Method for generating voxel-level structural damage supervision labels based on component-level discrete labels

Abstract

The application discloses a method for generating a voxel-level structural damage supervision label based on component-level discrete labels, which first performs three-dimensional voxelization processing on a target structure, establishes a mapping relationship between components and voxels, and obtains point source load event parameters and discrete damage level labels of each component. Then, the continuous vulnerability quantitative value of each voxel is calculated, which comprehensively reflects the load influence and material resistance capacity. For each component, the number of voxels to be marked as damaged is determined according to the damage level of the component, and the loss marking probability distribution inside the component is constructed based on the vulnerability value of the voxel. Finally, the damaged voxels are marked according to the probability distribution, and the voxel-level damage supervision label is generated after aggregation. The application uses coarse-grained engineering labels that are easy to obtain to automatically generate fine-grained supervision data that conforms to physical laws and has controllable randomness, effectively solving the problem of the lack of fine-grained labeled data when training a three-dimensional damage segmentation network.

Description

TECHNICAL FIELD

[0001] The present application relates to the intersection field of engineering intelligent simulation and computer vision, and particularly relates to a method for generating a voxel-level structure damage supervision label based on a component-level discrete label. BACKGROUND

[0002] In the field of structural engineering, safety assessment and disaster prevention and mitigation, accurately predicting the three-dimensional spatial damage distribution of complex engineering structures under the action of local extreme loads (such as impact, explosion, shock wave, etc.) is the key to safety rating, emergency decision-making and accident inversion. At present, the mainstream technical path is faced with the dual bottlenecks of high cost and high data dependence, and there is a significant data form fault between the two.

[0003] Firstly, the finite element simulation method based on physical mechanism can provide element-level precision damage field, but its calculation cost is high, the working condition coverage is limited, and the sensitive model and data are difficult to be publicly shared, so it is impossible to form a large-scale training data set.

[0004] Secondly, the three-dimensional segmentation method based on deep learning (such as 3D U-Net) has high efficient generalization potential, but its training is extremely dependent on a large number of voxel-level (pixel-level) fine labeling true value. Such data is almost impossible to obtain in reality, forming a fatal obstacle from method to application.

[0005] The reality dilemma is that only coarse-grained component-level damage labels (such as “XX beam-moderate damage”) can be widely obtained in engineering practice. If they are simply broadcast to the entire component geometric space as training labels, serious label noise and physical distortion will be generated:

[0006] Masking the real space mode: the uniform label cannot express the key physical laws that the damage diffuses from the load source and intensifies in the stress concentration area.

[0007] Damage model generalization ability: the model trained under such distorted supervision is difficult to generate reasonable spatial prediction for unseen working conditions.

[0008] Therefore, the existing technical system has a clear gap: under the condition of only coarse-grained component labels, how to synthesize high-quality, physically credible voxel-level supervision signals to reliably train three-dimensional damage segmentation models. The ideal synthesis method needs to consider two levels:

[0009] Physical constraints: the generated fine-grained labels need to meet the basic mechanical intuition of near-source vulnerability and gradient distribution.

[0010] Random reasonableness: controlled randomness needs to be introduced to simulate the uncertainty of materials and boundaries, so as to improve the generalization and robustness of the model.

[0011] Breaking through this bottleneck of generating supervision signals from coarse to fine is the key prerequisite for connecting limited engineering data with advanced artificial intelligence models and realizing the practicality of structural intelligent damage prediction. SUMMARY

[0012] To solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0013] The embodiment of the present application provides a method for generating voxel-level structural damage supervision labels based on component-level discrete labels, which comprises the following steps:

[0014] S100, performing three-dimensional voxelization processing on the target structure to obtain a three-dimensional voxel grid and a component-to-voxel mapping relationship; obtaining at least one point source load event parameter and a discrete damage level label of each component.

[0015] S200, based on the spatial relationship and attributes of each voxel and the point source load event parameter, calculating a continuous vulnerability quantitative value of each voxel in the three-dimensional voxel grid.

[0016] S300, for each component, determining the number of target voxels to be marked as damaged in the component according to the discrete damage level label corresponding to the component.

[0017] S400, for each component, constructing a loss labeling probability distribution on the voxel space inside the component based on the continuous vulnerability quantitative values of all voxels inside the component.

[0018] S500, according to the loss labeling probability distribution, performing probability sampling in each component to extract the number of voxels, and marking the extracted voxels as damaged voxels; aggregating the labeling results of all components to generate a voxel-level damage supervision label corresponding to the three-dimensional voxel grid.

[0019] The present application has at least the following beneficial effects:

[0020] 1. Effective conversion from coarse-grained engineering labels to fine-grained supervision data is achieved: the method of the present application, through three-dimensional voxelization of the structure and establishment of component mapping, calculation of continuous vulnerability based on the spatial relationship between load events and voxels, mapping of discrete damage levels to target voxel numbers, construction of probability distribution in the component based on vulnerability and probability sampling, systematically realizes for the first time that only based on component-level discrete damage labels, voxel-level supervision labels can be automatically generated for training three-dimensional segmentation network. This directly solves the core bottleneck that the model cannot be trained due to the lack of fine annotation data in this field.

[0021] 2. The generated supervision label has inherent physical rationality: the method calculates a continuous vulnerability value based on the spatial relationship and attributes of the point source load event parameters and each voxel, which is determined by the load effect and material strength, ensuring that the generated damage distribution follows the basic physical law that the area near the load source is more vulnerable and the low strength area is more vulnerable. Based on this vulnerability value and through a normalized function with a temperature coefficient, a probability distribution is constructed on the spatial of the voxels inside the component, so that the spatial pattern of the final label conforms to engineering intuition and is superior to simple uniform broadcast labels.

[0022] 3. Reasonable mesoscopic randomness is introduced under macroscopic constraints: the method maps each component's discrete damage level label to the target voxel number through a pre-set mapping rule, ensuring that the total number of damage voxels of each component conforms to its given macroscopic damage level, meeting the overall constraints of engineering evaluation. At the same time, through the mechanism of probability sampling according to the probability distribution, the selective randomization of the spatial position of the damage voxels inside the component is realized. This overall determination and local randomization feature allows the same coarse label to generate multiple reasonable fine labels, effectively increasing the diversity of training data and helping to improve the generalization ability of the model.

[0023] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0025] Figure 1 The flowchart of the method for generating voxel-level structural damage supervision labels based on component-level discrete labels provided by the embodiments of the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application. The use herein of the terms "and / or" includes a set of one or more associated listed items.

[0028] It is to be understood that some of the example embodiments are described in terms of a process or method depicted as a flowchart. Although a flowchart can describe operations as a sequential process, many of the operations can be performed in parallel, concurrently or simultaneously. In addition, the order of the operations can be re-arranged. A process is terminated when its operations are completed, but could also terminate without completing its operations due to, for example, a system shutdown. A process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0029] The present application provides a method for generating voxel-level structural damage supervision labels based on component-level discrete labels, aiming to solve the following technical problems:

[0030] Under the condition of only component-level discrete damage grades (coarse supervision), how to automatically and reasonably generate voxel-level (fine supervision) damage labels to provide the required supervision data for training three-dimensional segmentation networks.

[0031] Ensure that the generated voxel-level labels meet multiple physical and engineering constraints simultaneously:

[0032] Consistent with the spatial relationship of the load source (the closer to the load, the more affected and the more vulnerable);

[0033] Consistent with the material or component strength distribution (the lower the strength, the more vulnerable);

[0034] Consistent with the given component-level damage grade constraint (different grades correspond to different proportions or numbers of damaged voxels);

[0035] Possess controllable randomness, so that the same coarse label can generate multiple reasonable fine labels to avoid model overfitting and improve its generalization ability.

[0036] To solve the above technical problems, the present application provides a method for generating voxel-level structural damage supervision labels based on component-level discrete labels, the core of which is: through physical heuristic calculation, a continuous vulnerability representation quantity of each voxel is obtained, and then according to the component grade constraint, the specific damaged voxels are determined in a probabilistic sampling manner, so as to interpret the discrete component-level labels as fine-grained spatial distribution consistent with physical laws.

[0037] As Figure 1As shown, the method for generating a voxel-level structural damage supervision label based on component-level discrete labels provided by the application can include the following steps:

[0038] S100, performing three-dimensional voxelization processing on the target structure to obtain a three-dimensional voxel grid and a component-to-voxel mapping relationship; obtaining at least one point source load event parameter and a discrete damage level label of each component.

[0039] In the present application, the target structure refers to an engineering structure entity composed of components (such as beams, columns, plates, walls, etc.) with clear engineering semantics, which needs to be predicted for damage.

[0040] In the present application, the three-dimensional voxel grid is constructed based on the three-dimensional geometric information and physical attribute information of the target structure.

[0041] In the context of the present application, three-dimensional geometric information refers to a set of data used to uniquely determine the spatial shape, position and size of each component in the engineering structure, and is the basis for constructing the voxelization model and the component mapping relationship. Its specific forms and sources include but are not limited to:

[0042] Parametric geometric model: such as data generated by computer-aided design (CAD) software, in which components are defined by type (such as beams, columns) and key parameters (such as length, cross-sectional size, spatial coordinates).

[0043] Building information model (BIM) data: a digital building model containing complete geometric information and non-geometric attributes (such as materials, strength).

[0044] Three-dimensional surface mesh model: such as surface mesh data (such as.stl,.obj format files) composed of triangular facets, which can be obtained by three-dimensional laser scanning, photogrammetry or exported from simulation software.

[0045] Constructive solid geometry (CSG) representation or existing voxelized representation itself.

[0046] In the context of the present application, physical attribute information refers to engineering parameters related to the mechanical properties of structural materials or the damage resistance capacity of component cross sections. It does not determine the spatial shape and position of the component, but determines the intrinsic response characteristics of the component when subjected to load. Its specific content includes but is not limited to:

[0047] Material mechanical properties: such as the elastic modulus, yield strength, ultimate compressive / tensile strength, density, etc. of the material.

[0048] Cross-sectional geometric properties: such as cross-sectional area, moment of inertia, cross-sectional modulus, etc., which directly reflect the resistance of the component under bending moment, shear force, etc.

[0049] Other engineering indicators: such as pre-evaluation of component vulnerability index, durability grade, etc.

[0050] In the present application, the three-dimensional voxel grid is obtained by discretizing the continuous space of the engineering structure, which is a key step to convert the continuous geometry and attribute information in the real world into a regularized digital model that can be processed by a computer. The specific construction process is a standardized data processing pipeline, mainly including the following steps:

[0051] S11, input and analysis: input the three-dimensional geometric information of the structure (such as BIM model, CAD file or triangular mesh) and the corresponding physical attribute information (usually in the form of material table, component attribute list, etc. associated with the geometric model or provided separately). Analyze the input data, extract the geometric topological relationship and attribute mapping.

[0052] S12, spatial discretization and grid framework generation: according to the preset voxel space resolution (i.e. the physical size represented by each voxel, denoted as s, usually in meters), automatically calculate a regular three-dimensional axis-aligned bounding box that can completely enclose the input structure. Then, the bounding box is uniformly subdivided in three dimensions (depth D, height H, width W) according to the preset voxel space resolution s, thereby dynamically determining the dimensions of the grid DxHxW. Generate a three-dimensional voxel grid framework composed of DxHxW regular cubic units, initially empty. Each voxel is uniquely identified by its three-dimensional integer index (d, h, w), where:

[0053] d is the index of the depth direction (usually corresponding to the X axis or the front-back direction of the three-dimensional space), with a value range of [0, D-1];

[0054] h is the index of the height direction (usually corresponding to the Y axis or the vertical direction of the three-dimensional space), with a value range of [0, H-1];

[0055] w is the index of the width direction (usually corresponding to the Z axis or the left-right direction of the three-dimensional space), with a value range of [0, W-1].

[0056] The index triplet (d, h, w) corresponds to a regular cubic region defined by [x min +w×s, x min +(w+1)×s]×[y min +h×s, y min +(h+1)×s]×[z min +d×s, z min +(d+1)×s], where (x min , y min , z min ) is the minimum corner coordinate of the bounding box.

[0057] S13, assigning attributes and establishing relationships: This step is the core of voxelization, aiming to fill semantic and attribute information for each voxel in the grid framework. The specific operation is as follows:

[0058] (1) Component attribution determination: Traverse each voxel unit in the voxel grid, and determine its attribution to which component unit by calculating the spatial position relationship between its geometric center (or overall area) and each component geometric model. The present application can use various efficient algorithms to realize this determination, including but not limited to:

[0059] Accurate geometric determination method: Directly calculate the spatial position relationship between the geometric shape of the component (defined by the vertex coordinates, surface triangular mesh or parameterized geometric description of the component) and each voxel unit, and determine whether the voxel belongs to the component through geometric intersection test or inclusion determination.

[0060] Bounding box filling method: Calculate the axial bounding box (AABB) of each component in the voxel grid coordinate system, and then use voxelization filling algorithm (such as scan line filling) to quickly determine all voxels belonging to the component within the three-dimensional space region defined by the bounding box. This method is more efficient and suitable for the scene where the shape of the component is regular.

[0061] (2) Establishing mapping relationship and recording attributes: According to the above determination results:

[0062] Establishing mapping relationship: The system records the corresponding relationship between each voxel and its attribution component, forming the component-voxel mapping relationship. This relationship is the basis for all subsequent cross-scale operations.

[0063] Assigning geometric derived attributes: The voxel can record the center coordinates and other attributes derived from the original geometric information.

[0064] Associated physical attributes: According to the component to which the voxel belongs, query the physical attribute information obtained in step S11, and store the corresponding material strength, cross-sectional modulus and other parameters as the basic physical attribute values of the voxel.

[0065] Through the above process, a complete and structured three-dimensional voxel grid data is finally generated, which integrates spatial discrete framework (geometry), component semantic mapping (attribution) and basic physical attributes (material).

[0066] In the present invention, the component-to-voxel mapping explicitly records the component unit to which each voxel in the three-dimensional voxel grid belongs, constituting the key index basis for all subsequent cross-scale calculations (such as feature aggregation, statistical quantity calculation). In a preferred embodiment of the present invention, this mapping relationship is instantiated and stored in the form of a component semantic mask, a specific data structure. The component semantic mask is a three-dimensional integer tensor that is completely aligned with the three-dimensional voxel grid in spatial dimensions and has the same size (D, H, W). In this mask tensor, each voxel position stores a unique integer identifier that indicates which specific component unit this voxel belongs to. For example, all voxels belonging to "Beam No. 1" have their mask values set to 1; voxels belonging to "Column No. 2" have their mask values set to 2; background voxels that do not belong to any component are assigned a specific reserved value, such as 0. This representation converts the abstract membership relation into a dense, regular, and structured data that can be directly indexed and efficiently accessed through memory. The mask is derived from the geometry processing flow of this step and serves as the core data basis for implementing cross-scale information aggregation and consistency constraints in subsequent network training and prediction.

[0067] In the present invention, the point source load event parameter is used to quantify the external excitation acting on the target structure, which is the physical driving force source driving the damage generation simulation. A point source load event e is mathematically modeled as an excitation source applied at a specific spatiotemporal point with a certain intensity and physical meaning. Its parameterized representation at least includes:

[0068] Spatial position: the three-dimensional coordinates of the event occurrence point, used to calculate the spatial distance d between it and each voxel in the structure.

[0069] Intensity scalar: a non-negative real number Q, used to quantify the physical intensity of the event. Its specific engineering meaning is closely related to the load type, for example:

[0070] For explosion or shock wave events, Q can represent TNT equivalent (kg), shock wave overpressure peak value (Pa), or impulse (Pa·s).

[0071] For impact events, Q can represent the kinetic energy (J) or momentum (kg·m / s) of the impact body.

[0072] For local fire or energy release, Q can represent the heat release rate (kW) or total energy (J).

[0073] Directionality parameter: for loads with clear directionality (such as directional blasting, inclined impact), the directionality parameter can further include the action direction vector of the load, which is used to calculate the angle with the component surface to introduce a directional attenuation factor.

[0074] Load type identification: used to distinguish different physical mechanisms of load (e.g. "explosion", "impact", "fire") so as to select or adjust the corresponding physical impact model in the subsequent vulnerability calculation.

[0075] In specific implementation, a single point source load event can be processed, or a set of multiple simultaneous or sequential events E = {e1, e2, …, en} can be processed. m} and its comprehensive impact is calculated by superposition principle, and m is the total number of point source load events.

[0076] In the present application, the discrete damage level label of each component is a supervisory constraint condition for the method operation, representing the macroscopic evaluation results of the overall damage state of the component that can be obtained in engineering practice. The label is a discrete and graded description of the overall performance degradation or damage degree of the component after being subjected to the load event.

[0077] Each component c is assigned a label L(c) whose value comes from a pre-set, ordered discrete set, such as {0, 1, 2, 3} or {"undamaged", "mild damage", "moderate damage", "severe damage / destruction"}. The grade value is usually positively correlated with the severity of damage (e.g. 0 represents undamaged, and the larger the value represents the more severe the damage).

[0078] These discrete damage level labels usually come from:

[0079] Engineering evaluation report: given by professional engineers according to on-site investigation, detection data (such as crack width, deformation measurement) or experience.

[0080] Post-processing of simplified numerical simulation results: by running a fast, simplified model of mechanical analysis (such as component-level Pushover analysis), the overall response of the component (such as maximum displacement, plastic hinge state) is mapped to a discrete damage level.

[0081] Historical accident data record: derived from the actual damage of similar structures under similar disasters.

[0082] In the subsequent step, this level label will be used to determine the total number or proportion of voxels in the component that should be marked as damaged (such as S300 step), so as to ensure that the generated fine-grained label is consistent with the given engineering evaluation conclusion in macroscopic statistics, realizing the constraint transformation from coarse supervision to fine supervision.

[0083] Through the above steps, the preparation of all necessary input data is completed: the fine-grained spatial calculation framework (voxel grid and mapping) is established, the external excitation (load event parameters) is determined, and the macroscopic supervision target (component damage level) is obtained. These three are collectively used as the input of the method of the present application, supporting the subsequent vulnerability calculation and label generation process.

[0084] S200, based on the spatial relationship and properties of the point source load event parameters and each voxel, calculate the continuous vulnerability quantization value of each voxel in the three-dimensional voxel grid.

[0085] In S200, for each voxel v in the three-dimensional voxel grid (v takes values ​​from 1 to n, where n is the total number of voxels in the three-dimensional voxel grid), its continuous vulnerability quantization value U(v) satisfies the following condition:

[0086] U(v) = I(v) / (S(v) + ε);

[0087] Where I(v) is the cumulative influence of all point source load events on voxel v, S(v) is the strength property of voxel v, whose value is inherited from the material or section properties of the component to which the voxel belongs; ε is a very small positive constant, whose main function is to ensure the numerical stability of the formula and prevent calculation overflow when S(v) approaches zero. In one embodiment of the present invention, ε can take a value on the order of 1e-5 or smaller, and the specific value can be determined according to the floating-point precision of the calculation system.

[0088] Furthermore, I(v) is a linear superposition of the effects of each point source load event, satisfying the following condition:

[0089] ;

[0090] Among them, I i (v) represents the i-th point source load event e i The independent influence of voxel v, where i ranges from 1 to m, and m is the total number of point source load events. i (v) By point source load event e i Intensity Q i The product of this and a distance-based attenuation factor is given by the general formula: I i (v) = Q i ×φ(d(v,e i ), d(v, e) i () represents the distance from the center point of voxel v to e. i The distance to the location is specifically a three-dimensional Euclidean distance.

[0091] Furthermore, the attenuation factor φ(d) is a monotonically decreasing function, used to simulate the attenuation of physical influence with distance, where d is equivalent to d(v, e). i In different embodiments of the present invention, one of the following specific forms may be adopted (but is not limited to):

[0092] 1 / (1+k×d(v,e i ));

[0093] 1 / (1+k x d(v, e i )+ε);

[0094] 1 / (d(v, e i )+ε);

[0095] 1 / (1+k x (d(v, e i ) 2 );

[0096] A customized piecewise function. The piecewise function allows different decay laws in different distance intervals to more accurately fit specific physical phenomena. For example, a piecewise function available is defined as follows: ;

[0097] where k is the decay coefficient, used to control the speed of decay. D0 and D1 are distance thresholds, k1 and k2 are piecewise decay coefficients, and A is the continuity connection value at D1. This example indicates that the impact remains maximum at very close distances, adopts linear reciprocal decay in the medium distance interval, and adopts square inverse acceleration decay at long distances. Those skilled in the art can understand that the above example is only to illustrate the concept of the piecewise function, and different piecewise intervals and function forms can be defined according to specific engineering knowledge.

[0098] Among the various possible forms of the decay factor, 1 / (1+k x d(v, e i ) or its variant 1 / (1+k x d(v, e i )+ε) with a fine tuning constant ε is one of the common and good practical forms. This form can be physically explained as hyperbolic decay of the impact quantity with distance, and has good numerical stability and efficiency in calculation, which can effectively balance the high sensitivity at close distances and the rapid decay at long distances, so it is adopted in multiple embodiments of the present application.

[0099] Further, to more accurately simulate complex engineering scenarios, a weight factor η can also be introduced to modify the general formula of the independent impact quantity I i (v) when calculating, that is, it is also multiplied by a weight factor: I i (v)=Q i x φ(d(v, e i )) x η.

[0100] The weight factor is calculated based on one or more of the following engineering and physical factors, and its general form is η=η dir x η obs x η thick (multiplication or other combination when considering multiple factors):

[0101] Directional factor η dir: Related to the angle θ between the load action direction and the surface normal of the component that voxel v belongs to, a feasible calculation formula is: η dir = max(0, cos(θ)). This formula simulates the physical intuition that the impact is the largest when the incident is perpendicular (θ = 0°) and the impact is zero when the incident is grazing (θ ≥ 90°).

[0102] Occlusion factor η obs : Related to whether the spatial line of sight between event e i and voxel v is blocked by other components. A feasible implementation is to perform a ray casting check, and if the line of sight is completely blocked, set η obs = 0; if partially blocked or unblocked, set η obs = 1. A more refined model can assign a decay value between 0 and 1 according to the material and thickness of the occluding object.

[0103] Construction factor η thick : Related to the equivalent thickness t v of the component that voxel v belongs to, used to reflect the stronger local resistance of thicker components. A feasible calculation formula is: η thick = exp(-β × t v ) or η thick = 1 / (1 + γ × t v ), where β, γ are decay coefficients greater than zero. This factor shows that under the same external influence, the effective vulnerability of the area with greater thickness will be reduced.

[0104] Through this step, a continuous vulnerability quantification value that comprehensively reflects the spatial position (affected by the load), intrinsic properties (resistance capacity), and complex environmental factors (direction, occlusion, construction) of each voxel is calculated. The larger this value is, the more physically fragile the voxel is, and the more likely it is to be marked as damaged in the subsequent steps.

[0105] S300, for each component, according to the discrete damage level label corresponding to the component, determine the number of target voxels in the component to be marked as damaged.

[0106] The core of this step is to map the discrete damage level label of each component to a quantity guide value indicating the internal damage degree of the component through a preset mapping rule, and determine the target voxel number of the component based on the quantity guide value. The quantity guide value is an intermediate variable, and its specific form and acquisition method can be defined flexibly according to engineering needs.

[0107] In one illustrative embodiment, the quantity guidance value is a fixed damage ratio. By querying a pre-set static mapping table, a ratio corresponding to the member grade is obtained. The target voxel number Nc of member c satisfies the following condition:

[0108] Nc=rounddown(r L（c） ×|Vc|).

[0109] where r L（c） is the fixed damage ratio corresponding to the member damage grade L(c) obtained by querying the mapping table, and |Vc| represents the total number of voxels occupied by member c. rounddown() represents the rounding down operation, which ensures that Nc is an integer not exceeding the theoretical maximum value.

[0110] To further improve the rationality of the mapping, the pre-set mapping rule is a mapping table associated with the member type. For different types of members such as beams, columns, plates, and walls, the same discrete damage grade corresponds to different damage ratios to reflect the different mesoscopic damage characteristics that different types of members may exhibit under the same macroscopic grade. That is, the quantity guidance value depends not only on the damage grade but also on the member type (beams, columns, plates, walls, etc.), thus reflecting the mesoscopic damage differences of different types of members under the same macroscopic grade.

[0111] To increase the diversity of generated labels and improve the model generalization ability, the quantity guidance value can not be set as a fixed value, but a ratio value r min is randomly sampled from a pre-set interval [r max , r p ] according to a uniform distribution. At this time:

[0112] Nc=rounddown(r p ×|Vc|).

[0113] This method introduces controllable random disturbance under the premise of meeting the grade constraint.

[0114] To establish a more direct association between damage quantity and external excitation strength, the quantity guidance value is also scaled and adjusted according to the total strength of the point source load event. In an extended embodiment, the target voxel number Nc is not only related to the damage grade L(c) but also coupled with the overall strength Q total of the load event. A feasible implementation is:

[0115] Nc=rounddown(r L（c） ×|Vc|×f(Q total )).

[0116] where f(Q total) is a scaling function related to the total intensity, for example, f(Q total ) = min(1.0, a x Q total ), where a is a positive intensity-damage scaling factor. The min(·) function ensures that the scaling is upper bounded by 1.0. The value of a can be calibrated by engineering experience or simplified mechanical analysis. A typical calibration method is as follows: first, set a reference load intensity Q ref with clear physical meaning (e.g., design reference load or a certain typical accident load), and specify the target value f target (e.g., 0.8) that the scaling factor f(Q ref ) should reach when the total load intensity equals this reference value. Then, according to the formula f target = min(1.0, a x Q ref ), the value of a is calculated (i.e., a = f target / Q ref ). The values of Q ref and f target can be reasonably determined by the skilled person in the art according to the structural characteristics, material properties, and safety margin requirements of the actual application scenario, thereby completing the configuration of a.

[0117] By any one or a combination of the above methods, a clear target number of damaged voxels is determined for each damaged component based on the quantity guidance value, providing a quantitative constraint for the next step of probabilistic sampling within the component.

[0118] S400, for each component, based on the continuous vulnerability quantification values of all voxels inside the component, a loss marker probability distribution on the voxel space inside the component is constructed.

[0119] This step aims to convert the continuous vulnerability quantification values U(v) of all voxels within each component c into a spatial probability distribution P c (v), so that voxels with higher vulnerability are more likely to be selected as damaged samples, thereby incorporating physical prior into the sampling process.

[0120] In an embodiment, the loss marker probability distribution is constructed by a softmax function with a temperature coefficient. For a component c and the voxel set Vc it occupies, the probability P c (v) of any voxel v (v e Vc) being selected is calculated as follows:

[0121] ;

[0122] where U(g) is the continuous vulnerability quantification value of the gth voxel of component c, exp() is the natural exponential function. T is the temperature coefficient, T > 0, which is an adjustable hyper-parameter. T controls the concentration degree of the probability distribution. When T is small, the value difference of U(v) / T is amplified, which leads to the probability of high vulnerability voxels being significantly increased, and the sampling will be highly concentrated in the area with the highest vulnerability. When T is large, the value difference of U(v) / T is reduced, and the probability distribution tends to be flat, and the sampling will cover each voxel in the component more evenly. By adjusting T, the balance between determinism and randomness of the generated label can be flexibly controlled.

[0123] In actual calculation, to avoid numerical overflow of the exponential function exp(·) due to too large input value, a numerically stable softmax can be used. A general and effective method is to subtract the maximum value M from all U(v) / T before calculating the exponential: ; where M = max{U(g) / T} g∈Vc This mathematical transformation does not change the result of the probability distribution, but ensures the numerical stability of the calculation.

[0124] As a robust design, when the vulnerability quantification values U(v) of all voxels in component c are all very small (for example, they are all approximately zero, or the difference between them is much smaller than the numerical precision), the above softmax calculation may cause the probability distribution to lose meaning. In this case, it can be degenerated to a uniform distribution, i.e.: c (v) = 1 / |Vc|. This ensures that an effective probability distribution can be generated for subsequent sampling in any case.

[0125] Through the above method, a probability distribution matched with the internal physical vulnerability of each component is constructed, which provides a basis for the next step of performing constraint probability sampling.

[0126] S500, according to the damage label probability distribution, performing probability sampling in each component to extract the number of target voxels, and marking the extracted voxels as damage voxels; aggregating the marking results of all components to generate a voxel-level damage supervision label corresponding to the three-dimensional voxel grid.

[0127] The goal of this step is to determine the specific damage voxels in each component through sampling according to the probability distribution constructed in step S400 and the number of target voxels determined in step S300, and finally synthesize a complete voxel-level supervision label.

[0128] Specifically, based on the probability distribution P c(v), within each component c, probabilistic sampling is performed to draw Nc voxels, and these voxels are labeled as damage to form the damage voxel set Dc of the component. After performing this operation for all components, the global label matrix Y is synthesized by updating the values of the voxels in the matrix Y corresponding to the positions of the voxels in the damage voxel sets to 1 (indicating damage). Finally, the matrix Y is the generated voxel-level damage supervision label. If the damage level of component c is L(c) = 0, all voxels inside the component are labeled as undamaged, and there is no need for sampling. The damage voxel set of the component is an empty set, and all voxels of the component remain 0 in the matrix Y.

[0129] Initialize a global label matrix Y with the same size as the three-dimensional voxel grid, and all elements of the matrix Y are 0. Then, traverse all components and their damage voxel sets, and update the values of the voxels in the matrix Y corresponding to the positions of the voxels in the damage voxel sets to 1 (indicating damage). Finally, the matrix Y is the generated voxel-level damage supervision label. If the damage level of component c is L(c) = 0, all voxels inside the component are labeled as undamaged, and there is no need for sampling. The damage voxel set of the component is an empty set, and all voxels of the component remain 0 in the matrix Y.

[0130] In an embodiment, the probabilistic sampling employs non-replacement sampling. For each component c, if L(c) > 0, the non-replacement sampling is performed according to the probability distribution P c (v) to draw Nc voxels. This ensures that the damaged voxels are not repeated. In another embodiment, the probabilistic sampling employs a stratified sampling strategy to better control the spatial structure of the damage. The specific steps include:

[0131] Distance calculation: for each voxel v in component c, calculate the shortest distance d min (v) to all point source load event positions.

[0132] Ring division: define a set of increasing global distance thresholds 0 = d0< d1< … < d K-1 < d K = +∞. According to the distance of each voxel falling into the to-be-determined region, the voxel is classified into the corresponding ring, and the specific rules are as follows: if d min (v) ∈ [d K-1 , d K ], the voxel is classified into the hth distance ring. According to this, all voxels in the component are divided into K consecutive distance rings, for example, 0-d1, d1-d2, …, d K-1 -∞. The value of h is 1 to K.

[0133] Quota allocation: according to a preset rule (for example, according to the proportion of the number of voxels in each ring, or allocating a higher proportion to the near-field ring), the total sampling number Nc is allocated to each ring to obtain the quota N ch of each ring.

[0134] In-ring sampling: in each ring h, non-replacement sampling is performed according to the normalized probability distribution P ch (v) in the ring to draw N ch voxels.

[0135] To simulate the continuous regions rather than discrete points often presented in real damage, the sampling process can introduce a spatial connectivity constraint. One implementation is to perform a three-dimensional connected component analysis on the candidate damage voxels after they are preliminarily extracted according to the probability distribution, and then preferentially retain the voxels in the largest connected components until the quantity requirement is met. This is intended to make the generated damage labels form continuous clusters in space.

[0136] The voxel-level damage supervision label generated in step S500 is not limited to a binary form. In extended applications, various forms of supervision signals can be output to adapt to different training objectives:

[0137] Binary mask label: that is, Y(v)∈{0,1} as mentioned above.

[0138] Soft label / probability label: Y(v)∈[0,1], which can directly output the sampling probability P c (v) or the calibrated value for training a probabilistic segmentation model.

[0139] Hierarchical damage label: Y(v)∈{0,1,2,…}, for example, can be divided into no damage, mild damage, severe damage, etc. discrete levels according to the vulnerability quantization value U(v).

[0140] Regression field label: Y(v)∈R, which can output continuous damage indicators (such as damage factors, plastic strain) for regression tasks.

[0141] The sampling process is probabilistic. To support method reproducibility and achieve efficient data augmentation, the random seed used during sampling can be recorded. By fixing different random seeds, multiple voxel-level labels that are statistically reasonable but spatially diverse can be generated for the same set of inputs (structure, load, component label), greatly enriching the training data set.

[0142] The voxel-level damage supervision label generated by the present application provides key data support for training three-dimensional segmentation networks for engineering applications. The network trained based on such labels can be further used to perform various advanced structural safety analysis tasks, including but not limited to: structural residual carrying capacity evaluation, repair and reinforcement priority ranking, and rapid digital reconstruction of accident or disaster scenes.

[0143] To improve the processing efficiency of large and complex structures, the implementation of the present method can use a parallel computing framework. Thanks to the component-voxel mapping relationship established in step S100 and the data independence between steps, the system can naturally use components as the basic unit, or divide the three-dimensional space into multiple sub-regions, and perform parallel processing of the vulnerability calculation, probability distribution construction and sampling process in steps S200 to S500, thereby significantly shortening the overall label generation time.

[0144] Further, the method provided by the present application further comprises the following steps:

[0145] S600, the generated voxel-level damage supervision label is mapped to a voxel grid with different resolutions through upsampling or downsampling to generate a multi-resolution damage supervision label.

[0146] To adapt to the training needs of three-dimensional segmentation networks with different resolutions, or to realize progressive analysis from coarse to fine, the present application can further process the voxel-level damage supervision label generated by S500 to generate a multi-resolution damage supervision label.

[0147] The core operation of this step is resolution mapping: the damage supervision label at the original resolution (denoted as s0) is converted to a voxel grid at another target resolution (s1) through upsampling or downsampling techniques.

[0148] Downsampling: when the target resolution s1 is coarser than the original resolution s0 (i.e., s1 > s0), the label information of multiple fine-grained voxels is aggregated into a coarse-grained voxel through pooling (such as max pooling, average pooling) or subsampling operations to generate a low-resolution, large-receptive-field supervision label. For example, each 2x2x2 fine-grained voxel block is combined into 1 coarse-grained voxel.

[0149] Upsampling: when the target resolution s1 is finer than the original resolution s0 (i.e., s1 < s0), the coarse-grained label is assigned to a finer voxel grid through interpolation (such as nearest neighbor interpolation, trilinear interpolation) or transposed convolution operations to generate a high-resolution, more detailed supervision label. For binary mask labels, nearest neighbor interpolation is commonly used to maintain class boundaries.

[0150] The application value of generating multi-resolution damage supervision labels includes:

[0151] Multi-scale network training: the generated labels with different resolutions can be used to train three-dimensional segmentation networks with encoder-decoder structure or multi-scale feature fusion module, so that the network can learn both the global context information and the local fine features of the damage, and improve the model performance.

[0152] Progressive analysis and computational optimization: a model can be first trained on low-resolution labels for preliminary screening or coarse segmentation, and then fine-tuned using high-resolution labels for key areas to achieve a balance between accuracy and efficiency.

[0153] Data adaptability: it can provide matching supervision signals for various downstream models with different input resolutions, enhancing the generality of the method output data.

[0154] By integrating the step S600, the present application can output a multi-resolution supervised label set, providing a more abundant and flexible data basis for intelligent analysis of structural damage.

[0155] Further, the method further comprises: verifying the generated supervised label by using the trained three-dimensional segmentation model, calculating the uncertainty of the label; and adjusting the probability distribution in the step S400 or the number of target voxels in the step S300 according to the uncertainty feedback.

[0156] To continuously improve the physical rationality of the generated supervised label and the training efficiency of the downstream model, in an extended embodiment of the present application, a verification and feedback adjustment closed-loop mechanism based on the trained model can be introduced. The closed-loop mechanism combines the method of the present application with the downstream three-dimensional segmentation model, realizing self-optimization of the label quality. The closed-loop mechanism includes the following stages:

[0157] Initial stage: running the method of the present application (S100-S500) to generate an initial voxel-level damage supervised label.

[0158] Model training stage: training a three-dimensional segmentation model using the initial label set.

[0159] Verification and feedback stage: using the trained model to predict the input (structure, load) corresponding to the newly generated or historically generated label, calculating the label uncertainty by analyzing the difference or inconsistency between the model prediction and the generated label. The uncertainty is quantified as a feedback signal.

[0160] Parameter adjustment stage: according to the feedback label uncertainty, adaptively adjusting the key parameters in the core steps of the present application (such as the number mapping rule in S300 or the probability distribution temperature coefficient T in S400).

[0161] Iteration stage: using the adjusted parameters to regenerate improved supervised labels, and optionally retraining or fine-tuning the model, forming an iterative optimization cycle.

[0162] The label uncertainty is used to quantify the reliability of the generated label or the consistency with the model prediction. A feasible calculation method is based on the prediction-label difference:

[0163] (1) For a set of samples, use the trained three-dimensional segmentation model to perform forward propagation to obtain the predicted damage probability map of each voxel.

[0164] (2) The generated binary supervised label Y(v) ∈ {0, 1} is regarded as the true value.

[0165] (3) Calculate the mean absolute error (MAE), cross-entropy (CE) or variance of the predicted probability around the model decision boundary (e.g. 0.5) between the predicted probability and the hard label within each sample or each component. This difference value can be used as a measure of the uncertainty of the label of the sample or component. The greater the difference, the higher the uncertainty.

[0166] According to the calculated uncertainty, the previous parameters can be automatically adjusted:

[0167] Adjust the temperature coefficient T in S400: If the label uncertainty of a certain component or a certain type of sample is consistently high, it may indicate that the internal probability distribution is too concentrated or ambiguous. The system can adjust the temperature coefficient T corresponding to this part of data to make the probability distribution more uniform and introduce more diversity to explore more reasonable spatial patterns when generating labels next time.

[0168] Adjust the mapping rule in S300: If the overall uncertainty indicates that the label is too damaged or under-damaged, the mapping relationship between the damage level and the damage ratio r can be fine-tuned. For example, if the model tends to predict more damage in low-level components than to generate labels, the damage ratio value corresponding to this level can be appropriately increased.

[0169] This closed-loop mechanism gives the method the ability to improve itself, and the final output is a verified and optimized high-quality supervised label set that is more compatible with downstream models. It not only improves the reliability of the label itself, but also indirectly improves the performance upper limit of all downstream analysis models that rely on this label for training.

[0170] In summary, the method for generating voxel-level structural damage supervision labels based on component-level discrete labels provided by the present application has at least the following advantages:

[0171] 1. Breaks through the data bottleneck and realizes the transformation from coarse supervision to fine supervision: successfully converts the easily accessible component-level discrete damage evaluation (coarse label) in engineering into the three-dimensional voxel-level supervision signal (fine label) required by deep learning, greatly reducing the dependence on scarce and expensive fine annotation data, and providing a data foundation for training high-performance segmentation models.

[0172] 2. The generated label has clear physical meaning and high quality: by calculating the voxel-level vulnerability quantification value, the generated label strictly follows the basic mechanical law of near load source vulnerability and low intensity zone vulnerability, and has a reasonable spatial distribution, which can provide correct supervision signals for the model and guide it to learn the real damage pattern.

[0173] 3. The method realizes the unification of constraints and randomness: the method introduces probability-based sampling randomness in the component while strictly meeting the component-level overall damage constraint. This controllable randomness simulates the uncertainty of real damage, can generate multiple reasonable spatial distribution variants for the same training sample, effectively acts as data augmentation, and significantly improves the generalization ability and robustness of the trained model.

[0174] 4. The method is flexible and has strong scalability: core modules such as vulnerability calculation model, grade-number mapping rule, and sampling strategy can be customized and extended according to specific engineering knowledge (such as introducing direction factor, connectivity constraint, and multi-resolution output), which can adapt to different structure types, load conditions, and evaluation standards.

[0175] 5. High automation and strong practicability: the method realizes the end-to-end automatic generation process from coarse label to fine label, and its calculation efficiency is much higher than manual labeling or full physical simulation, which provides a feasible technical means for quickly and low-cost construction of large-scale structure damage training dataset, and effectively promotes the practical application of structure intelligent evaluation technology.

[0176] The embodiment of the application also provides an electronic device, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method described in the embodiment of the application.

[0177] The embodiment of the application also provides a computer readable storage medium storing computer executable instructions, and the computer executable instructions are used to execute the method described in the embodiment of the application.

[0178] It should be understood that the various forms of flow shown above can be used to reorder, add or delete steps. For example, the steps described in the present application can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in the present application can be achieved, and the present application does not limit this.

[0179] The above specific embodiments do not constitute a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent replacement and improvement within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for generating voxel-level structural damage monitoring tags based on component-level discrete tags, characterized in that, The method includes the following steps: S100: Perform three-dimensional voxelization on the target structure to obtain a three-dimensional voxel mesh and the mapping relationship between components and voxels; obtain at least one point source load event parameter and discrete damage level labels for each component; S200, based on the spatial relationship and properties of the point source load event parameters and each voxel, calculate the continuous vulnerability quantization value of each voxel in the three-dimensional voxel mesh; S300, for each component, determine the number of target voxels to be marked as damage within the component based on the discrete damage level label corresponding to the component; S400: For each component, based on the continuous vulnerability quantization values ​​of all voxels within the component, construct the loss label probability distribution in the voxel space within the component; S500, based on the loss label probability distribution, perform probability sampling within each component to extract the target number of voxels, and label the extracted voxels as damage voxels; aggregate the labeling results of all components to generate voxel-level damage supervision labels corresponding to the three-dimensional voxel mesh.

2. The method according to claim 1, characterized in that, In S200, for each voxel v in the three-dimensional voxel mesh, the continuous vulnerability quantization value U(v) of voxel v satisfies the following condition: U(v) = I(v) / (S(v) + ε); Where I(v) is the cumulative effect of all point source load events on voxel v, S(v) is the intensity attribute of voxel v, and ε is a positive constant.

3. The method according to claim 2, characterized in that, I(v) satisfies the following condition: ； Among them, I i (v) represents the i-th point source load event e i Independent effect of I on voxel v i (v) By point source load event e i Intensity Q i It consists of a product of a distance-based attenuation factor, where i ranges from 1 to m, and m is the total number of point source load events.

4. The method according to claim 3, characterized in that, The attenuation factor is any of the following functions: 1 / (1+k×d(v,e i )); 1 / (1+k×d(v,e) i )+ε); 1 / (d(v),e i )+ε) 1 / (1+k×(d)(v,e) i )) 2 ); Where d(v, e) i ) for voxels v to e i The distance to the location, where k is the attenuation coefficient.

5. The method according to claim 3, characterized in that, When calculating the independent influence, a weighting factor is also multiplied, which is determined based on at least one of the following factors: The angle between the direction of the load and the normal to the surface of the component to which the voxel belongs; Spatial occlusion relationship between events and voxels; The equivalent thickness of the component to which the voxel belongs.

6. The method according to claim 1, characterized in that, In S300, a preset mapping rule is used to map the discrete damage level label of each component to a quantity guide value that indicates the degree of damage within it, and the number of target voxels of the component is determined based on the quantity guide value.

7. The method according to claim 1, characterized in that, In S400, the loss label probability distribution is constructed using a softmax function with a temperature coefficient.

8. The method according to claim 1, characterized in that, In S500, spatial connectivity constraints are introduced during sampling so that voxels marked as damage form one or more connected clusters in space, simulating the continuous morphology of actual damage.

9. The method according to claim 1, characterized in that, The voxel-level damage supervision label is any one of the following: a binary damage mask label, a soft label representing damage probability, a hierarchical label representing multi-level damage degree, or a regression field label representing continuous damage intensity.

10. The method according to claim 1, characterized in that, The method further includes the following steps: The S600 generates voxel-level damage supervision labels and maps them to voxel grids of different resolutions through upsampling or downsampling to generate multi-resolution damage supervision labels.

Images