pixels within the eroded region.

[0079] The calculated centroid coordinates are mapped back to the original image space coordinate system, assigned a unique node identifier, and associated with the area proportion and average gray value attributes of the region.

[0080] Traverse all independent eroded connected domains, repeat the above centroid localization and attribute encapsulation process, and aggregate to generate a set of eroded region nodes that represent the erosion distribution density and spatial location.

[0081] By using precise morphological centroid positioning, discrete eroded connected domain boundaries are transformed into graph nodes with clear spatial indices and physical properties, realizing an abstract mapping from pixel-level description to topological node level, and providing standardized input for subsequent diffusion adjacency relationship construction.

[0082] S2.4: Perform diffusion adjacency analysis based on the set of nodes in the corrosion region, apply Euclidean distance threshold determination to establish the connection relationship between adjacent corrosion pits, and calculate the diffusion intensity coefficient as the edge weight based on the oxide film evolution gradient to generate a set of corrosion adjacency edges that characterize the corrosion spread trend.

[0083] Receive the set of corrosion region nodes generated by S2.3, and extract the two-dimensional spatial coordinates of the center of each corrosion pit and the corresponding porosity gradient feature value.

[0084] Perform a fully connected distance matrix calculation on the node set, traverse all node pairs and calculate their Euclidean geometric distances, and construct an initial distance map representing spatial proximity relationships.

[0085] A dynamic threshold parameter based on image resolution and the average diameter of typical pits is set to filter out node pairs with a distance less than the threshold as a set of candidate adjacent edges.

[0086] For each pair of nodes in the candidate adjacent edge set, extract the estimated oxide film thickness or the difference in grayscale mean between the two endpoints corresponding to the pit region.

[0087] The diffusion intensity coefficient is calculated using an exponential decay function model, mapping the gray-scale differences between nodes to probability weights for corrosion propagation, as shown in the following formula:

[0088] Where w is the diffusion intensity coefficient, and I and I j σ represents the mean gray value of two connected erosion pit nodes, and σ is the local gray standard deviation constant.

[0089] The calculated diffusion intensity coefficient is assigned to the corresponding candidate adjacent edge to form a directed edge data structure with weight attributes.

[0090] Merge all weighted directed edges to generate a set of erosion adjacent edges that represent the erosion propagation trend and spatial connectivity.

[0091] Through the above diffusion adjacency analysis and weight quantization, discrete corrosion region nodes are transformed into graph structure edge data with physical evolution significance, achieving the technical effect of corrosion damage from morphological recognition to mechanism modeling.

[0092] S2.5: Perform stress feature mapping on the wear gradient grayscale band in the original damage semantic annotation data, use grayscale gradient change rate extreme value search to locate the wear start point, end point and stress concentration point, combine the force trajectory to define the path connection relationship between nodes and assign contact pressure attenuation coefficient, and generate a multi-type damage subgraph structure set that integrates mechanical evolution features.

[0093] Pixel-level gradient field calculations are performed on the wear gradient grayscale region in the original damage semantic annotation data. The Sobel operator is used to convolve in the horizontal and vertical directions to generate a two-dimensional gradient magnitude map that characterizes the degree of drastic change in surface brightness.

[0094] Based on the gradient magnitude map, a local extremum search process is performed, and an adaptive threshold is set to filter out pixels with gradient magnitudes significantly higher than background noise, marking them as potential stress concentration candidate points.

[0095] Morphological refinement is used to extract the skeleton of candidate point clusters, identify linear structures distributed along the force trajectory, and determine the coordinates of the start and end points of the wear path.

[0096] Calculate the Euclidean distance and directional angle between the starting endpoint, the ending endpoint, and each candidate stress concentration point, and construct a node connection matrix that reflects the geometric topological relationship of wear propagation.

[0097] Based on the formation erosion coefficient in the working condition label, a contact pressure attenuation model is established, and the weight coefficient of the edge between adjacent nodes is calculated using the following formula:

[0098] Where w is the edge weight, P0 is the initial contact pressure, k is the formation erosion attenuation coefficient, and d is the path length between nodes along the wear trajectory.

[0099] The calculated weight coefficients are assigned to the corresponding edges to form a set of directed weighted edges that incorporate the characteristics of mechanical evolution.

[0100] By integrating node coordinates, connectivity relationships, and edge weights, a multi-type damage subgraph structure is generated, which includes wear start points, end points, and stress concentration points.

[0101] Through the stress feature mapping and weight assignment process described above, the geometric annotations from the previous step are transformed into graph structure data with physical and mechanical meaning, realizing the semantic upgrade of wear damage from visual appearance to mechanical essence.

[0102] like Figure 2 As shown, step S3 involves performing normalized scale mapping and spatial coordinate alignment operations based on the multi-type damage subgraph structure set to generate a standardized damage semantic topology map carrying multimodal local descriptors. Specifically, this involves performing normalized scale mapping and spatial coordinate alignment operations based on the multi-type damage subgraph structure set, and extracting the microstructure gradient histogram, edge direction entropy, local contrast variation coefficient, and surface roughness estimate of each node to obtain the standardized damage semantic topology map. This includes: S3.1: Perform affine transformation processing based on a reference datum on the spatial coordinates of the nodes in the multi-type damage subgraph structure set, use the least squares method to fit the standard geometric contour of each structural component as an alignment datum, and map the original coordinates of crack bifurcation points, pit centers and stress concentration points to a unified standard metric space to generate a normalized damage node location dataset with spatial coordinate consistency.

[0103] S3.2: Based on the normalized damage node location dataset, perform a local neighborhood grayscale field reconstruction operation. Use a sliding window gradient operator to traverse the pixel region around each node and calculate the grayscale change rate vector. Use histogram statistics to quantify the distribution frequency of the microstructure to generate a microstructure gradient histogram that characterizes the micro-texture undulation characteristics of the node surface.

[0104] Based on the normalized damage node location dataset, a local sliding window with a preset radius is constructed with each node as the center, and the original gray-level pixel matrix within the window coverage area is extracted as the basic data source for microstructure analysis.

[0105] The Sobel operator is used to perform convolution operation on the local gray-level pixel matrix, and the gradient components in the horizontal and vertical directions are calculated respectively. The resulting gradient vector field represents the magnitude and direction of the gray-level change rate of the pixel.

[0106] The magnitude of the gradient vector is quantized into a microstructure strength index. Based on the gradient direction angle, the local region is divided into several discrete direction intervals, and the distribution frequency of the gradient vector in each direction interval is statistically analyzed.

[0107] The frequency data obtained from the statistics are normalized to eliminate the influence of the absolute value difference of light intensity on texture feature extraction, and a histogram sequence reflecting the distribution law of microscopic surface undulation morphology is formed.

[0108] Through the above processing method, the normalized coordinate data from the previous step is transformed into a microstructure gradient histogram that characterizes the micro-texture undulations of the node surface, thus realizing a quantitative description of the subtle geometric morphology of the damaged area.

[0109] For example, the sliding window size is set to 16×16 pixels with a stride of 1 pixel. A Sobel convolution is performed on the neighborhood of the crack bifurcation point to obtain the gradient magnitude matrix. The 0 to 360 degree orientation angles are divided into 8 bins, and the number of gradient vectors within each bin is counted. After normalization, an 8-dimensional microstructure gradient histogram vector [0.12, 0.45, 0.08, 0.05, 0.10, 0.03, 0.09, 0.08] is generated. This vector significantly highlights the main directional features of the crack orientation and effectively distinguishes random noise from structural damage texture.

[0110] S3.3: Perform edge direction disorder measurement processing based on the microstructure gradient histogram, apply the Shannon entropy calculation formula to solve the information entropy value of the edge direction angle distribution in the node neighborhood, quantify the disorder and evolution complexity of the damage boundary direction, and generate edge direction entropy that reflects the randomness of the damage morphology and texture evolution.

[0111] The microstructure gradient histogram and the corresponding node neighborhood pixel matrix generated by S3.2 are received as the input data source for edge direction disorder measurement.

[0112] The Sobel operator is applied to each pixel in the neighborhood of the node to calculate the gradient, and the horizontal and vertical gradient components are obtained to construct a local gradient field.

[0113] The edge orientation angle of each pixel is calculated based on the gradient components, and the angle between the gradient vector and the horizontal axis is solved by the arctangent function to form the edge orientation angle distribution matrix.

[0114] The Shannon entropy formula is used to solve for the information entropy value of the edge direction angular distribution matrix, quantifying the degree of disorder and evolutionary complexity of the damage boundary orientation.

[0115]

[0116] Where H is the edge direction entropy, N is the total number of discretized intervals of the direction angle, and p i It represents the probability value of the number of pixels in the i-th direction interval relative to the total number of pixels in the neighborhood.

[0117] The calculated original entropy value is normalized and mapped to the [0,1] interval to eliminate the dimensional differences caused by different neighborhood sizes and generate a standardized edge direction entropy.

[0118] Through the above processing method, the microstructure gradient features of the previous step are transformed into edge direction entropy that reflects the randomness of the texture evolution of damage morphology. This enables the quantitative characterization of crack bifurcation complexity, corrosion boundary irregularity and wear trajectory disorder, providing a key texture disorder index for subsequent multimodal feature fusion.

[0119] For example, a 32×32 pixel neighborhood window is selected for a crack bifurcation point on the surface of the tunnel boring machine cutterhead flange. After calculation using the Sobel operator, the pixel distribution probabilities for the eight main directional intervals are statistically obtained as 0.15, 0.10, 0.25, 0.05, 0.20, 0.05, 0.10, and 0.10, respectively. Substituting these values ​​into the Shannon entropy formula, the original entropy value is 2.75 bits. After normalization to the maximum value (the maximum theoretical entropy value is 3.0), the edge direction entropy is obtained as 0.917. This high entropy value accurately reflects the high degree of disorder in the edge direction at the crack bifurcation point, which is different from the lower directional entropy value (usually 0.4-0.6) in the wear region, significantly improving the model's ability to distinguish damage types.

[0120] S3.4: Perform local contrast fluctuation analysis using the edge direction entropy, select the pixel grayscale sequence within a preset radius around the node and calculate the ratio of its mean to standard deviation, evaluate the degree of contrast dispersion caused by uneven illumination and material differences in the local area through coefficient of variation statistical processing, so as to generate a local contrast variation coefficient that characterizes the visual saliency stability of the damaged area.

[0121] Receive the edge direction entropy and normalized damaged node location dataset generated by S3.3, and determine the local analysis neighborhood centered on the current node. Set the preset radius R to 5 pixels, construct a circular sliding window covering the area around the node, and extract the gray value sequence of all pixels within the window to form a local gray value distribution set.

[0122] Calculate the arithmetic mean of the local gray-level distribution set to represent the baseline brightness level within that neighborhood. The mean is calculated using the following formula:

[0123] Where N is the total number of valid pixels within the window, x i Let be the grayscale value of the i-th pixel.

[0124] Based on the mean, the standard deviation of the local grayscale distribution is calculated to quantify the dispersion of pixel grayscale values ​​relative to the mean. The standard deviation is calculated using the following formula:

[0125] Where σ is the standard deviation, which reflects the fluctuation range of local texture.

[0126] Statistical processing of the coefficient of variation was performed, dividing the standard deviation by the arithmetic mean to eliminate the influence of the absolute value of illumination intensity on contrast evaluation. The local contrast coefficient of variation (CV) was calculated using the following formula:

[0127] This indicator characterizes the relative contrast fluctuation under unit brightness, effectively suppressing misjudgments caused by non-uniform lighting in the tunnel boring machine shaft.

[0128] By using the above processing method, the edge direction entropy and local gray-scale statistical features of the previous step are transformed into local contrast variation coefficients, which realizes the quantification of the visual saliency stability of the damaged area and enhances the generalization robustness of the model under varying illumination conditions.

[0129] For example, for the center node of a corrosion pit on the surface of a cutterhead flange, a radius R = 5 pixels and an effective number of pixels N = 78 are set within the window. The calculated local grayscale mean μ = 120.5 and standard deviation σ = 18.2. Substituting these values ​​into the formula, the local contrast coefficient of variation CV = 0.151 is obtained. Compared to the CV value of 0.045 for the surrounding intact metal surface, this index is significantly increased, accurately reflecting the drastic grayscale fluctuations caused by increased surface roughness in the corrosion area, and effectively distinguishing between low-contrast oil stain interference and high-contrast actual damage.

[0130] S3.5: Perform surface physical property inversion processing based on the local contrast variation coefficient, establish the mapping relationship between grayscale response and roughness by combining the camera imaging model and the principle of bidirectional reflection distribution function, derive the surface roughness estimate of the node region, and perform multi-dimensional fusion and encapsulation of microstructure gradient histogram, edge direction entropy, local contrast variation coefficient and surface roughness estimate to generate a standardized damage semantic topology map carrying multimodal local descriptors.

[0131] The local contrast variation coefficient and the original grayscale data of the node neighborhood are received from the output of S3.4 and used as the input basis for the inversion of surface physical properties.

[0132] Based on the camera imaging model, a mapping relationship between pixel gray values, surface normal vectors, and illumination angles is established, and a bidirectional reflectance distribution function (BRDF) is introduced to describe the scattering characteristics of the shield machine's metal surface under specific illumination.

[0133] A roughness inversion equation based on the Oren-Nayar diffuse reflection model is constructed. The local contrast variation coefficient is used as the standard deviation parameter of the surface microfacet slope distribution. The surface roughness estimate is calculated using the following formula:

[0134] Where R is the estimated surface roughness, σ is the standard deviation of the microfacet slope, C is the coefficient of variation of local contrast, and μ n denoted as the mean gray level of the neighborhood, k is the camera gain coefficient, I is the incident light intensity, and θ is the incident angle.

[0135] The above nonlinear equations are solved using the iterative least squares method. By combining the edge direction entropy with the correction term for illumination anisotropy, the microscopic concavity and convexity height distribution matrix of the node region is derived.

[0136] The calculated surface roughness estimate is normalized and mapped to the [0,1] interval to form a roughness scalar that characterizes the physical tactile features.

[0137] The microstructure gradient histogram, edge orientation entropy, local contrast variation coefficient, and surface roughness estimate are multidimensionally spliced ​​and standardized.

[0138] Generate a standardized damage semantic topology map carrying multimodal local descriptors to achieve cross-modal enhancement from two-dimensional visual features to three-dimensional physical properties.

[0139] like Figure 3 As shown, step S4 involves using the standardized damage semantic topology graph as input to construct a graph neural network model containing three hierarchical graph convolutional modules. The first layer performs a first-level hierarchical graph convolutional aggregation operation; the second layer introduces an adaptive edge weight learning mechanism and performs a second-level hierarchical graph convolutional propagation process based on the edge weights generated by this mechanism to output the semantic similarity between different types of damage; the third layer performs a third-level graph pooling operation to generate the global graph embedding vector. Specifically, this includes: S4.1: Perform initial linear mapping processing on the multimodal local descriptors of nodes in the standardized damage semantic topology graph to convert the microstructure gradient histogram, edge orientation entropy, local contrast variation coefficient and surface roughness estimate into an initial node hidden state vector of uniform dimension, providing a standardized feature input basis for subsequent hierarchical graph convolution calculation.

[0140] S4.2: Based on the initial node hidden state vector, perform the first-level graph convolution aggregation operation, use the sparse tensor computing architecture to traverse the adjacency relationship in the damaged subgraph structure set, and perform weighted summation and nonlinear activation transformation on the initial node hidden state vectors of adjacent nodes to generate a primary node aggregation feature vector containing geometric continuity and texture evolution information.

[0141] The initial node hidden state vector and the adjacency matrix of the normalized damage semantic topology graph output by S4.1 are used as input data for the first-level graph convolution aggregation operation.

[0142] A graph message passing mechanism based on sparse tensor computation is constructed, which traverses the k-order neighborhood of each node in the damage subgraph and extracts the hidden state vectors of neighboring nodes and their corresponding geometric connection weights.

[0143] Perform a neighbor feature weighted aggregation operation, and sum the feature vectors of each node in the neighborhood with the normalized edge weights by performing a dot product, according to the linear transformation rules of graph convolutional networks.

[0144] The aggregation feature of the i-th node in layer l is calculated using the following formula:

[0145] in, Let d be the feature vector of node i in the l-th layer, N(i) be the set of neighbors of node i, and d i and d j W represents the degree of nodes i and j, respectively. ij For edge weights, Let σ be the bias term for node i in the l-th layer, and σ be the nonlinear activation function. Let be the input feature vector of node j in the (l-1)th layer.

[0146] The LeakyReLU activation function is introduced to perform nonlinear mapping on the aggregated linear combination results, thereby enhancing the model's ability to capture features such as crack curvature abrupt changes, corrosion edge diffusion, and wear gradient changes.

[0147] Through the hierarchical graph convolution aggregation operation described above, the initial node hidden state of the previous step is transformed into a primary node aggregation feature vector containing information on local geometric continuity and texture evolution, thereby achieving effective encoding of the microscopic morphology of the damaged area.

[0148] S4.3: Based on the aggregated feature vectors of the primary nodes, an adaptive edge weight learning mechanism is constructed. The semantic correlation between the crack segment curvature weight, pit diffusion adjacency weight, and wear path orientation weight is dynamically calculated through a learnable parameter matrix to update the edge connection strength defined in the original damage semantic annotation data and generate an enhanced edge weight matrix with cross-damage type semantic similarity representation.

[0149] The primary node aggregated feature vector and the topological connection structure of the original damaged subgraph output by S4.2 are used as the input data source for adaptive edge weight learning.

[0150] A learnable semantic association mapping network is constructed, and three sets of independent parameter matrices are initialized, including the crack curvature sensitivity matrix, the corrosion diffusion coupling matrix, and the wear stress attenuation matrix, which correspond to the physical evolution characteristics of the three types of damage, respectively.

[0151] The aggregated feature vectors of primary nodes are input into a fully connected layer for dimensional transformation to generate a high-dimensional semantic latent space representation, thereby capturing the potential texture and geometric relationship patterns between nodes.

[0152] A bilinear attention mechanism is used to calculate the interaction strength between any two connected nodes in the latent space. The semantic similarity of node features is quantified by dot product operation to generate an initial semantic association score matrix.

[0153] A non-linear activation function is introduced to normalize the semantic association score, eliminating the influence of dimensional differences on weight allocation and ensuring that the weight values ​​are distributed within the effective range of 0 to 1.

[0154] For the crack subgraph, the initial semantic association score and the geometric curvature weights between nodes are weighted and fused. The contribution ratio of curvature to edge weights is dynamically adjusted using learnable coefficients to strengthen the continuity representation of the crack extension direction and generate the curvature weights of the crack segments.

[0155] For the corrosion subgraph, the initial semantic association score is multiplied by the inverse of the Euclidean distance between corrosion pits and the oxidation gradient coefficient. The physical adjacency effect of corrosion diffusion is simulated by an exponential decay function to generate a corrosion pit diffusion adjacency weight that reflects the corrosion spread trend.

[0156] For the wear subgraph, the initial semantic association score is combined with the directional consistency coefficient of the force trajectory and the contact pressure estimate. Cosine similarity is used to measure the alignment of the path direction and generate wear path direction weights that characterize the evolution of wear mechanics.

[0157] The aforementioned crack segment curvature weights, pit diffusion adjacency weights, and wear path orientation weights are integrated into a unified global adjacency matrix, replacing the original fixed topological connection strengths, to form an enhanced edge weight matrix with semantic similarity representation across damage types.

[0158] By using an adaptive edge weight learning mechanism, the static topology structure of the previous step is transformed into a dynamic semantic association network, which enables refined modeling of the physical evolution of different damages and significantly improves the model's generalization ability and recognition accuracy for collaborative classification of multiple types of damage under complex working conditions.

[0159] For example, the dimension of the initial node aggregation feature vector is set to 64, and the size of the learnable parameter matrix is ​​64×64. For crack node pairs, if the curvature weight is 0.8, the semantic association score is 0.9, and the weighted coefficient learning value is 0.7, then the weighted edge weights after fusion are calculated as 0.7×0.9+0.3×0.8=0.87. For erosion node pairs, if the Euclidean distance is 5 pixels, the oxidation gradient is 0.6, and the semantic score is 0.85, the edge weights are significantly improved after exponential decay processing. The resulting enhanced edge weight matrix greatly improves the message passing efficiency between damaged nodes of the same type and effectively suppresses misclassification caused by noise interference.

[0160] S4.4: The enhanced edge weight matrix is ​​used to perform a second-level graph convolutional propagation process on the aggregated feature vector of the primary node. The enhanced edge weight matrix is ​​fused into the node feature update formula as a message passing adjustment factor to output a high-level node semantic embedding vector that can distinguish the essential differences between the splitting point, the pit center and the stress concentration point.

[0161] The enhanced edge weight matrix generated by S4.3 and the primary node aggregate feature vector output by S4.2 are received as input data for the second-level graph convolutional propagation processing.

[0162] A message passing update equation based on an attention mechanism is constructed, and the semantic correlation coefficient in the enhanced edge weight matrix is ​​mapped to a dynamic adjustment factor in the message passing process, which is used to control the contribution intensity of neighboring node information to the feature update of the central node.

[0163] For any node i in the graph and its set of neighboring nodes N(i), perform a weighted feature aggregation operation and calculate the hidden state of the nodes after convolution of the l-th layer graph using the following formula:

[0164] in, Let σ represent the hidden state vector of node i after the l-th layer processing, where σ is a non-linear activation function such as ReLU, and α is the hidden state vector of node i. ij W represents the adaptive attention coefficients between nodes i and j in the enhanced edge weight matrix. (l) Let be the learnable linear transformation weight matrix of the l-th layer. Let be the feature vector of node j in the previous layer.

[0165] Through the above aggregation operation, the high curvature features of crack bifurcation points, the diffusion texture features of pit centers, and the gradient abrupt change features of stress concentration points are differentially fused in the neighborhood, suppressing noise interference and enhancing category-specific signals.

[0166] Batch normalization is performed on the aggregated feature vectors to eliminate the numerical distribution shift caused by differences in physical dimensions between different damage types, thus ensuring the convergence stability of model training.

[0167] By introducing sparsity constraints through a nonlinear activation function, redundant neighbor information that is irrelevant to the semantics of the current node is filtered out, and a high-level node semantic embedding vector that can accurately distinguish the essential differences between cracks, corrosion, and wear is output.

[0168] S4.5: Perform a third-level graph pooling operation based on the high-level node semantic embedding vector, and use topology-aware global average pooling to compress the spatial dimension of the multi-type damage subgraph structure set to generate a fixed-length global graph embedding vector that represents global semantics. This global graph embedding vector will be used as the direct input for cross-graph contrastive learning for subsequent model training.

[0169] It receives high-level node semantic embedding vectors after the second-level graph convolutional propagation processing. This vector set contains deep semantic features of crack bifurcation points, pit centers, and stress concentration points.

[0170] A topology-aware global average pooling layer is constructed, and an attention masking mechanism based on node degree is introduced to address the dynamic change in the number of nodes in a set of multi-type damaged subgraph structures.

[0171] Calculate the normalized importance weight of each node in the graph, and generate a weight coefficient sequence that reflects the saliency of the local structure based on the number of edges connected to the node and its contribution to the enhanced edge weight matrix.

[0172] By applying weighted average aggregation, the semantic embedding vector of high-level nodes is multiplied element-wise with the corresponding normalized importance weights to achieve differentiated preservation of feature information of each node.

[0173] The weighted node feature vectors are summed to eliminate the impact of differences in the number of nodes on the global feature representation, thus initially compressing the spatial dimension.

[0174] A fixed-dimensional fully connected mapping layer is introduced to project the aggregated variable-length feature vectors onto a preset fixed-length latent space, ensuring the consistency of the output vector dimensions.

[0175] Batch normalization is used to standardize the distribution of the projected vectors, which accelerates model convergence and improves the stability of feature representation.

[0176] Through the aforementioned topology-aware global average pooling operation, the high-level node semantic embedding vector from the previous step is transformed into a fixed-length global graph embedding vector that represents global semantics. This achieves feature sublimation from local micro-features to global macro-damage morphology, providing a unified input benchmark for subsequent cross-graph comparative learning.

[0177] For example, for a composite damage map on the surface of a cutterhead flange, the input is a high-level node semantic embedding vector containing 45 nodes, each with a feature dimension of 128. The degree of each node is calculated, with a maximum degree of 6 and a minimum degree of 1. After normalization, a sequence of weight coefficients is obtained. The 128-dimensional node vector is multiplied by the weights and then summed to obtain a 128-dimensional aggregate vector. Through a 128x64 fully connected layer, a 64-dimensional global graph embedding vector is output. This vector effectively integrates information on crack propagation trend and corrosion diffusion range. In subsequent comparative learning, the Euclidean distance between the embedding vectors of the same type of damage map is significantly reduced, and the discriminative power of the embedding vectors of different types of damage maps is significantly improved, verifying the effectiveness of pooling operation in feature decoupling.

[0178] Step S5: Based on the global graph embedding vector, perform cross-graph comparative learning to construct positive sample pairs between subgraphs of the same type of damage and negative sample pairs between subgraphs of different types of damage. By jointly optimizing graph structure similarity constraints and node feature consistency constraints, a multi-type damage collaborative classification model with cross-damage type feature decoupling capability is generated. Specifically, this includes: S5.1: Perform sample pair construction processing based on the global graph embedding vector, combine the global graph embedding vectors between damage subgraphs of the same type into positive sample pairs, and combine the global graph embedding vectors between damage subgraphs of different types into negative sample pairs, so as to generate a cross-graph contrastive learning training dataset containing a set of positive sample pairs and a set of negative sample pairs.

[0179] S5.2: Calculate graph structure similarity using the cross-graph contrastive learning training dataset. Use graph kernel functions to perform topological structure matching metric on the global graph embedding vectors in the positive sample pair set and topological structure difference metric on the global graph embedding vectors in the negative sample pair set to output graph structure similarity constraints that characterize the correlation strength between damage geometric continuity and texture evolution.

[0180] It receives positive and negative sample pairs from the cross-graph contrastive learning training dataset as global graph embedding vectors, which serve as the input basis for graph structure similarity measurement.

[0181] For the set of positive sample pairs, the topological connection matrix and edge weight distribution features of the nodes in each pair of graphs are extracted, and a graph similarity calculation model based on the random walk kernel is constructed.

[0182] The Weisfeiler-Lehman subtree graph kernel is used to process the iterative aggregation of node neighborhood labels, generating a subtree hash sequence that reflects the local topology, thus quantifying the structural consistency of similar damages in terms of geometric continuity.

[0183] The graph structure similarity score of positive sample pairs is calculated using the following formula:

[0184] Where S is the similarity score, N is the number of iterations, γ is the decay factor, K is the subtree matching count in the i-th iteration, and G1 and G2 are two damage maps in the positive sample pair.

[0185] For the set of negative sample pairs, the same topological feature extraction and kernel function calculation are performed to obtain structural difference measures between different types of damage.

[0186] The similarity scores of positive sample pairs and the difference measures of negative sample pairs are normalized and mapped to the [0,1] interval to generate graph structure similarity constraints.

[0187] Through the above processing method, the global graph embedding vector of the previous step is transformed into a graph structure similarity constraint that represents the correlation strength between the geometric continuity of damage and the evolution of texture, thereby realizing the feature decoupling and differentiation of multiple types of damage at the topological level.

[0188] S5.3: Perform node feature consistency verification based on the graph structure similarity constraint. Based on the maximum mean difference, evaluate the distribution alignment of the microstructure gradient histogram, edge direction entropy and local contrast variation coefficient of each node in the positive sample pair set to generate node feature consistency constraints that characterize the dimension of distinguishing the essence of similar damage.

[0189] Receive the graph structure similarity constraints output by S5.2 and the set of positive sample pairs constructed by S5.1, and extract the node feature matrix of each damage subgraph in the positive sample pairs.

[0190] For the crack, corrosion or wear sub-graphs in the positive sample pairs, the microstructure gradient histogram, edge direction entropy and local contrast variation coefficient of the nodes are extracted to form a high-dimensional node feature distribution sequence.

[0191] The distance between node feature distributions of positive sample pairs is calculated based on the maximum mean difference (MMD) to quantify the feature alignment of the same type of damage under different working conditions.

[0192] The Gaussian kernel function is used as the metric kernel for MMD, and the difference value of the node feature distribution is calculated using the following formula:

[0193] Among them, X p and X q Let n and m be the node feature sets of the two damage subgraphs in the positive sample pair, respectively, where n and m are the number of nodes, k(·,·) is the Gaussian kernel function, and x is the node feature set of the two damage subgraphs in the positive sample pair. i and y j This refers to the feature vector of a specific node.

[0194] The calculated MMD value is used as a node feature consistency constraint. This index reflects the distribution consistency of similar damage in microtexture and geometry.

[0195] By minimizing the MMD constraint, the model is forced to bring the node distribution of similar damage closer together in the feature space, suppressing feature drift caused by changes in illumination or surface rust, and achieving precise locking of the essential distinguishing dimension of similar damage.

[0196] S5.4: Based on the graph structure similarity constraint and the node feature consistency constraint, perform joint loss function optimization, introduce an adaptive weight allocation mechanism to backpropagate and update the parameters of the graph neural network model, so as to minimize the feature distance of positive sample pairs and maximize the feature distance of negative sample pairs, thereby generating an intermediate multi-type damage collaborative classification model with cross-damage type feature decoupling capability.

[0197] The system receives graph structure similarity constraints and node feature consistency constraints, and constructs a joint loss function to drive model parameter updates.

[0198] Define a contrastive learning loss function, which includes a distance minimization term for positive sample pairs and a distance maximization term for negative sample pairs.

[0199] The InfoNCE loss framework is used to calculate the lower bound of mutual information of positive sample pairs in the embedding space.

[0200]

[0201] Where sim is the cosine similarity, τ is the temperature coefficient, and z i z j This is the embedding vector for positive sample pairs.

[0202] A node feature consistency regularization term is introduced to constrain the distribution alignment of features of nodes with the same type of damage.

[0203]

[0204] Where, μ pos μ neg These are the mean vectors of the features of the positive and negative sample nodes, respectively.

[0205] An adaptive weight allocation mechanism is constructed to dynamically balance the contribution ratio of structural constraints and feature constraints.

[0206] The weight coefficients are adjusted based on the number of training iterations, with an initial focus on structural topology matching and a later focus on fine-grained feature decoupling.

[0207]

[0208] in, The dynamic weight coefficients for comparison loss are a function that monotonically decreases with time (number of training iterations t), and their values ​​are usually in the range of (0,1). The dynamic weight coefficients for consistency loss are a function that monotonically increases with time (training iteration number t). and This ebbs and flows, usually satisfying... + =1.

[0209] The gradient of the joint loss function with respect to the parameters of a lightweight graph neural network is calculated using the backpropagation algorithm.

[0210] The Adam optimizer is used to update the network weights, with a focus on correcting the parameters of the adaptive edge weight matrix in the second layer.

[0211] By using a joint optimization process, the constraint indicators of the previous step are transformed into model parameter update quantities, thereby enhancing the ability to solve feature problems across damage types.

[0212] S5.5: Perform generalization performance verification and finalization on the intermediate multi-type damage collaborative classification model, use the standardized damage semantic topology map under the reserved working condition label to perform forward inference calculation, obtain the joint confidence distribution vector, and solidify the model parameters according to the accuracy convergence of the joint confidence distribution vector, and finally generate a multi-type damage collaborative classification model that can be used for collaborative identification of multiple damage types of shield machine structural components.

[0213] Obtain the intermediate-state multi-type damage co-classification model and the standardized damage semantic topology map validation set under reserved working condition labels. Input the map from the validation set into the model, perform forward inference calculation, and generate the joint confidence distribution vector of cracks, corrosion, and wear.

[0214] Based on the joint confidence distribution vector and the truth value verified manually, a confusion matrix is ​​constructed, and the recognition accuracy, recall, and F1 score of various types of impairments are calculated. The fluctuation of the loss function value during the validation process is monitored, and the slope of the accuracy convergence curve is recorded over consecutive epochs.

[0215] Set a convergence threshold; when the accuracy increment is lower than a preset small value and the loss function tends to stabilize, the model parameters are considered to have reached their optimal state. Freeze the weight parameters of all layers in the lightweight graph neural network, including the first-layer feature mapping matrix, adaptive edge weight learning parameters, and pooling layer coefficients.

[0216] Perform a model serialization and saving operation to package the solidified network structure definition file and binary weight file to generate the final multi-type damage co-classification model file.

[0217] Through the above generalization performance verification and parameter solidification processing, the intermediate model is transformed into a fixed model with stable reasoning ability, realizing high-precision collaborative identification of multiple damage types of shield tunneling components under complex working conditions.

[0218] For example, 500 standardized damage semantic topology maps under water-rich soft soil conditions were selected as the validation set. The convergence threshold was set to an accuracy increment of less than 0.001 for 5 consecutive epochs. Inference showed that the crack identification accuracy was 98.2%, corrosion 96.5%, and wear 97.1%. The loss function stabilized around 0.032 after the 45th epoch. The model parameters were frozen and saved as an .onnx file, with a file size controlled within 4.5MB to meet embedded deployment requirements.

[0219] Step S6: Extract the semantic topological map of the damage to be tested from the image of the tunnel boring machine structural component to be inspected, and input it into the multi-type damage collaborative classification model to output the joint confidence distribution result containing multiple types of damage. Specifically, the multiple types of damage include three types of damage: cracks, corrosion, and wear. Input the image of the tunnel boring machine structural component to be inspected into the map generation engine, and use fast edge-guided topological skeleton extraction processing to reconstruct the semantic topological map of the damage to be tested in the current scene in real time. This includes: S6.1: Acquire high-resolution surface images of the structural components of the tunnel boring machine to be inspected, and perform multi-scale gradient field analysis and non-maximum suppression processing on the images using fast edge-guided topological skeleton extraction to generate an initial damage geometric feature map containing a candidate set of crack centerlines, the boundary contour of the corrosion area, and the gray-scale gradient band of the wear trajectory.

[0220] S6.2: Based on the initial damage geometric feature map, perform a semantic connected region construction operation, abstract the crack centerline candidate set into a directional node chain, aggregate the corrosion region boundary contour into a connected domain with porosity gradient, and map the wear trajectory grayscale gradient band into a stress concentration point sequence distributed along the force path, thereby generating an original damage topology node set with physical semantic attributes.

[0221] Receive the initial damage geometry feature map output by S6.1, which includes a candidate set of crack centerlines, the boundary contour of the corrosion area, and the gray-scale gradient band of the wear trajectory.

[0222] Skeleton refinement and orientation field analysis are performed on the candidate set of crack centerlines. The local curvature of the image is calculated using the Hessian matrix, and the curvature extrema are extracted as topological nodes. Adjacent nodes are connected based on the consistency of gradient direction to form a directional vector chain, thus establishing the physical semantic path of crack propagation.

[0223] Morphological closure operation is performed to fill the boundary contour of the corrosion region. The geometric centroid and area moment of each independent connected domain are calculated. The centroid is mapped to corrosion diffusion nodes. The node attributes are quantified based on the porosity gradient distribution characteristics to construct a discrete point cloud set representing the oxidation erosion density.

[0224] Principal component analysis is performed on the gray-scale gradient band of the wear trajectory. The principal axis direction with the largest gray-scale change rate is extracted as the force trajectory. The extreme points of the second derivative of gray-scale are searched along the trajectory to locate the stress concentration area. The extreme points are serialized into wear evolution nodes distributed along the mechanical load path.

[0225] By integrating the above three types of node data, each node is assigned a corresponding physical type label and spatial coordinate attribute, generating a set of original damage topology nodes with clear mechanical and geometric semantics.

[0226] Through the semantic connectivity region construction process described above, pixel-level geometric features are transformed into topological node data carrying physical evolution information, realizing the mapping from visual representation to structural semantics and providing standardized input for subsequent graph structure construction.

[0227] S6.3: Perform adjacency relationship inference and edge weight assignment on the original damage topology node set, calculate curvature weight, diffusion adjacency coefficient and contact pressure attenuation factor based on the Euclidean distance and texture similarity between nodes, and construct a semantic topology map of the damage to be tested that represents geometric continuity and texture evolution.

[0228] Based on the original set of damaged topological nodes generated by S6.2, spatial adjacency relationship inference processing between nodes is performed. The Euclidean distance between any two damaged nodes of the same type is calculated, and node pairs with a distance less than a preset neighborhood threshold are selected as candidate connection edges to construct a preliminary spatial topological skeleton.

[0229] For crack-type node pairs, the direction vector of the connecting line segment in the image coordinate system is extracted, and the rate of curvature change of the line segment is calculated by combining it with the local pixel gradient field. The curvature weight of the crack edge is calculated using the following formula:

[0230] Among them, W c Here, k represents the normalized average curvature value, and k is the curvature weight. This weight is used to characterize the geometric continuity of crack propagation; the greater the curvature, the lower the weight, reflecting the abrupt change in crack direction.

[0231] For corrosion-prone node pairs, the variance of pixel grayscale distribution within the region connecting the centers of two corrosion pits is analyzed to quantify the uniformity of oxide film diffusion. The diffusion adjacency coefficient is calculated using the following formula:

[0232] Among them, W eD is the diffusion adjacency coefficient. max For the maximum permissible diffusion distance, D ij σ is the Euclidean distance between nodes. gray This represents the standard deviation of grayscale values ​​in the connected region. This coefficient reflects the intensity of the corrosion propagation trend; the closer the distance and the more uniform the grayscale, the stronger the adjacency relationship.

[0233] For wear-type node pairs, the gray-level gradient decay sequence between nodes is extracted along the force trajectory direction, and the contact pressure transmission efficiency is calculated. The contact pressure decay factor is determined based on the integral value of the gray-level gradient change rate, the physical-mechanical mapping relationship of the wear path is established, and edge weights characterizing texture evolution are generated.

[0234] By integrating the above three types of edge weight data, the node set and the weighted edge set are encapsulated into a graph data structure to form a semantic topology graph of the damage to be tested.

[0235] By fusing multimodal features and assigning edge weights constrained by physical mechanisms, the discrete nodes from the previous step are transformed into graph structure data with geometric continuity and texture evolution semantics. This achieves semantic dimensionality enhancement of damage morphology from the pixel level to the topological level, providing high discriminative input for subsequent graph neural network inference.

[0236] S6.4: Input the semantic topology graph of the damage to be tested into a pre-trained multi-type damage collaborative classification model, and use the hierarchical graph convolution module inside the model to perform hierarchical graph convolution aggregation operation and adaptive edge weight propagation operation to generate a global graph embedding vector representing the global damage semantics.

[0237] The system receives the semantic topology map of the damage to be tested, extracts the multimodal local descriptor matrix of the nodes and the sparse tensor of the adjacency relationship, and uses it as the input data for the hierarchical graph convolutional network.

[0238] A linear projection transformation is performed on the input node features, and a learnable weight matrix is ​​used to map heterogeneous features such as microstructure gradient histogram and edge direction entropy to a hidden state space of uniform dimension, eliminating the influence of dimensional differences on subsequent aggregation operations.

[0239] In the first-layer graph convolution module, a message passing mechanism is executed based on the adjacency matrix, the hidden state vectors of the first-order neighbors of each node are aggregated, and a primary node embedding representation containing local geometric continuity is generated through weighted summation and nonlinear activation function.

[0240] An adaptive edge weight learning mechanism is introduced to calculate the semantic correlation between crack curvature, corrosion diffusion coefficient and wear pressure attenuation factor, and dynamically update the edge weight matrix to enhance the texture evolution similarity representation across damage types.

[0241] In the second-layer graph convolution module, the message passing intensity is adjusted by using the updated enhanced edge weight matrix, and higher-order neighbor information is further aggregated to output a high-level node semantic embedding vector that can distinguish the essential differences between splitting forks, pit centers and stress concentration points.

[0242] The third-layer graph pooling operation is performed, and topology-aware global average pooling is used to compress the spatial dimension of the semantic embedding vectors of high-level nodes, generating a fixed-length global graph embedding vector that represents the global damage semantic distribution.

[0243] By using hierarchical graph convolution and adaptive weight propagation, the topological map of the damage to be tested is transformed into a global graph embedding vector, which enables the decoupling and generalization of multi-type damage features under complex working conditions, providing a highly robust semantic foundation for subsequent classification decisions.

[0244] For example, the semantic topological map of the damage to be tested generated from the surface image of the cutterhead flange contains 128 nodes and 350 edges. The input node feature dimension is 64-dimensional, which is aggregated by the first layer of graph convolution to generate a 64-dimensional primary embedding. The adaptive edge weight learning mechanism calculates the semantic similarity coefficient between the crack segment and the wear path to be 0.85, which is significantly higher than the 0.32 between the crack and corrosion. After the second layer of propagation and the third layer of global average pooling, a global graph embedding vector of length 128-dimensional is output. This vector significantly improves the Euclidean distance clustering density on the test set, effectively distinguishing similar texture interference under illumination changes, and verifying the model's feature extraction capability for collaborative identification of multiple damage types under limited computing power.

[0245] S6.5: Based on the global graph embedding vector, perform fully connected layer mapping and soft maximization normalization to calculate the probability distribution values ​​of the three types of damage, namely crack, corrosion and wear, under the current working conditions, and output the final inference data containing the joint confidence distribution results of the three types of damage.

[0246] Step S7: Based on the joint confidence distribution results, perform a graph structure integrity assessment. According to the assessment results, generate a low-confidence warning signal and mark areas recommended for manual review, or output the final damage classification conclusion. Specifically, calculate the connectivity of various damage subgraphs in the joint confidence distribution results; if the connectivity of a certain type of damage subgraph is determined to be lower than a preset threshold, generate the low-confidence warning signal and mark areas recommended for manual review; otherwise, output the final damage classification conclusion. This includes: S7.1: The confidence vectors of the three types of damage (crack, corrosion, and wear) in the joint confidence distribution results are analyzed and processed to extract the corresponding high-confidence node set and edge connection relationship data, so as to reconstruct the crack centerline topology chain, corrosion connected domain boundary graph and wear gradient grayscale map to be evaluated as the input objects for graph structure integrity evaluation.

[0247] S7.2: Based on the reconstructed crack centerline topological chain, corrosion connectivity domain boundary graph, and wear gradient grayscale graph, graph theory connectivity component analysis is used to calculate the maximum connected subgraph node ratio and average path length of each damage subgraph to generate an original connectivity sequence characterizing the geometric continuity of various damage types. The damage subgraphs include the crack subgraph, corrosion subgraph, and wear subgraph.

[0248] The input objects are the crack centerline topology chain, the corrosion connected domain boundary map, and the wear gradient grayscale map reconstructed by S7.1, which correspond to the crack sub-map, corrosion sub-map, and wear sub-map, respectively.

[0249] Perform a depth-first search traversal on the cracked subgraph to identify all interconnected sets of nodes in the graph, and calculate the connected component N containing the most nodes. max,c and the total number of its nodes N total,c .

[0250] The percentage of nodes in the largest connected subgraph of the crack is calculated using the following formula:

[0251] in, The number of nodes in the largest connected component. P represents the total number of nodes in the crack subgraph, an index that characterizes the continuity of crack fracture. c This represents the percentage of nodes in the largest connected subgraph of the crack.

[0252] The shortest path length between all connected node pairs in the crack subgraph is calculated based on Dijkstra's algorithm, and the average value is taken as the average crack path length, reflecting the spatial span of crack propagation.

[0253] Perform the same connectivity component analysis on the corrosion subgraph, calculate the percentage of nodes in the largest connected subgraph in the corrosion region, and quantify the cluster density and diffusion range of the corrosion pit group.

[0254] The average Euclidean distance between adjacent pit nodes in the corrosion subgraph is calculated as the average path length, which characterizes the tightness of corrosion propagation.

[0255] Extract the stress concentration point sequence along the force trajectory from the wear subgraph, calculate the node connectivity ratio on the main wear path, and assess the integrity of the wear zone.

[0256] The average geometric distance between adjacent stress concentration points on the wear path is calculated as the average path length, reflecting the linear characteristics of wear evolution.

[0257] The original connectivity sequence is generated by arranging the percentage of nodes in the maximum connected subgraph of the crack, the average path length of the crack, the percentage of nodes in the maximum connected subgraph of the corrosion region, the average Euclidean distance between adjacent pit nodes in the corrosion subgraph, the percentage of nodes connected on the main wear path, and the average geometric distance between adjacent stress concentration points on the wear path in order of damage type.

[0258] By using graph theory connected component analysis and shortest path algorithm, the reconstructed damage topology graph is transformed into a quantified sequence of geometric continuity indicators, enabling a preliminary quantitative assessment of the morphological integrity of various damage types.

[0259] S7.3: Perform normalized weighted fusion processing on the original connectivity sequence, and calculate a comprehensive connectivity evaluation value that reflects the integrity of the damage morphology under the current working condition by combining the texture evolution weight coefficient and deformation conduction attenuation factor specific to the damage type.

[0260] The system receives the original connectivity sequence, including the percentage of nodes in the crack's maximum connected subgraph, the average path length of corrosion, and the number of wear connected components. It extracts the current working condition label of the tunnel boring machine (TBM) to identify the stratum type and tunneling parameters, serving as the context for weight calculation. It retrieves the texture evolution weight coefficient library specific to each damage type to obtain the benchmark values ​​for the geometric continuity weight of cracks, the diffusion and propagation weight of corrosion, and the stress transmission weight of wear. It then calls the deformation transmission attenuation factor calculation module to dynamically correct the physical sensitivity parameters for each damage type based on the load level and vibration frequency in the working condition label. Finally, it constructs a comprehensive connectivity evaluation value calculation model, performing a weighted fusion calculation of the original connectivity and dynamic weights, using the following formula:

[0261] Where C is the comprehensive connectivity evaluation value, w i Let be the texture evolution weight coefficient for the i-th type of damage. Let M be the deformation conduction attenuation factor for the i-th type of damage under the current operating conditions. i Let represent the original connectivity of the i-th type of damage. Normalization is performed, mapping the weighted summation result to a standard interval of 0 to 1 to eliminate the influence of dimensional differences on the evaluation results. The output is a comprehensive connectivity evaluation value reflecting the integrity of the damage morphology under the current operating condition, achieving quantitative unification of multi-source damage characteristics.

[0262] S7.4: Compare the comprehensive connectivity evaluation value with the various damage connectivity dynamic thresholds preset in the shield machine structural health monitoring database. If the comprehensive connectivity evaluation value is determined to be lower than the preset threshold of the corresponding type, an abnormal state marker is triggered and a low confidence warning signal containing spatial coordinate information is generated.

[0263] Obtain the comprehensive connectivity evaluation value calculated by S7.3, and retrieve the corresponding dynamic threshold parameter set from the tunnel boring machine structural health monitoring database based on the damage type index. For the three types of damage—cracks, corrosion, and wear—extract their preset lower connectivity threshold and confidence attenuation coefficient, and construct a multidimensional threshold comparison vector.

[0264] An element-wise comparison process is used to logically operate the comprehensive connectivity evaluation value against the corresponding dynamic threshold. If the evaluation value is lower than the threshold, the topology of the damaged area is determined to have a break or sparsity anomaly, triggering a low-confidence labeling state.

[0265] Based on the marked state, the spatial coordinate backtracking module is invoked to extract the set of key nodes that lead to reduced connectivity and their neighborhood boundary coordinates from the original damage semantic topology graph. The convex hull algorithm is then used to calculate the bounding box of the key node coordinates, generating a suggested geometric descriptor for the manually reviewed region containing the minimum bounding rectangle.

[0266] The geometric descriptor of the review area and the low confidence warning signal are encapsulated into a standard JSON format data packet, which includes damage type identifier, confidence score, anomaly cause code and vertex coordinate sequence of the review area.

[0267] Through the above comparison and encapsulation process, the overall connectivity of the previous step is transformed into a low-confidence early warning signal with spatial positioning capabilities, enabling accurate marking of potential false detection areas and guidance for manual intervention, significantly improving the reliability of the detection system under complex working conditions.

[0268] S7.5: Perform a branch decision operation based on the presence or absence of the low confidence warning signal. If a low confidence warning signal exists, output a structural component surface damage detection report indicating the area to be manually reviewed. If no low confidence warning signal exists, directly output the final damage classification judgment conclusion, which includes the specific categories of cracks, corrosion, and wear.

[0269] Step S8: Based on the final damage classification judgment and low confidence warning signal, dynamically update the damage evolution records in the shield machine structural health monitoring database, and fine-tune and optimize the adaptive edge weights of the multi-type damage collaborative classification model according to the newly entered working condition labels and damage true values, so as to achieve iterative improvement of the model's generalization ability under complex working conditions.

[0270] In this embodiment, the damage truth value refers to the standard damage category label, which is the original damage semantic annotation data generated in step S1 after the domain engineer interactively delineates the semantically connected region based on the high-definition surface image, and has undergone abstract mapping and standardization processing. The damage truth value specifically characterizes the actual damage category (such as cracks, pits, wear, etc.) existing on the surface of the key structural components of the tunnel boring machine under specific service conditions.

[0271] S8 specifically includes: S8.1: Obtain the final damage classification judgment, low confidence warning signal, and corresponding newly entered working condition label and damage true value. Perform spatiotemporal alignment and structured encapsulation processing on the above multidimensional data to generate a damage evolution incremental data package containing complete contextual information, providing a standardized input source for subsequent database updates.

[0272] S8.2: Based on the damage evolution incremental data package, perform version verification and conflict detection processing, merge the verified incremental data into the historical sequence of the tunnel boring machine structural health monitoring database, so as to generate an updated damage evolution record with temporal continuity and state integrity, ensuring that the stored data can accurately reflect the damage evolution trajectory of the structural components throughout their entire life cycle.

[0273] S8.3: Using the damage true value in the updated damage evolution record as a supervision signal, and combining it with the newly entered working condition label, construct an error feedback matrix for a specific geological environment, calculate the prediction deviation gradient of the multi-type damage co-classification model under the current working condition, and generate a weight adjustment direction vector that characterizes the model's insufficient adaptability.

[0274] The working condition labels and corresponding ground truth damage values ​​under specific geological conditions are extracted from the updated damage evolution records and used as the baseline data source for error feedback calculation. The joint confidence distribution output by the multi-type damage co-classification model under the current working condition is aligned and matched with the ground truth damage values ​​to construct prediction bias vectors for three types of damage: cracks, corrosion, and wear. The difference between the predicted probability distribution and the true labels is calculated using the cross-entropy loss function to quantify the classification error magnitude of the model under specific geological conditions. The error feedback component for the k-th type of damage is calculated using the following formula:

[0275] Among them, y k This represents the true label value of the k-th type of damage. This represents the confidence probability value predicted by the model. Let be the error feedback component for the k-th type of damage. Weighted summation of the error feedback components for each type of damage is performed to generate a global error scalar reflecting the overall prediction bias. The global error scalar is backpropagated to the adaptive edge weight layer of the lightweight graph neural network using the chain rule, and the gradient of the loss function with respect to the edge weight parameters is calculated. The adjustment direction vector of the edge weights between the i-th node and the j-th node is calculated using the following formula:

[0276] Among them, G ij Let w be the edge weight ijThe gradient direction is given by E, where E is the global error scalar. Momentum accumulation processing is performed on the calculated gradient vector, incorporating historical gradient information to smooth noise interference under current conditions, generating a stable weight adjustment direction vector. Through gradient backpropagation, prediction bias is transformed into a weight adjustment direction vector characterizing insufficient model adaptability, achieving precise model parameter positioning and optimization guidance for specific geological environments.

[0277] S8.4: Based on the weight adjustment direction vector, execute the backpropagation optimization strategy under the adaptive learning rate constraint to locally fine-tune and update the adaptive edge weight parameters in the second layer of the multi-type damage collaborative classification model, so as to generate an optimized adaptive edge weight set with the ability to map new working conditions, thereby enhancing the model's semantic similarity modeling accuracy for unknown damage patterns.

[0278] Receive the weight adjustment direction vector generated by S8.3, which represents the bias gradient of the model in modeling the semantic similarity of the three types of damage, crack, corrosion and wear, under the current specific working condition.

[0279] An adaptive learning rate scheduler is constructed, which dynamically calculates the base learning rate decay coefficient for the current step size based on the slope of the convergence curve of the loss function in the historical iterations, so as to balance the fine-tuning speed and stability.

[0280] The weight adjustment direction vector is projected onto the adaptive edge weight parameter space of the second layer of the multi-type damage collaborative classification model to identify the edge connection subset that is most relevant to the currently newly entered working condition label.

[0281] For the identified subset of highly correlated edges, the momentum optimization method is applied to update their weight parameters, suppress high-frequency noise interference, and accelerate the convergence process along the gradient descent direction.

[0282] The updated adaptive edge weights are calculated using the following formula:

[0283] Among them, W t+1 W is the updated adaptive edge weight set. t η represents the edge weights at the current moment, η is the dynamically calculated base learning rate, v is the momentum coefficient, and g is the weight adjustment direction vector.

[0284] The updated edge weights are subjected to L2 regularization constraints to prevent overfitting caused by small sample fine-tuning and to ensure that the generalization boundary of the model is not over-compressed under unknown damage modes.

[0285] By using a backpropagation optimization strategy under an adaptive learning rate constraint, the weight adjustment direction vector is transformed into an optimized adaptive edge weight set with the ability to map features of new working conditions, thereby significantly enhancing the model's accuracy in semantic similarity modeling of unknown damage modes under complex and variable shield tunneling construction environments.

[0286] S8.5: Replace the old weight parameters in the original multi-type damage collaborative classification model with the optimized adaptive edge weight set, and perform a generalization performance evaluation test on the validation set to output an instance of the multi-type damage collaborative classification model that has completed the iterative upgrade, so as to realize the continuous self-learning and dynamic improvement of detection accuracy of the system in the complex and variable shield tunneling construction environment.

[0287] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0288] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0289] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision, characterized in that, include: S1: Obtain high-resolution surface images and corresponding condition labels of key structural components of the tunnel boring machine under different service conditions, and generate original damage semantic annotation data based on the semantic connectivity regions drawn interactively by domain engineers. S2: The original damage semantic annotation data is abstracted and mapped to construct a set of multi-type damage subgraph structures; S3: Based on the set of multi-type damage subgraph structures, perform normalized scale mapping and spatial coordinate alignment operations to generate a standardized damage semantic topology map carrying multimodal local descriptors; S4: Using the standardized damage semantic topology map as input, a global graph embedding vector is generated based on a preset graph neural network model; S5: Based on the global graph embedding vector, perform cross-graph comparison learning, construct positive sample pairs between subgraphs of the same type of damage and negative sample pairs between subgraphs of different types of damage, and generate a multi-type damage collaborative classification model by jointly optimizing graph structure similarity constraints and node feature consistency constraints. S6: Extract the semantic topology map of the damage to be tested from the image of the structural component of the tunnel boring machine to be inspected, and input the semantic topology map of the damage to be tested into the multi-type damage collaborative classification model, and output the joint confidence distribution result containing multiple types of damage; S7: Based on the joint confidence distribution results, perform a graph structure integrity assessment. Based on the assessment results, generate a low confidence warning signal and mark areas recommended for manual review, or output the final damage classification judgment conclusion.

2. The method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision according to claim 1, characterized in that, After step S7, the following is included: S8: Based on the final damage classification judgment and low confidence warning signal, dynamically update the damage evolution records in the shield machine structural health monitoring database, and fine-tune and optimize the adaptive edge weights of the multi-type damage collaborative classification model according to the newly entered working condition labels and damage true values.

3. The method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision according to claim 1, characterized in that, The abstract mapping process of the original damage semantic annotation data to construct a set of multi-type damage subgraph structures includes: defining crack bifurcation points and endpoints as nodes, crack segments as edges with curvature weights, pit centers as nodes, diffusion adjacency relationships as edges, stress concentration points as nodes, and wear path directions as edges, thus obtaining the set of multi-type damage subgraph structures.

4. The method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision according to claim 3, characterized in that, Step S3 specifically includes: Affine transformation based on a reference datum is performed on the spatial coordinates of nodes in the multi-type damage subgraph structure set. The standard geometric contour of each structural component is fitted using the least squares method as an alignment datum. The original coordinates of the crack bifurcation point, pit center and stress concentration point are mapped to a unified standard metric space to generate a normalized damage node location dataset. Based on the normalized damage node location dataset, a local neighborhood grayscale field reconstruction operation is performed. A sliding window gradient operator is used to traverse the pixel region around each node and calculate the grayscale change rate vector. The distribution frequency of the microstructure is quantified by the histogram statistical method to generate a microstructure gradient histogram. The edge direction disorder is calculated based on the microstructure gradient histogram. The Shannon entropy calculation formula is applied to solve the information entropy value of the edge direction angle distribution in the node neighborhood, quantifying the disorder and evolution complexity of the damage boundary direction, and generating edge direction entropy. The local contrast fluctuation analysis operation is performed using the edge direction entropy. The pixel grayscale sequence within a preset radius around the node is selected and its mean to standard deviation ratio is calculated. The contrast dispersion caused by uneven lighting and material differences in the local area is evaluated by the coefficient of variation statistical processing, and the local contrast variation coefficient is generated. Surface physical property inversion processing is performed based on the local contrast variation coefficient. The mapping relationship between grayscale response and roughness is established by combining the camera imaging model and the principle of bidirectional reflection distribution function. The surface roughness estimate of the node region is derived. The microstructure gradient histogram, edge direction entropy, local contrast variation coefficient and surface roughness estimate are multidimensionally fused and encapsulated to generate the standardized damage semantic topology map carrying multimodal local descriptors.

5. The method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision according to claim 4, characterized in that, The graph neural network model includes three hierarchical graph convolution modules. The first layer is used to perform a first-level graph convolution aggregation operation. The second layer is used to introduce an adaptive edge weight learning mechanism and perform a second-level graph convolution propagation process based on the edge weights generated by the mechanism to output the semantic similarity between different types of damage. The third layer is used to perform a third-level graph pooling operation to generate the global graph embedding vector.

6. The method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision according to claim 5, characterized in that, Step S4 specifically includes: An initial linear mapping process is performed on the node multimodal local descriptors in the standardized damage semantic topology map to convert the microstructure gradient histogram, edge orientation entropy, local contrast variation coefficient and surface roughness estimate into an initial node hidden state vector of uniform dimension. Based on the initial node hidden state vector, the first hierarchical graph convolution aggregation operation is performed. The sparse tensor computing architecture is used to traverse the adjacency relationships in the set of damaged subgraph structures. The initial node hidden state vectors of adjacent nodes are weighted and nonlinearly activated to generate primary node aggregation feature vectors. An adaptive edge weight learning mechanism is constructed based on the aggregated feature vectors of the primary nodes. The semantic correlation between the crack segment curvature weight, pit diffusion adjacency weight, and wear path orientation weight is dynamically calculated through a learnable parameter matrix. The edge connection strength defined in the original damage semantic annotation data is updated and an enhanced edge weight matrix is ​​generated. The enhanced edge weight matrix is ​​used to perform the second-level graph convolutional propagation process on the aggregated feature vector of the primary node. The enhanced edge weight matrix is ​​fused into the node feature update formula as a message passing adjustment factor to output a high-level node semantic embedding vector that can distinguish the essential differences between the splitting point, the pit center and the stress concentration point. The third-level graph pooling operation is performed based on the high-level node semantic embedding vector. Topology-aware global average pooling is used to compress the spatial dimension of the multi-type damage subgraph structure set to generate the global graph embedding vector.

7. The method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision according to claim 1, characterized in that, Step S5 specifically includes: Based on the global graph embedding vector, sample pair construction processing is performed, combining the global graph embedding vectors between damage subgraphs of the same type into positive sample pairs, and combining the global graph embedding vectors between damage subgraphs of different types into negative sample pairs, generating a cross-graph contrastive learning training dataset containing a set of positive sample pairs and a set of negative sample pairs; The graph structure similarity metric is calculated using the cross-graph contrastive learning training dataset. A graph kernel function is used to perform topological structure matching metric on the global graph embedding vectors in the positive sample pair set and topological structure difference metric on the global graph embedding vectors in the negative sample pair set, so as to output graph structure similarity constraints that characterize the correlation strength between damage geometric continuity and texture evolution. Node feature consistency verification is performed based on the graph structure similarity constraint. The distribution alignment of the microstructure gradient histogram, edge direction entropy and local contrast variation coefficient of each node in the positive sample pair set is evaluated based on the maximum mean difference, and node feature consistency constraints are generated. Based on the graph structure similarity constraint and the node feature consistency constraint, a joint loss function optimization is performed. An adaptive weight allocation mechanism is introduced to backpropagate and update the parameters of the graph neural network model in order to minimize the feature distance of positive sample pairs and maximize the feature distance of negative sample pairs, thereby generating an intermediate multi-type damage collaborative classification model. The generalization performance of the intermediate multi-type damage collaborative classification model is verified and finalized. Forward inference calculation is performed using the standardized damage semantic topology map under the reserved working condition label to obtain the joint confidence distribution vector. The model parameters are then solidified based on the accuracy convergence of the joint confidence distribution vector to generate the multi-type damage collaborative classification model.

8. The method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision according to claim 1, characterized in that, The various types of damage include: cracks, corrosion, and wear.

9. The method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision according to claim 1, characterized in that, The original damage semantic annotation data includes a crack centerline topological chain with directional and curvature attributes, a corrosion connected domain boundary characterizing the porosity gradient, and a wear gradient grayscale band distributed along the stress trajectory.

10. The method for detecting surface damage of key structural components of a tunnel boring machine based on computer vision according to claim 1, characterized in that, Based on the joint confidence distribution results, a graph structure integrity assessment is performed. According to the assessment results, a low-confidence warning signal is generated and areas recommended for manual review are marked, or a final damage classification conclusion is output, including: Calculate the connectivity of each type of damage subgraph in the joint confidence distribution result; if the connectivity of a certain type of damage subgraph is determined to be lower than a preset threshold, generate the low confidence warning signal and mark the area recommended for manual review; otherwise, output the final damage classification judgment conclusion.

Images