Method and apparatus for predicting three-dimensional voxel damage based on component consistency constraints

Abstract

This invention provides a method and apparatus for predicting three-dimensional voxel damage based on component consistency constraints. The method first acquires structural information, constructs a three-dimensional voxel mesh and a mapping relationship between components and voxels, and integrates structural properties and load effects to construct a multi-channel input tensor. This tensor is then input into a specially trained three-dimensional segmentation network model to obtain initial voxel damage probabilities. The core of this invention lies in introducing cross-scale engineering consistency constraints: based on the mapping relationship, voxel predictions are aggregated into component-level damage representations, and a consistency correction is calculated according to a reasonable range defined by the component's preset damage level. This allows for the coordinated adjustment of the initial probabilities of all voxels within the component, ultimately outputting a three-dimensional damage distribution map that conforms to component-level engineering rationality. This method effectively solves three major engineering problems: class imbalance in voxel-level prediction, overall component inconsistency, and overfitting to noise labels.

Description

Technical Field

[0001] This invention relates to the field of engineering structure damage prediction technology, and in particular to a three-dimensional voxel damage prediction method and apparatus based on component consistency constraints. Background Technology

[0002] In critical scenarios such as structural safety assessment, impact analysis of blasting demolition, and impact assessment of mining and industrial accidents, accurately predicting the spatial damage distribution of structures under external loads (such as blast impacts and energy releases) is of great significance for disaster prevention and control, scheme optimization, and safety decision-making. In recent years, with breakthroughs in deep learning technology, three-dimensional convolutional neural networks (3D CNNs), especially 3D segmentation architectures represented by 3D U-Net, have been introduced into this field due to their powerful spatial feature extraction capabilities. These methods typically discretize the structure into a voxel grid and classify it at the voxel level, ultimately outputting a damage probability distribution or a binary mask, providing a new approach for automatically learning damage patterns from massive amounts of data.

[0003] However, when this paradigm, derived from biomedical image segmentation, is directly transferred to engineering physical scenarios, existing technologies reveal significant limitations, mainly in the following three aspects:

[0004] First, extreme voxel-level class imbalance makes model training prone to getting stuck in trivial solutions. In real-world engineering structures, damaged regions typically occupy only a tiny fraction of the overall volume, resulting in a significant disparity between the number of damaged and undamaged voxels. If strong voxel-level classification losses such as Binary Cross-Entropy (BCE) or Focal Loss are used for optimization at the initial training stage, the loss function will quickly be dominated by the predominantly undamaged voxels, driving the model to converge to a degenerate state where all voxels are predicted as "undamaged." This makes the model training process unstable, or even completely fail, making it difficult to learn meaningful damage features.

[0005] Secondly, existing methods generally neglect the inherent component-level semantics and overall constraints of engineering structures. Engineering structures are composed of component units with clearly defined functions and mechanical behaviors, such as beams, columns, slabs, and walls. In engineering practice, the core object of safety assessment is the component, and its conclusion is usually given in the form of the overall damage level of the component. Current mainstream 3D voxel segmentation methods only perform independent prediction and optimization at the pixel level, lacking explicit modeling of the component as a basic engineering unit. This leads to contradictions in the prediction results at the component scale that often violate common sense in engineering: for example, a component judged to be undamaged overall may contain a large number of voxels predicted to be damaged; or vice versa. This inconsistency seriously weakens the interpretability and practical engineering value of the prediction results.

[0006] Finally, in weakly supervised or synthetic data scenarios, the model is overly sensitive to label noise. Obtaining high-precision, voxel-level 3D damage ground truth values ​​from real-world scenarios is extremely difficult and costly. Therefore, training data often relies on labels generated through numerical simulations or synthesized using simplified rules. These labels inevitably contain uncertainties and noise at the voxel level. If existing methods rely solely on such noisy voxel-level labels for supervision, the model is highly susceptible to overfitting to local, random label errors, thus learning spurious damage patterns, making it difficult to generalize to real-world situations, and failing to guarantee the reasonableness of the predicted overall damage state of the component.

[0007] In summary, the fundamental problem with current structural damage prediction methods based on 3D voxel segmentation lies in treating complex engineering physics problems as purely data-driven, intensive classification tasks, failing to effectively integrate prior knowledge from the engineering domain (such as component partitioning) and evaluation logic (such as overall consistency). Therefore, a new technical framework is urgently needed to introduce structured engineering constraints into deep learning models to overcome class imbalance, improve prediction consistency, and enhance robustness to noisy data, thereby truly achieving a leap from voxel classifiers to intelligent structural damage inference engines. Summary of the Invention

[0008] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0009] According to a first aspect of the present invention, a three-dimensional voxel damage prediction method based on component consistency constraints is provided, the method comprising the following steps:

[0010] S100: Obtain the three-dimensional geometric and physical property information of the structure to be evaluated, construct a three-dimensional voxel mesh, and establish the mapping relationship between component units and voxels.

[0011] S200, construct a multi-channel input tensor, wherein the multi-channel input tensor includes at least a strength channel input tensor reflecting the inherent properties of the structure and a spatial distribution channel input tensor reflecting the influence of external loads.

[0012] S300, the multi-channel input tensor is input into a pre-trained three-dimensional segmentation network model to obtain the initial damage prediction probability of each voxel; wherein, the three-dimensional segmentation network model is optimized by an overall loss function that includes component consistency constraint loss, and the component consistency constraint loss applies an interval constraint based on a preset damage level to the component-scale aggregation result of the voxel-level prediction through the mapping relationship between component units and voxels.

[0013] S400, based on the mapping relationship between the component unit and the voxel, the initial damage prediction probability is aggregated into a component-level damage characterization quantity for each component unit.

[0014] S500, calculate the component consistency correction amount based on the component-level damage characterization quantity of each component unit and the damage ratio reference range defined by the preset damage level corresponding to each component unit.

[0015] S600, based on the component consistency correction amount, adjust the initial damage prediction probability of voxels belonging to the same component unit, generate and output the final three-dimensional voxel damage distribution map that conforms to component-level engineering consistency.

[0016] According to a second aspect of the present invention, a three-dimensional voxel damage prediction device based on component consistency constraints is provided, comprising:

[0017] The preprocessing unit is used to acquire the three-dimensional geometric and physical property information of the structure to be evaluated, construct a three-dimensional voxel mesh, and establish the mapping relationship between component elements and voxels.

[0018] An input building unit is used to construct a multi-channel input tensor, which includes at least a strength channel input tensor reflecting the inherent properties of the structure and a spatial distribution channel input tensor reflecting the influence of external loads.

[0019] The damage prediction unit is used to input the multi-channel input tensor into a pre-trained three-dimensional segmentation network model to obtain the initial damage prediction probability for each voxel. The three-dimensional segmentation network model is optimized by an overall loss function that includes component consistency constraint loss. The component consistency constraint loss applies interval constraints based on a preset damage level to the component-scale aggregation result of the voxel-level prediction through the mapping relationship between component units and voxels.

[0020] A scale aggregation unit is used to aggregate the initial damage prediction probability into a component-level damage characterization quantity for each component unit based on the mapping relationship between the component unit and the voxel.

[0021] The consistency correction unit is used to calculate the component consistency correction amount based on the component-level damage characterization quantity of each component unit and the damage ratio reference range defined by the preset damage level corresponding to each component unit.

[0022] The result generation unit is used to adjust the initial damage prediction probability of voxels belonging to the same component unit according to the component consistency correction amount, and generate and output the final three-dimensional voxel damage distribution map that conforms to the component-level engineering consistency.

[0023] The present invention has at least the following beneficial effects:

[0024] 1. Overcoming extreme class imbalance and improving training stability: By designing a phased weight scheduling strategy, classification loss is suppressed in the early stage of training, and the optimization direction is stabilized by relying on the overlap loss which is insensitive to the class. This effectively avoids the model from falling into trivial solutions where all predictions are of undamaged values ​​due to the extreme sparsity of damaged voxels, and ensures reliable convergence of the training process.

[0025] 2. Achieving cross-scale prediction consistency and enhancing engineering rationality: The system creatively uses "components" as prior knowledge as the basic unit of engineering evaluation. Through a component consistency constraint loss function, the network output is forced to meet the overall requirements at the component scale. This makes the detailed voxel-level prediction results logically consistent with the macroscopic component damage level assessment, eliminating contradictory engineering irrationality and significantly improving the interpretability and direct usability of the prediction results.

[0026] 3. Enhancing the robustness and generalization ability of the model in weakly supervised scenarios: By fusing relatively easily obtainable component-level damage level labels as strong constraint signals and co-supervising them with voxel-level loss, the dependence on high-precision, complete voxel-level labels is reduced. This mechanism enables the model to resist local noise in voxel-level simulations or synthetic labels, guiding the model to learn more generalized overall component damage patterns and enhancing the data adaptability and practicality of the method in engineering practice.

[0027] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A flowchart of a three-dimensional voxel damage prediction method based on component consistency constraints provided in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] 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 invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0032] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0033] This invention provides a three-dimensional voxel damage prediction method based on component consistency constraints, which addresses the following technical problems existing in the application of existing three-dimensional voxel segmentation technology for damage prediction in engineering structures:

[0034] 1. The severe class imbalance caused by the extremely sparse voxel-level damage region makes it easy for the model to fall into trivial solutions of "all predictions are as no damage" in the early stage of training, resulting in unstable convergence or even failure.

[0035] 2. The prediction results lack engineering semantic constraints. Existing methods only optimize at the voxel level, ignoring the holistic requirement of components as basic units in engineering evaluation, which often leads to unreasonable contradictions in the prediction results at the component scale.

[0036] 3. Under weak supervision or synthetic data conditions, the model is sensitive to label noise and is prone to overfitting to local uncertainties and random errors in voxel-level labels, making it difficult to guarantee the robustness and rationality of the overall damage state prediction of components.

[0037] The core of the three-dimensional voxel damage prediction method based on component consistency constraints provided in this invention lies in the application of a three-dimensional segmentation network model trained using a specific paradigm. This training paradigm is key to solving the engineering problems described in the background section.

[0038] The training paradigm of the described 3D segmentation network model innovatively embeds the component-level evaluation logic from the engineering field into the deep learning optimization process in the form of a differentiable loss function. Its core idea is to guide model learning through an overall loss function that incorporates cross-scale constraints. This loss function not only includes a voxel-level loss function to supervise voxel-level prediction accuracy, but also innovatively introduces a component consistency constraint loss that forces the prediction results to conform to engineering rationality at the component scale.

[0039] Through this training, the model not only learns the complex mapping from structural properties and load information to local damage, but also internalizes the engineering prior knowledge of the component as a whole evaluation unit. Therefore, the model possesses a cross-scale reasoning capability: its directly output voxel-level predictions are implicitly constrained by the overall state of the component, laying a good foundation for subsequent explicit component-level consistency correction.

[0040] It should be noted that this invention does not limit the specific internal architecture of the 3D segmentation network model. Its core protection lies in the methodological idea of ​​optimizing the voxel-level prediction output of the model through the component consistency constraint loss. Therefore, any 3D deep learning model capable of outputting voxel-level damage predictions and optimized end-to-end through the component consistency constraint loss is within the scope of this invention.

[0041] To facilitate understanding, the following are several illustrative network architecture examples that can realize this concept:

[0042] Example 1: Single Segmentation Network Architecture

[0043] This architecture consists of a standard encoder-decoder segmentation network (such as 3D U-Net) that directly outputs voxel damage prediction logistic values. Component consistency constraints are implemented entirely through an external loss function; that is, during training, the component consistency constraint loss is calculated based on the network's output and backpropagated along with the voxel-level loss, implicitly guiding the network to learn feature representations that conform to component consistency. This architecture is the simplest and most efficient.

[0044] Example 2: Cascaded Two-Stage Architecture

[0045] This architecture divides the prediction process into two differentiable stages:

[0046] Damage segmentation network: As the first stage, it receives multi-channel input tensors and outputs an initial voxel damage probability map.

[0047] Consistency Optimization Module: As the second stage, it receives the probability map and component semantic mask output from the first stage. This module (e.g., a lightweight fully connected network or a differentiable optimization layer) aggregates component statistics based on component mapping relationships and generates a global correction signal (such as scaling or bias parameters for each component) based on component consistency constraints. It then adaptively adjusts the initial probability map and outputs the corrected final probability map.

[0048] During training, the architecture allows for joint training of the two phases, with the overall loss simultaneously monitoring the quality of the final output and the output of the first phase, thus making the cross-scale consistency logic explicit.

[0049] Example 3: Feature Modulation Architecture

[0050] This architecture introduces a component-aware feature modulation module on top of a standard segmentation network. This module takes a component semantic mask or a component state predicted by an auxiliary sub-network as conditional input and performs spatially adaptive modulation (e.g., through conditional normalization layers or attention mechanisms) on the feature maps during the main segmentation network's decoding process. Its core idea is to enable the network to integrate component-level overall contextual information when reconstructing details, thereby guiding the generation of scale-consistent predictions from the feature learning stage.

[0051] Example 4: Multi-task collaborative architecture (a preferred implementation).

[0052] This architecture explicitly adds component-level semantic understanding capabilities to a single segmentation network, achieving deeper cross-scale feature fusion through multi-task learning. It comprises two branches that work together:

[0053] Main branch (segmentation branch): As the core, it adopts an encoder-decoder structure (e.g., 3D U-Net) and is responsible for extracting multi-level spatial and semantic features from the multi-channel input tensor. Finally, it outputs the damage prediction logic value of each voxel at the end of the decoder.

[0054] Auxiliary Branch (Component Rating Branch): This branch, acting as a parallel task module, connects to a deeper layer of the main encoder, receiving its semantically rich intermediate layer feature maps. Its core operation involves aggregating the responses of all voxels belonging to the same component on the feature map using the mapping relationship between the component units and voxels (e.g., a component semantic mask), employing global average pooling to generate a fixed-dimensional component-level feature vector. This vector is then passed through a lightweight prediction head (e.g., a fully connected layer) to directly output the predicted damage level for each component unit.

[0055] The training advantage of this architecture lies in the fact that the component rating task of the auxiliary branch provides a direct component-level supervision signal through an additional auxiliary loss term (such as cross-entropy loss). This signal works synergistically with the voxel-level loss of the main branch and the component consistency constraint loss acting on the output of the main branch, jointly guiding the features learned by the network to facilitate both fine voxel segmentation and macroscopic component evaluation, thereby strengthening the consistency of cross-scale predictions from the internal mechanism of the model.

[0056] Regardless of the architecture used, the specific training process follows the unified paradigm proposed in this invention, mainly including the following steps:

[0057] S1. Obtain the training dataset, which includes a three-dimensional voxel mesh of the sample structure, the mapping relationship between component units and voxels, multi-channel input tensors, voxel-level damage ground truth labels, and preset damage level labels of component units.

[0058] Each sample in the training dataset corresponds to an independent engineering structure evaluation scenario and contains the following interrelated data elements:

[0059] 1. Three-dimensional voxel mesh of sample structure: constructed based on the three-dimensional geometric and physical property information of the sample structure.

[0060] In the context of this invention, three-dimensional geometric information refers to a set of data used to uniquely determine the spatial shape, position, and dimensions of each component in an engineering structure, and is the foundation for constructing the mapping relationship between voxelized models and components. Its specific forms and sources include, but are not limited to:

[0061] Parametric geometric models: For example, data generated by computer-aided design (CAD) software, in which components are defined by type (such as beam, column) and key parameters (such as length, cross-sectional dimensions, spatial coordinates).

[0062] Building Information Modeling (BIM) data: A digital building model containing complete geometric information and non-geometric attributes (such as materials and strength).

[0063] Three-dimensional surface mesh models: For example, surface mesh data composed of triangular facets (such as .stl, .obj format files) can be obtained through three-dimensional laser scanning, photogrammetry, or exported from simulation software.

[0064] Constructive solid geometry (CSG) representation or existing voxelized representation itself.

[0065] In the context of this invention, physical property information refers to engineering parameters related to the mechanical properties of structural materials or the damage resistance of component sections. It does not determine the spatial shape and location of a component, but rather its inherent response characteristics under load. Its specific content includes, but is not limited to:

[0066] Material mechanical properties: such as the elastic modulus, yield strength, ultimate compressive / tensile strength, density, etc.

[0067] Cross-sectional geometric properties, such as cross-sectional area, moment of inertia, and section modulus, directly reflect the resistance of a component to bending moment, shear force, and other forces.

[0068] Other engineering indicators: such as the pre-assessed component vulnerability index and durability level.

[0069] In this invention, the three-dimensional voxel mesh is obtained by discretizing the continuous space of the engineering structure, which is a key step in transforming continuous geometric and attribute information in the real world into a computer-processable, regularized digital model. Its specific construction process is a standardized data processing pipeline, mainly including the following steps:

[0070] S11, Input and Parsing: Input the three-dimensional geometric information of the structure (such as a BIM model, CAD file, or triangular mesh) and the corresponding physical attribute information (usually provided in the form of a material list, component attribute list, etc., associated with the geometric model or provided separately). Parse the input data to extract geometric topological relationships and attribute mappings.

[0071] S12, Spatial Discretization and Mesh Frame Generation: Based on the preset voxel spatial resolution (i.e., the physical size represented by each voxel, denoted as s, usually in meters), a regular 3D axis-aligned bounding box that can completely enclose the input structure is automatically calculated. Then, this bounding box is uniformly subdivided in three dimensions (corresponding to depth D, height H, and width W, respectively) according to the preset voxel spatial resolution s, thereby dynamically determining the mesh dimensions D×H×W. An initially empty 3D voxel mesh frame is generated, consisting of D×H×W regular cubic units. Each voxel is uniquely identified by its 3D integer index (d, h, w), where:

[0072] d is the index of the depth direction (usually corresponding to the X-axis or front-back direction in three-dimensional space), and its value range is [0, D-1].

[0073] h is the index of the height direction (usually corresponding to the Y-axis or vertical direction in three-dimensional space), and its value range is [0,H-1].

[0074] w is the index of the width direction (usually corresponding to the Z-axis or left and right directions in three-dimensional space), and its value range is [0, W-1].

[0075] The index triple (d, h, w) corresponds to an index in real space consisting of [x min +w×s,x min +(w+1)×s]×[y min +h×s,y min +(h+1)×s]×[z min +d×s,z min The regular cubic region defined by [+(d+1)×s], where (x min ,y min ,z min ) represents the coordinates of the smallest corner point of the bounding box.

[0076] S13, Voxel Attribute Assignment and Relationship Establishment: This step is the core of voxelization, aiming to fill each voxel in the mesh framework with semantic and attribute information. The specific operations are as follows:

[0077] (1) Component Attribution Determination: By traversing each voxel element in the voxel mesh and calculating the spatial positional relationship between its geometric center (or the entire region) and the geometric models of each component, it is determined which component element it belongs to. This invention can employ various efficient algorithms to achieve this determination, including but not limited to:

[0078] Precise geometric determination method: Directly calculate the geometric shape of the component (defined by the component's vertex coordinates, surface triangular mesh, or parametric geometric description) and the spatial positional relationship of each voxel element. Determine whether the voxel belongs to the component through geometric intersection test or inclusion judgment.

[0079] Bounding box filling method: Calculate the axial bounding box (AABB) of each component in the voxel mesh coordinate system, and then use a voxelized filling algorithm (such as scanline filling) to quickly determine all voxels belonging to the component within the 3D space defined by the bounding box. This method is more efficient and suitable for scenarios with regular component shapes.

[0080] (2) Establish mapping relationship and record attributes: Based on the above judgment results:

[0081] Establish mapping relationships: Record the correspondence between each voxel and its associated component, forming a mapping relationship between component units and voxels. This relationship is the basis for all subsequent cross-scale operations.

[0082] Assign geometrically derived properties: You can record properties derived from the original geometric information, such as the center coordinates, for voxels.

[0083] Associating physical properties: Based on the component to which the voxel belongs, query the physical property information obtained in step S11, and store the corresponding material strength, section modulus and other parameters as the basic physical property values ​​of the voxel.

[0084] Through the above process, a complete and structured 3D voxel mesh data is finally generated, integrating spatial discrete framework (geometry), component semantic mapping (attribution), and basic physical properties (material). This data volume serves as a unified spatial benchmark and attribute data source for subsequent steps, including multi-channel input tensor construction (intensity channel values ​​are directly derived from the physical properties of voxels), network forward prediction, and component-scale statistical aggregation, ensuring the consistency of spatial benchmark and attribute information throughout the entire process from data preprocessing to model prediction.

[0085] 2. Mapping Relationship between Component Units and Voxels: This relationship explicitly records the component unit to which each voxel in the 3D voxel mesh belongs, forming the key indexing basis for all subsequent cross-scale calculations (such as feature aggregation and statistical calculations). In a preferred embodiment of the present invention, this mapping relationship is instantiated and stored using a component semantic mask as a specific data structure. The component semantic mask is a 3D integer tensor with the same dimensions (D, H, W) that is fully aligned with the 3D voxel mesh in spatial dimensions. In this mask tensor, each voxel position stores a unique integer identifier, which indicates which specific component unit the voxel belongs to. For example, all voxels belonging to "beam 1" have a mask value set to 1; voxels belonging to "column 2" have a mask value set to 2; and background voxels that do not belong to any component are assigned a specific reserved value, such as 0. This representation transforms the abstract attribution mapping relationship into a dense, regular, and efficiently accessible structured data that can be directly indexed from memory. This mask originates from the geometric processing flow of this step and serves as the core data foundation for achieving cross-scale information aggregation and consistency constraints in subsequent network training and prediction.

[0086] 3. Multichannel Input Tensor: Composed of at least two spatially aligned, independent data channels stacked together, providing a fused input that integrates structural characteristics and load environment for subsequent 3D segmentation networks. The specific construction method of the multichannel input tensor is as follows:

[0087] (1) Generation of the intensity channel input tensor

[0088] This strength channel aims to provide the network with prior knowledge of its structural resistance. The strength eigenvalues ​​of each voxel are obtained by querying and mapping physical property information. The specific implementation process includes the following steps:

[0089] Step a1, determine the component to which the voxel belongs: For each voxel in the 3D voxel mesh, based on the established component element-voxel mapping relationship, query and determine one or more component elements to which the voxel belongs.

[0090] Step a2, query component attribute values: Based on the component affiliation determined in step a1, obtain the values ​​of the corresponding component's material mechanical properties (such as yield strength, ultimate compressive strength, or elastic modulus) or cross-sectional geometric properties (such as section modulus) from the physical attribute information.

[0091] Step a3, Calculate and assign intensity feature values: Normalize the attribute values ​​obtained in step a2 (e.g., divide by the maximum value of all components in the training dataset for that attribute), and use the processed result as the intensity feature value of the voxel. For different cases, follow these rules:

[0092] If a voxel belongs uniquely to a component, then the attribute value of that component is used directly.

[0093] If a voxel is located at the geometric boundary of multiple components (i.e., the mapping relationship indicates that it belongs to multiple components), then the arithmetic mean of the attribute values ​​of these components is taken.

[0094] If the voxel does not belong to any component (such as the background or cavity region), a preset background value (such as 0) is assigned.

[0095] Step a4, Generate Intensity Channel Tensor: Traverse all voxels in the entire 3D voxel mesh, repeating steps a1 to a3, assigning intensity feature values ​​to each voxel, and finally generating the intensity channel input tensor I∈R. D×H×W This tensor spatially characterizes the inherent material or cross-sectional capacity threshold at each point of the structure to resist damage.

[0096] (2) Generation of spatially distributed channel input tensors

[0097] This spatially distributed channel is designed to encode the physical field information of specific external load events. The load intensity value for each voxel is obtained through physical calculations based on load parameters. The specific implementation process includes the following steps:

[0098] Step b1, determine the load parameters: Based on the evaluation scenario, determine the type of external load (e.g., point source explosion impact, local impact, etc.) and its key physical parameters. Taking a point source explosion as an example, the parameters include at least: the coordinates of the explosion center in three-dimensional space (x0, y0, z0), the explosion equivalent E (usually expressed as TNT equivalent), and the attenuation characteristic parameters of the shock wave propagating in the medium (e.g., attenuation coefficient k).

[0099] Step b2, Define the load attenuation model: Based on the load type, select or establish a physical model or empirical formula describing the attenuation of load intensity with spatial distance. For example, for a free-field air-based explosion shock wave, the attenuation relationship of the peak overpressure POP of the free-field air-based explosion shock wave with distance r can be expressed as POP(r)∝E 1 / 3 / r or the more precise Friedlander equation, etc. In this model, the independent variable *r* represents the Euclidean distance from any point in space to the load source (e.g., the explosion center). ∝ is a proportional sign.

[0100] Step b3, calculate the voxel load strength value: traverse each voxel in the 3D voxel mesh. For the current voxel v, first, calculate the geometric center point (x) of that voxel. v y v , z v The Euclidean distance r from the load source (e.g., the center of the explosion) v Subsequently, r vSubstituting into the above load attenuation model, the theoretical load intensity POP(r) at this voxel location is calculated. v This POP(r) v This is initially used as the load strength value of the voxel.

[0101] Step b4, Normalization and Tensor Generation: After calculating all voxels, a raw three-dimensional array of load intensity distribution is obtained. To optimize network training stability, this array is usually normalized (e.g., linearly scaling all values ​​to the [0, 1] interval, or applying standard scores). The normalized three-dimensional array is the final spatial distribution channel input tensor L∈R. D×H×W The spatially distributed channel input tensor L physically characterizes the input intensity distribution field of the external excitation targeted in this evaluation throughout the entire structural space.

[0102] (3) Integration of multi-channel tensors

[0103] Stack the generated I and L channel tensors along the channel dimension to construct the final multi-channel input tensor X∈R. C×D×H×W Where C≥2. Each channel is strictly aligned in the spatial dimension to ensure that the value of each voxel position in different channels corresponds to different properties (intrinsic strength and external load) of the same spatial point, thereby providing the fusion information necessary for the 3D segmentation network to learn the damage generation mechanism (i.e., the interaction between properties and excitation).

[0104] 4. Voxel-level damage ground truth labels: A three-dimensional data array that is spatially strictly aligned with a three-dimensional voxel mesh. It is used to provide the network with accurate voxel-scale supervision signals and is crucial for training the model to learn damage localization and morphology. It is usually a binary array (each voxel value is 0 to represent "no damage" and 1 to represent "damage"), or it can be a probability array (representing the uncertainty of damage existence).

[0105] During training, the label is compared point by point with the voxel predictions output by the network. The error is calculated using voxel-level loss functions (such as Dice Loss and Focal Loss) to drive the optimization of network parameters, thereby teaching the model how to identify the precise spatial distribution of damage from the input features.

[0106] The acquisition of this label reflects the idea of ​​multi-source data fusion, and its sources include:

[0107] High-precision numerical simulation: Through mechanical simulations such as finite element analysis (FEA) and discrete element method, damage processes can be simulated, generating a voxel-level damage field that is comprehensive and has a clear physical mechanism.

[0108] Experimental observation data: By using non-destructive testing technologies such as industrial computed tomography (CT) and three-dimensional optical scanning to scan and reconstruct real damaged components, highly realistic damage data can be obtained.

[0109] Expert manual annotation: The damaged area is manually drawn and annotated on the 3D model or slice image by domain experts. It is suitable for small-scale, high-value or complex damage patterns of samples.

[0110] While one of the core innovations of this invention lies in using component-level labels to constrain and improve the rationality of voxel predictions, high-precision voxel-level ground truth labels (when available) remain a valuable source of supervision for guiding the model to learn basic damage features and ensuring spatial accuracy of predictions in the early stages of training. In weakly supervised scenarios, these labels may be incomplete or noisy; in such cases, the framework proposed in this invention can compensate for this lack of information through component consistency constraints.

[0111] 5. Preset Damage Level Labels for Component Units: A discrete level label corresponding one-to-one with each component unit in the structure, used to provide monitoring signals for component dimensions. This label typically uses a finite number of discrete levels to represent the overall damage state of the component, such as {no damage, minor damage, moderate damage, severe damage}. Each level corresponds to a qualitative or semi-quantitative evaluation of the overall damage degree of that type of component in engineering practice.

[0112] During training, this label does not directly supervise voxels, but is used for two purposes:

[0113] Calculate the component consistency constraint loss: As a benchmark, compare it with the component scale statistics obtained by network prediction aggregation to calculate whether it conforms to the engineering rationality.

[0114] (If present) Supervised auxiliary branch: In a multi-task network architecture, the learning of the component rating branch is directly supervised to enhance the component-level semantics of feature representation.

[0115] This tag is usually more aligned with engineering practice and mainly comes from:

[0116] Simulation Summary: The damage level of a component is comprehensively evaluated based on the overall response (such as maximum displacement, plastic strain distribution, and bearing capacity reduction factor) from the post-processing results of high-fidelity numerical simulation.

[0117] Inspection report: Based on on-site inspection and monitoring data of real projects or post-disaster assessment reports, experts make a grade judgment on the overall condition of the components.

[0118] Engineering standards: The assessment is conducted in accordance with the damage level classification criteria defined in relevant industry specifications and standards.

[0119] Each preset damage level L is associated with a reasonable damage ratio reference range [α] in this invention. L ,β L The interval defines a component belonging to level L, whose internal damage voxels should generally fall within this range. This association is the mathematical basis for transforming qualitative levels into computable constraints (for component consistency constraint loss) and a key bridge connecting engineering experience and data-driven models. The interval boundary can be preset based on engineering experience, obtained statistically from training data, or used as a learnable parameter.

[0120] In this invention, the lower limit α of the damage ratio reference interval is... L and upper limit value β L It can be set in any of the following ways:

[0121] (1) Preset based on engineering experience: Differentiated settings based on component type (beam, column, etc.), material properties or functional importance.

[0122] (2) Based on data statistics: In the training dataset, the distribution of the true damage ratio (i.e. the proportion of damaged voxels in the truth mask) of all components marked as level L is statistically analyzed, and a certain confidence interval (such as 10%-90% quantile) is taken as the damage ratio reference interval.

[0123] (3) As learnable parameters: the lower limit and upper limit are initialized to reasonable values ​​and optimized together with other network parameters through gradient descent during training, so that the model can adaptively learn the optimal range.

[0124] The above five elements together constitute a complete training sample, providing comprehensive information for the training of the three-dimensional segmentation network model, from input and prior knowledge to multi-granular supervision.

[0125] S2 inputs the multi-channel input tensor of the training dataset into the 3D segmentation network model to be trained, performs forward computation to obtain the network's prediction output, and obtains the corresponding prediction result.

[0126] The prediction result refers to the output generated after the forward computation of the 3D segmentation network model. In one embodiment, when the network is a multi-task architecture, the prediction result includes at least the damage prediction logic value of each voxel output by the main branch, and optionally includes the damage level prediction value of each component unit output by the auxiliary branch. The damage prediction logic value is used to calculate the voxel-level loss, and the damage level prediction value is used to calculate the auxiliary loss.

[0127] In an embodiment employing the multi-task network architecture, the forward computation process is completed jointly by two cooperating branches:

[0128] (1) Main branch (segmentation branch): As the core feature extraction and reconstruction path, it encodes and decodes the input multi-channel tensor, and finally outputs the damage prediction logic value of each voxel. This logic value is then transformed by an activation function (such as Sigmoid) to represent the independent damage probability of each voxel.

[0129] (2) Auxiliary Branch (Component Rating Branch): This branch connects to the intermediate layer of the main branch to obtain intermediate layer feature maps containing rich semantics. When performing component-level prediction, this branch needs to access the component unit and voxel mapping relationship pre-bound to the current training sample (usually stored and loaded into memory in the form of component semantic masks or index tables). Using this mapping relationship, the auxiliary branch aggregates the intermediate features of all voxels belonging to the same component (e.g., through pooling operations) to form a component-level feature vector, and then calculates the damage level prediction value of each component unit through subsequent network layers (such as fully connected layers).

[0130] It should be noted that the 3D segmentation network model follows different output paradigms during the training and inference phases: during training, intermediate logical values ​​and auxiliary information are output to optimize parameters; during inference, the converted voxel damage probabilities are directly output to provide the final result. Essentially, these are the same model operating in different stages.

[0131] S3, Based on the voxel-level damage prediction information in the prediction results and the voxel-level damage ground truth labels Y∈{0,1} in the training dataset. D×H×W Calculate the voxel-level loss function.

[0132] In the training process of this invention, the voxel-level damage prediction information refers to the numerical data obtained from the main branch output of the 3D segmentation network model, used to characterize the damage state of each voxel. In a specific embodiment, this information is the damage prediction logical value Z∈R output by the main branch of the network. D×H×W .

[0133] Furthermore, the specific calculation process of the voxel-level loss function is as follows:

[0134] S31, the damage prediction logic value Z in the prediction result is converted into a voxel damage probability map P=σ(Z) through the Sigmoid activation function σ. P and Z have the same dimension, and each element p∈[0,1] in P represents the probability that the corresponding voxel is predicted as damaged.

[0135] S32, calculate the loss term based on the overlap between the predicted region and the actual damage region (hereinafter referred to as the overlap loss term, denoted as L) based on the voxel damage probability map P and the voxel-level damage ground truth label Y. overlap ) and voxel-level classification loss term L focal .

[0136] Among them, the overlap loss term measures the degree of overlap between the predicted region and the real region, and is relatively robust to class imbalance. Taking Dice loss as an example, L... overlap =1-[(2×∑ i (p) i ×y i )+ε) / (∑ i p i +∑ i y i +ε)], where p i Let y represent the predicted probability of the i-th voxel in P. i p in Y i Binary label at the corresponding position (y i (∈{0,1}, where 1 represents damage and 0 represents no damage). ∑ i (p) i ×y i The expression represents the sum of the predicted probabilities of all Q=D×H×W voxels in a 3D voxel mesh and their true labels, characterizing the intersection of the predictions and the true values ​​(only damaged voxels y). i p at =1 i It will participate in effective calculations and will not damage voxels. i =0 (contribution is 0 at point 0); ∑ i p i ∑ represents the sum of predicted probabilities over all voxels, characterizing the overall size of the predicted damage area; i y i This represents the summation of the true labels of all voxels in the 3D voxel mesh, characterizing the overall size of the true damage region (due to y). i This is a binary label, and its value is actually equal to the number of real damaged voxels. The value of i ranges from 1 to Q. ε is a small positive constant (e.g., 1 × 10⁻⁶). -6 (), used to prevent calculation errors where the denominator is zero.

[0137] The voxel-level classification loss term is designed to address the voxel class imbalance problem in voxel damage prediction scenarios. It weakens the loss weights of easily classified samples and strengthens the loss weights of difficult-to-classify samples (such as damage voxels with low probability of prediction) through a modulation factor. Taking Focal Loss as an example, its calculation formula is: L focal =-[∑ i y i ×(1-p i ) γ ×logp i +(1-y i )×p i γ ×log(1-p) i ], where γ≥0 is the focusing parameter, used to adjust the degree of attention to easy and difficult samples (usually γ=2).

[0138] S33, based on overlap loss term L overlap And voxel-level classification loss term L focal Calculate the voxel-level loss function.

[0139] In this invention, the voxel-level loss function L voxel The overlap loss term L overlap And voxel-level classification loss term L focal It is composed of weighted combinations. Its expression is: L voxel =L overlap +λ(t)·L focal λ(t) is a weight scheduling function that varies with the training period or the number of iterations t.

[0140] To address the issue that the model's predictions of trivial solutions are easily caused by the extreme sparsity of damaged voxels in the early stages of training, this step employs a phased weight scheduling strategy to dynamically adjust the weights of the voxel-level classification loss term, i.e., λ(t):

[0141] Phase 1 (Stable Initialization): Within a pre-defined phase after training begins (e.g., the first K% of the total training epochs, where K is preferably 10%; if the proportion of damaged voxels in the structure to be evaluated is less than 1%, K is 15%~20%; if the proportion of damaged voxels is greater than 5%, K is 5%~10%), λ(t) is set to 0 or a very small value (e.g., λ(t) < 0.1). For example, if the total number of training epochs is 1000, the pre-defined stable initialization phase is the first 100 epochs (K=10%). In this phase, λ(t) is set to 0.01 (a very small value), and model optimization relies only on the class-insensitive overlap loss term to stabilize the three-dimensional spatial features of the learned structure and avoid being misled by the proportion of undamaged voxels.

[0142] The second stage (refined classification): After the preset stage, the weight λ(t) is started from 0 and monotonically increased to a target value (e.g., 1) according to a preset scheduling function. In this stage, supervision of the classification loss is gradually introduced to enhance the model's ability to distinguish the minority class (damage voxels).

[0143] The implementation methods of the weighted scheduling function include, but are not limited to:

[0144] (1) Pre-defined monotonically increasing functions, such as linear function λ(t)=t / T, square function λ(t)=(t / T) 2 The cosine function λ(t) = 0.5 × (1 - cos(πt / T)) or the exponential function, etc., where T is the total planned duration (number of cycles or steps) of the second stage.

[0145] (2) A piecewise function that is dynamically adjusted based on the validation set evaluation metrics (such as loss value, segmentation accuracy), for example, when the validation set metrics tend to stabilize before it starts to increase.

[0146] In one specific embodiment, a normalized training progress t can be used. norm =min(epoch / (epoch max -1),1), and define λ(t) = (t norm ) 2 In this context, epoch represents the current training cycle number, i.e., which training cycle (round) is currently being executed. It is a positive integer starting from 0 or 1. It is the basic unit of progress in the training process. max This indicates the preset total number of training epochs. Training will continue until all epochs are completed. max It ends after one cycle.

[0147] In this invention, (t) norm ) 2 The design allows for training in the early stages (t norm When the weights approach zero, the weights increase very slowly, thus maintaining a weakly supervised state for the classification loss term for a longer period, fully ensuring the stable initialization of the model based on the overlap loss term under extreme class imbalance. Entering the later stages of training (t... norm When the size is large, the weights increase rapidly to enhance the ability to distinguish difficult-to-differentiate samples (damaged voxels). The entire scheduling process is smooth, differentiable, and requires no additional hyperparameter tuning, making it a preferred implementation scheme that balances stability and performance.

[0148] It should be noted that the voxel-level damage prediction information can also be a converted voxel damage probability map P or other equivalent forms. As long as it can be used to calculate the voxel-level loss function with the truth label Y, it is within the scope of this step.

[0149] S4. Based on the mapping relationship between component units and voxels, the voxel-level damage prediction information is aggregated into the component-level damage characterization quantity of each component unit, and the component consistency constraint loss is calculated according to the damage ratio reference interval corresponding to the preset damage level label of the component unit.

[0150] In this step, based on the voxel damage probability map P obtained in S3, the mapping relationship between component units and voxels provided in step S1 (preferably provided in the form of a component semantic mask M), and the preset damage level label L(c) of each component unit, the component consistency constraint loss L is calculated. comp The purpose of this loss is to inject the prior knowledge of the overall damage level assessment of components in the engineering process as a constraint signal into the network training, so as to ensure that the network output conforms to the engineering rationality at the component scale.

[0151] Furthermore, the specific calculation process for the component consistency constraint loss includes the following steps:

[0152] (1) Aggregation of component-level damage representation parameters: For each component unit c, all voxels satisfying M[j]=c are selected from the voxel damage probability map P using the component semantic mask M, resulting in the set of probability values ​​{p} of all voxels satisfying M[j]=c. j} j∈Vc Where M[j] represents the integer identifier stored at the corresponding position in the component semantic mask M for the j-th voxel in the voxel damage probability map P, and p j Let Vj represent the damage prediction probability of the j-th voxel, and Vc represent the set of voxel indices belonging to component unit c, where c ranges from 1 to n, and n is the total number of component units. Then, these probability values ​​are aggregated into a component-level damage characterization Sj for component unit c. c This is used to characterize the overall predicted damage level of the component. The specific aggregation method can be:

[0153] Mean: S c =(1 / |Vc|)×∑ j∈Vc (p) j ), |Vc| represents the number of elements in Vc, that is, the total number of voxels contained in the component unit c. In other words, S c It is the arithmetic mean of the predicted probabilities of all voxels within the component.

[0154] Weighted mean: S c =(1 / ∑ j∈Vc (w) j ))×∑ j∈Vc (p) j ×w j ), where w j The weighting coefficient for the j-th voxel can be determined based on factors such as the material properties of the voxel, its distance from the load source, or a pre-calculated vulnerability index, to reflect the differences in damage contribution or structural importance of voxels at different locations within the component.

[0155] Quantiles: For example, the median or a specified quantile (such as the 75th quantile) of the predicted probabilities of voxels within a component is taken as S. c It is more robust to outliers.

[0156] Engineering Explanation: S c This can be intuitively explained as the proportion of voxels inside a component that are predicted to be damaged. It is a key bridge for collapsing fine voxel-level predictions into macroscopic component-level assessments.

[0157] (2) Calculation of component consistency constraint loss: For each component unit c, based on its preset damage level label L(c) (such as mild, moderate, severe damage, etc.), a reasonable damage ratio reference range [α] is set. L（c） ,β L（c） This interval defines the reasonable range within which the overall damage proportion of the component should fall under this damage level (i.e., the component-level damage characterization quantity S). c (A reasonable range of values). Component consistency constraint loss L comp The component-level damage characterization S is obtained by penalizing component element c. c Calculated by the degree of deviation from this range:

[0158] L comp =∑ n c=1 (max(0, α) L（c） -S c )+max(0,S c -β L（c） )).

[0159] Where, α L（c） and β L（c） Let α be the lower and upper limits of the damage proportion reference interval corresponding to L(c), respectively, and satisfy 0≤α L（c） ≤β L（c） ≤1.

[0160] When S c When the damage ratio falls within the reference range, the calculation results of both penalty terms are 0, and the consistency constraint loss term corresponding to component element c is 0; when S c Below the lower limit α L（c） or higher than the upper limit β L（c） When this occurs, the corresponding penalty term will produce a positive penalty value, and the component element will contribute a non-zero consistency constraint loss term. Component consistency constraint loss L comp This is the sum of the consistency constraint loss terms corresponding to all component units.

[0161] To facilitate efficient training on large-scale 3D meshes, the aggregation operation can be optimized in the following ways:

[0162] (a) The mapping relationship of the semantic mask representation of the components is preprocessed into a list of voxel indices or a sparse representation corresponding to each component.

[0163] The purpose of this approach is to transform spatial query-based aggregation logic into efficient data operations based on direct indexes, thereby fully utilizing the parallel computing capabilities of modern GPUs.

[0164] The input to this method is a component semantic mask M∈Z. D×H×WEach voxel location stores the ID (integer identifier) ​​of its associated component unit, while the background region is typically represented by specific values ​​such as 0 or -1. Directly traversing the mask M in each training iteration to find all voxels belonging to a given component incurs enormous computational overhead. Therefore, this step aims to pre-parse and organize all component-voxel relationships, transforming the spatial scanning-based query process into a high-speed data access operation via direct memory indexing.

[0165] Preprocessing methods mainly employ one of the following two forms for conversion:

[0166] Form 1: Generate a list of component voxel indices.

[0167] This method generates a unique voxel location list for each component unit. The specific process is as follows: First, the 3D component semantic mask M is flattened into a one-dimensional array. Then, this one-dimensional array is traversed, and based on the component ID stored at each location, the linear index of that location is added to the corresponding component's index list. Finally, a data structure (such as a list of lists or a dictionary) is obtained, with component IDs as keys and the linear index array as values. In subsequent training, when calculating the statistics of a certain component, all relevant data can be efficiently aggregated directly from the similarly flattened voxel prediction probability array using its index list.

[0168] Form 2, constructing a sparse aggregation matrix:

[0169] This method, from a linear algebraic perspective, abstracts the aggregation operation from voxel probabilities to component statistics into a matrix multiplication. Specifically, it constructs a sparse matrix A with the number of rows equal to the total number of component elements n and the number of columns equal to the total number of voxels Q. This matrix is ​​designed such that the c-th row of the sparse matrix A intersects with the flattened voxel probability vector p. flat The result of multiplication is exactly the component-level damage characterization quantity S of component element c. c (For example, when the position corresponding to the voxel belonging to component element c in this row of the matrix is ​​1 / |Vc|, and the rest are 0, the averaging operation is achieved.) This matrix is ​​extremely sparse and can be stored using formats such as Compressed Sparse Rows (CSR).

[0170] After the above preprocessing, in the core loop of training:

[0171] Eliminates real-time parsing overhead: There is no need to repeatedly traverse the 3D mask for each loss calculation.

[0172] It achieves extreme parallelism: whether it is clustering through an index list or performing sparse matrix multiplication, it can rely on the parallel computing architecture of the GPU to complete the statistical calculation of all components at once, which is an order of magnitude improvement in efficiency compared to the traditional component-by-component loop.

[0173] Memory access has been optimized: the data access pattern has been changed from irregular spatial jumps to continuous index access or regular matrix operations, which is more in line with the efficient memory access pattern of modern hardware.

[0174] (b) Utilize the tensor indexing and aggregation operations (such as torch.scatter or tf.segment_sum) provided by deep learning frameworks (such as PyTorch and TensorFlow) to implement voxel probabilistic batch parallel aggregation of all components on the GPU, thereby avoiding inefficient per-component looping.

[0175] The component consistency constraint loss and voxel-level loss calculated in this step jointly guide network optimization, achieving the goal of accurately capturing local damage details while ensuring consistency between macroscopic component-scale prediction results and engineering evaluation standards.

[0176] S5, the voxel-level loss function and the component consistency constraint loss are weighted and combined to form the overall loss function, and the parameters of the three-dimensional segmentation network are updated through the backpropagation algorithm using the overall loss function to complete the model training.

[0177] In one embodiment of the present invention, the overall loss function L total The expression is: L total =L voxel +λ comp ×L comp Wherein, λ comp A pre-defined positive hyperparameter is used to balance the relative importance between voxel-level prediction accuracy and component scale consistency constraints. Through this loss function, the optimization process of the network parameters is simultaneously subject to two layers of constraints: L voxel To ensure that the prediction matches the local truth as closely as possible at the voxel scale, L comp This forces predictions to conform to overall engineering semantics at the component scale.

[0178] In an embodiment employing a multi-task network architecture that includes auxiliary branches, the overall loss function is further expanded to:

[0179] L total =L voxel +λ comp ×L comp +λ aux ×L aux .

[0180] Among them, L aux To compensate for the loss between the predicted component damage level output by the auxiliary branch and the preset damage level label (such as cross-entropy loss), λ aux For L auxThe corresponding balancing weights. Adding this loss aims to improve the consistency of cross-scale predictions by making the features extracted from the main branch more conducive to component-level overall evaluation through multi-task collaborative learning, thereby further enhancing the consistency of cross-scale predictions from within the network.

[0181] After constructing the overall loss function, the gradient of this loss with respect to all trainable parameters of the 3D segmentation network model (including parameters of the main branch and possible auxiliary branches) is calculated using the backpropagation algorithm. Subsequently, gradient descent or its variants (such as Adam and SGD) are used to update the network parameters based on the gradients. This process is repeated iteratively until the model converges or reaches the preset training stopping condition, thus completing the model training.

[0182] By repeatedly executing steps S1 to S5, the network parameters of the 3D segmentation network model are iteratively optimized until the preset training stopping conditions are met (e.g., reaching the maximum number of training cycles or the loss function converges), thereby completing the training of the 3D segmentation network model.

[0183] The method described in this invention is particularly suitable for scenarios where the voxel-level ground truth labels Y are incomplete, scarce, or contain noise. In such weakly supervised or semi-supervised training, due to L... voxel L in focal The term heavily relies on accurate voxel-level labels, which reduces its reliability. In this case, training can primarily rely on the overlap loss term (which is relatively robust to noise) and L as a strong supervision signal. comp (It relies on more readily available component-level labels). This can be achieved by appropriately setting weights (e.g., lowering λ(t) accordingly, or directly setting L). voxel =L overlap The component consistency constraint loss can work in conjunction with the overlap loss to provide effective supervision signals, guiding the network to learn reasonable and logically sound damage distribution patterns even in the absence of perfect voxel labels. This significantly improves the practicality and robustness of the method under real-world engineering data conditions.

[0184] like Figure 1 As shown, based on the above-mentioned trained model, the complete prediction method provided in this embodiment of the invention integrates the solution of the above-mentioned technical problems into a unified framework by introducing component-level engineering priors and cross-scale consistency constraints. Its main application steps include:

[0185] S100: Obtain the three-dimensional geometric and physical property information of the structure to be evaluated, construct a three-dimensional voxel mesh, and establish the mapping relationship between component units and voxels.

[0186] In this invention, the structure to be evaluated refers to an engineering structural entity composed of components with clear engineering semantics (such as beams, columns, slabs, walls, etc.) that requires damage prediction. It corresponds to the sample structure in the training process, but does not require ground truth damage labels.

[0187] The process of constructing the 3D voxel mesh in this step is consistent with the method used in the training phase of this invention, specifically including:

[0188] Spatial discretization: Based on the preset voxel resolution, a regular three-dimensional axis-aligned bounding box that can surround the entire structure is generated and uniformly divided into (D, H, W) cubic units to form the initial mesh framework.

[0189] Geometry and Attribute Voxelization: For each voxel, geometric calculations (such as intersection testing and axial bounding box filling) are used to determine which component element it belongs to, thereby establishing a mapping relationship between component elements and voxels. At the same time, based on the physical property information of the component to which it belongs (such as material strength), basic physical property values ​​are assigned to the voxel.

[0190] In a preferred embodiment, the mapping relationship between component elements and voxels is stored in the form of a component semantic mask. This mask is a three-dimensional integer tensor of the same size as the voxel mesh, and the value at each position identifies the component element ID to which the corresponding voxel belongs.

[0191] S200, construct a multi-channel input tensor, wherein the multi-channel input tensor includes at least a strength channel input tensor reflecting the inherent properties of the structure and a spatial distribution channel input tensor reflecting the influence of external loads.

[0192] In this step, the process of constructing the multi-channel input tensor is consistent with the method used in the training phase of this invention:

[0193] Strength channel input tensor: The value of this channel is directly derived from the basic physical property value (such as the material yield strength) assigned to each voxel in step S100, and is formed after normalization. It spatially encodes the distribution of the structure's inherent ability to resist damage.

[0194] Spatial distribution channel input tensor: The value of this channel is calculated by substituting external load parameters (such as point source location and energy) into a preset physical attenuation model (such as the shock wave overpressure attenuation formula), and it characterizes the relative intensity or energy density distribution of the load in the structural space.

[0195] Aligning and stacking the aforementioned channels in spatial dimensions forms the final multi-channel input tensor X∈R. C ×D×H×W Where channel C≥2, it serves as the unified input for subsequent network predictions.

[0196] S300, the multi-channel input tensor is input into the pre-trained three-dimensional segmentation network model to obtain the initial damage prediction probability of each voxel.

[0197] The multi-channel input tensor constructed in step S200 is input into the pre-trained 3D segmentation network model. The model performs forward computation on the input and directly outputs the initial damage prediction probability of each voxel, forming an initial 3D damage probability map aligned with the input space. This probability map will serve as the input basis for subsequent steps (S400 and S500) to perform component-scale aggregation and engineering consistency correction.

[0198] S400, based on the mapping relationship between the component unit and the voxel, the initial damage prediction probability is aggregated into a component-level damage characterization quantity for each component unit.

[0199] This step, based on the mapping relationship between component units and voxels established in S100 (preferably provided in the form of a component semantic mask), aggregates the initial damage prediction probability map obtained in step S300 into a component-level damage characterization quantity for each component unit.

[0200] In this step, the component semantic mask M serves as a crucial prior index. The initial damage prediction probabilities are filtered through the component semantic mask to aggregate the component-level damage representation for each component element. Specifically, for each component element k of the structure to be evaluated, the unique integer identifier of that component element is used to filter out all voxel position index sets Vk satisfying M[g]=k from the component semantic mask M. Subsequently, based on the voxel position index set Vk, all probability values ​​{p} at the corresponding positions are extracted from the initial damage probability map. g} g∈Vk Where M[g] represents the integer identifier stored at the corresponding position of the g-th voxel in the component semantic mask M in the initial damage probability map, and p g Let Vg represent the damage prediction probability of the g-th voxel, where Vk is the set of voxel position indices belonging to component element k, and k ranges from 1 to B, where B is the total number of component elements in the structure to be evaluated. Finally, a predefined aggregation function F is applied to these probability values. agg (), calculate the component-level damage characterization quantity S of component element k. k :S k =F agg ({p g} g∈Vk The aggregation function can be the mean, weighted average, or quantile, and its specific form has been determined and maintained during the network training phase.

[0201] This step is a crucial bridge connecting the voxel-level fine-grained predictions output by the network with subsequent component-level engineering consistency corrections. Through this aggregation operation, massive amounts of voxel-level probabilistic information are compressed into a series of component-level statistical descriptions with clear engineering semantics, providing direct numerical basis for imposing reasonable constraints and corrections at the component scale.

[0202] When the mapping relationship between the component unit and the voxel is provided in the form of a component semantic mask, in this step, the initial damage prediction probability is filtered through the component semantic mask to aggregate and obtain the component-level damage characterization of each component unit.

[0203] S500, calculate the component consistency correction amount based on the component-level damage characterization quantity of each component unit and the damage ratio reference range defined by the preset damage level corresponding to each component unit.

[0204] This step aims to compare the component-level damage characterization with its reasonable engineering reference range and quantify its deviation, thereby generating a global correction signal. The core design principle of the correction is: if the component-level damage characterization falls within the reasonable range, the correction approaches zero, indicating that the prediction is basically reasonable and no significant adjustment is needed; if the component-level damage characterization deviates from the reasonable range, a non-zero correction is generated, the direction and magnitude of which drive the component-level damage characterization to move closer to the reasonable range.

[0205] In this invention, one possible implementation method for calculating the correction amount is as follows:

[0206] 1. Calculation deviation: For component element k, first calculate its component-level damage characterization quantity S. k Relative to the damage proportion reference range [α] L（k） ,β L（k） The deviation d k .

[0207] If S k <α L（k） , then d k =α L（k） -S k (If it is below the lower limit, it needs to be corrected upwards).

[0208] If S k >β L（k） , then d k =β L（k） -S k (If it exceeds the upper limit, it needs to be adjusted downwards).

[0209] If α L（k） ≤S k ≤β L（k） , then d k =0 (within the interval, no correction is needed).

[0210] 2. Generate component consistency correction amount: △, component consistency correction amount for component element k. k It can be directly taken as the deviation d k Alternatively, it can be processed using a smoothing function g() (such as scaling, sigmoid, etc.) to make the correction process more stable: △ k =g(d k For example, △ k =b×d k b is a scaling factor that controls the intensity of the correction, 0 < b < 1.

[0211] The component consistency correction is a scalar value that acts on the entire component. In the next step (S600), this correction is distributed uniformly or according to a specific rule to each voxel within the component to systematically adjust its initial damage prediction probability, thereby bringing the overall statistics of the corrected component closer to a reasonable range.

[0212] S600, based on the component consistency correction amount, adjust the initial damage prediction probability of voxels belonging to the same component unit, generate and output the final three-dimensional voxel damage distribution map that conforms to component-level engineering consistency.

[0213] Specifically, the component consistency correction amount Δ for each component unit is calculated based on step S500. k The initial damage prediction probabilities of all voxels belonging to the same component unit are adjusted collaboratively. The aim is to transform the consistency judgment based on the overall component into a fine-grained calibration of the predictions of all voxels within the component, thereby generating a final result that is reasonable overall at the component scale and accurate in detail at the voxel scale.

[0214] The specific adjustment process is as follows:

[0215] For any component element k, a differentiable adjustment function f() is used to apply the correction amount of the component element to the initial prediction probability p of the u-th (u∈Vk) voxel within that component element. init u The corrected final damage probability p is obtained. final u :p final u =f(p init u, △ k ), where p final u Let be the final damage probability after correction for the u-th voxel.

[0216] The design of the adjustment function must meet two core requirements: first, to ensure that the overall statistics of the adjusted components approach the reference interval; and second, to ensure that the corrected probability value of each voxel still falls within the meaningful interval [0,1].

[0217] A typical implementation of the adjustment function is additive adjustment and truncation: p final u =clip(p init u +△ k ,0,1).

[0218] Where clip(x, 0, 1) is the cutoff function, and x is the input parameter of the cutoff function. In this formula, x = p init u +△ k The truncation function operates according to the following rules: output 0 when x < 0, output 1 when x > 1, and otherwise output x. Its significance lies in adjusting the component-level component consistency correction Δ. k The predicted value is applied uniformly to each voxel within the component. If △ k If Δ > 0, then the overall damage probability level of the component is increased; if Δ k If the value is less than 0, the overall probability of damage to the component is reduced. Ensure that the corrected probability value remains within the valid interval [0, 1].

[0219] After performing the above adjustment operations on all component units and all voxels contained in the structure, the final three-dimensional voxel damage probability distribution map covering the entire structure is obtained. This result not only preserves the spatial details and discrimination confidence at the voxel scale from the deep learning model, but also forces its statistical characteristics at the component scale to fall within the reasonable range defined by engineering experience or standards through component consistency correction, thereby fundamentally eliminating unreasonable contradictions at the component-wide level that may occur in voxel-level predictions.

[0220] This final distribution map can be directly used for 3D visualization, intuitively showing the spatial location and severity of damage; or a threshold (such as 0.5) can be set to binarize it into a damage mask, which can be used to quantitatively count the proportion of damage volume of each component, providing a direct, reliable and intuitive digital basis for structural safety assessment and decision-making.

[0221] Based on the same inventive concept, another embodiment of the present invention provides a three-dimensional voxel damage prediction device based on component consistency constraints. This device can be implemented using software, hardware, or a combination of both, and deployed in a server, workstation, or dedicated computing device. The device may include:

[0222] The preprocessing unit is used to acquire the three-dimensional geometric and physical property information of the structure to be evaluated, construct a three-dimensional voxel mesh, and establish the mapping relationship between component elements and voxels.

[0223] The input construction unit, connected to the preprocessing unit, is used to construct a multi-channel input tensor. The multi-channel input tensor includes at least a strength channel input tensor reflecting the inherent properties of the structure and a spatial distribution channel input tensor reflecting the influence of external loads.

[0224] The damage prediction unit, connected to the input construction unit, is used to input the multi-channel input tensor into a pre-trained 3D segmentation network model to obtain the initial damage prediction probability for each voxel. The 3D segmentation network model is optimized by an overall loss function that includes component consistency constraint loss. The component consistency constraint loss applies an interval constraint based on a preset damage level to the component-scale aggregation result of the voxel-level prediction through the mapping relationship between component units and voxels.

[0225] The scale aggregation unit, connected to the damage prediction unit and the preprocessing unit, is used to aggregate the initial damage prediction probability into a component-level damage characterization quantity for each component unit based on the mapping relationship between the component unit and the voxel.

[0226] The consistency correction unit, connected to the scale aggregation unit, is used to calculate the component consistency correction amount based on the component-level damage characterization amount of each component unit and the damage ratio reference interval defined by the preset damage level corresponding to each component unit.

[0227] The result generation unit, connected to the damage prediction unit and the consistency correction unit, is used to adjust the initial damage prediction probability of voxels belonging to the same component unit according to the component consistency correction amount, and generate and output the final three-dimensional voxel damage distribution map that conforms to the component-level engineering consistency.

[0228] The device may further include a training subsystem for training the 3D segmentation network model before deployment or during online updates. This subsystem includes:

[0229] Data Management Module: Used to manage training datasets, including multi-channel input tensors of sample structures, voxel-level ground truth labels, and component-level damage level labels.

[0230] Loss calculation module: used to calculate voxel-level loss (including classification loss of phased scheduling) and component consistency constraint loss.

[0231] Optimization module: Used to weight and combine the above losses into a total loss function, and update the network parameters through the backpropagation algorithm.

[0232] The device functions by one or more processors executing instructions (computer programs) stored in a memory. The memory stores executable instructions, which, when executed by the processors, cause the processors to control the aforementioned functional units to work together to complete the entire process from data input to the generation of the final damage distribution map.

[0233] It should be noted that the device embodiment and the method embodiment are based on the same inventive concept, and the functions of each unit correspond one-to-one with the method steps. Therefore, the descriptions of technical details, implementation methods, and technical effects in the method embodiment are also applicable to the device embodiment. Those skilled in the art will understand that the functional division of the above-mentioned units is logical and can be combined or further split in actual implementation, as long as their combination can achieve the overall function described in this invention.

[0234] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0235] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0236] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0237] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. 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 modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A three-dimensional voxel damage prediction method based on component consistency constraints, characterized by, The method includes the following steps: S100: Obtain the three-dimensional geometric and physical property information of the structure to be evaluated, construct a three-dimensional voxel mesh, and establish the mapping relationship between component units and voxels; S200, construct a multi-channel input tensor, wherein the multi-channel input tensor includes at least a strength channel input tensor reflecting the inherent properties of the structure and a spatial distribution channel input tensor reflecting the influence of external loads; S300, the multi-channel input tensor is input into a pre-trained 3D segmentation network model to obtain the initial damage prediction probability for each voxel; wherein, the 3D segmentation network model is optimized through an overall loss function that includes component consistency constraint loss, and the component consistency constraint loss applies interval constraints based on a preset damage level to the component-scale aggregation result of the voxel-level prediction through the mapping relationship between component units and voxels; the component consistency constraint loss satisfies the following conditions: L comp =∑ n c=1 (max(0,α L（c） -S c )+max(0,S c -β L（c） )); Among them, L comp For component consistency constraint loss, L(c) represents the preset damage level of component element c, α L（c） and β L（c） These are the lower and upper limits of the damage proportion reference interval corresponding to L(c), respectively, and S c Let c be the component-level damage characterization quantity for component element c, where c takes values ​​from 1 to n, and n is the total number of component elements. S400, based on the mapping relationship between the component unit and the voxel, the initial damage prediction probability is aggregated into a component-level damage characterization quantity for each component unit; S500, calculate the component consistency correction amount based on the component-level damage characterization quantity of each component unit and the damage ratio reference range defined by the preset damage level corresponding to each component unit; S600, based on the component consistency correction amount, adjust the initial damage prediction probability of voxels belonging to the same component unit, generate and output the final three-dimensional voxel damage distribution map that conforms to component-level engineering consistency.

2. The method of claim 1, wherein, In S300, the overall loss function also includes a voxel-level loss function, which is a weighted combination of an overlap loss term and a voxel-level classification loss term. During the training of the 3D segmentation network model, a phased weight scheduling strategy is used to dynamically adjust the weight of the voxel-level classification loss term.

3. The method of claim 2, wherein, The 3D segmentation network model is obtained through the following training steps: S1. Obtain the training dataset, which includes a three-dimensional voxel mesh of the sample structure, the mapping relationship between component units and voxels, multi-channel input tensors, voxel-level damage ground truth labels, and preset damage level labels of component units. S2, input the multi-channel input tensor of the training dataset into the 3D segmentation network model to be trained, perform forward computation to obtain the corresponding prediction results; S3, Calculate the voxel-level loss function based on the voxel-level damage prediction information and the voxel-level damage truth label in the prediction result; S4. Based on the mapping relationship between component units and voxels, the voxel-level damage prediction information is aggregated into the component-level damage characterization quantity of each component unit, and the component consistency constraint loss is calculated according to the damage ratio reference interval corresponding to the preset damage level label of the component unit. S5, the voxel-level loss function and the component consistency constraint loss are weighted and combined to form the overall loss function, and the parameters of the three-dimensional segmentation network model are updated through the backpropagation algorithm using the overall loss function to complete the model training.

4. The method of claim 3, wherein, In S4, the aggregation of voxel-level damage predictions into a component-level damage characterization quantity for each component unit specifically involves: for any component unit, based on the damage prediction probabilities of all voxels within that component unit, calculating a statistic characterizing the overall damage level of that component unit, which serves as the component-level damage characterization quantity for that component unit.

5. The method of claim 3, wherein, The voxel-level loss function satisfies the following condition: L voxel =L overlap +λ(t)·L focal , where L voxel L is the voxel-level loss function. overlap For the overlap loss term, L focal For voxel-level classification loss, λ(t) is the weight scheduling function that varies with the training period or number of iterations t.

6. The method of claim 5, wherein, The weight scheduling function is any one of a monotonically increasing linear function, a square function, a cosine function, or an exponential function; or, the weight scheduling function is a piecewise function that is dynamically adjusted according to the evaluation index of the validation set.

7. The method according to claim 1, characterized in that, The mapping relationship between the component unit and the voxel is provided in the form of a component semantic mask; in S400, the initial damage prediction probability is filtered through the component semantic mask to aggregate and obtain the component-level damage representation quantity of each component unit.

8. The method of claim 3, wherein, When the voxel-level damage truth labels are incomplete or contain noise, the component consistency constraint loss and the overlap loss term together constitute the main supervision signal for the model, forming a weakly supervised or semi-supervised training process.

9. A three-dimensional voxel damage prediction apparatus based on component consistency constraints, characterized by, include: The preprocessing unit is used to acquire the three-dimensional geometric and physical property information of the structure to be evaluated, construct a three-dimensional voxel mesh, and establish the mapping relationship between component elements and voxels. An input construction unit is used to construct a multi-channel input tensor, wherein the multi-channel input tensor includes at least a strength channel input tensor reflecting the inherent properties of the structure and a spatial distribution channel input tensor reflecting the influence of external loads; The damage prediction unit is used to input the multi-channel input tensor into a pre-trained 3D segmentation network model to obtain the initial damage prediction probability for each voxel. The 3D segmentation network model is optimized using an overall loss function that includes a component consistency constraint loss. This component consistency constraint loss applies an interval constraint based on a preset damage level to the voxel-level predicted component-scale aggregation result through the mapping relationship between component units and voxels. The component consistency constraint loss satisfies the following condition: L comp =∑ n c=1 (max(0,α L（c） -S c )+max(0,S c -β L（c） )); Among them, L comp For component consistency constraint loss, L(c) represents the preset damage level of component element c, α L（c） and β L（c） These are the lower and upper limits of the damage proportion reference interval corresponding to L(c), respectively, and S c Let c be the component-level damage characterization quantity for component element c, where c takes values ​​from 1 to n, and n is the total number of component elements. A scale aggregation unit is used to aggregate the initial damage prediction probability into a component-level damage characterization quantity for each component unit based on the mapping relationship between the component unit and the voxel. The consistency correction unit is used to calculate the component consistency correction amount based on the component-level damage characterization quantity of each component unit and the damage ratio reference range defined by the preset damage level corresponding to each component unit. The result generation unit is used to adjust the initial damage prediction probability of voxels belonging to the same component unit according to the component consistency correction amount, and generate and output the final three-dimensional voxel damage distribution map that conforms to the component-level engineering consistency.

Images