Method and system for evaluating ancient building wood structure based on three-dimensional modeling

Abstract

The application provides a three-dimensional modeling-based ancient building wood structure evaluation method and system, relates to the technical field of ancient building evaluation, and comprises the following steps: obtaining multi-source sensing data of an ancient building wood structure, identifying a damage area of a wood structure surface and extracting damage parameters based on surface image data and three-dimensional geometric data, the damage parameters comprising the position, geometric shape and depth level of damage; dynamically adjusting the reconstruction strategy of a three-dimensional reconstruction algorithm between the damage area and the non-damage area according to the damage parameters; mapping the damage parameters to a finite element analysis model as initial defect boundary conditions, and performing mechanical simulation on the finite element model with damage to output a structure safety evaluation result; and the method realizes high-confidence evaluation of ancient building wood structure damage, efficient modeling of damage areas in priority, and scientific multi-target repair decision-making.

Description

Technical Field

[0001] This invention relates to the field of ancient building assessment technology, and in particular to a method and system for assessing the timber structure of ancient buildings based on three-dimensional modeling. Background Technology

[0002] Structural safety assessment and restoration of ancient wooden buildings is a crucial task in the field of cultural relic preservation. Currently, damage detection and assessment of ancient wooden buildings typically employ various techniques, including optical image acquisition, 3D laser scanning, internal defect detection (such as infrared thermal imaging and ultrasound), and finite element numerical simulation. However, in practical applications, the integrated use of these techniques still has the following shortcomings:

[0003] Data acquired by different sensors (such as surface images, 3D point clouds, and internal detection data) are usually processed separately after acquisition, with weak correlation between them, making it difficult to form a comprehensive description of damage characteristics, resulting in low utilization of data resources.

[0004] Existing assessment methods rely heavily on single data sources or empirical estimations to determine key indicators such as damage depth and extent. These methods are easily affected by factors such as the field environment, material properties, and measurement noise, resulting in limitations in the stability and reliability of the assessment results.

[0005] In the process of generating 3D digital models based on point cloud data, existing reconstruction methods usually adopt globally uniform parameter settings, which makes it difficult to effectively control the overall computational overhead while ensuring the accuracy of local details. It is often necessary to make trade-offs between detail preservation and computational efficiency.

[0006] When introducing the detected damage information into the finite element model for structural analysis, existing methods often use approximation methods such as element deletion, uniform stiffness reduction, or equivalent loads. These methods fail to fully reflect the mechanical properties of wood as an anisotropic material and make it difficult to accurately simulate the actual impact of damage on the stress performance of components and the overall structure.

[0007] Existing assessment processes are typically based on one-time data collection and modeling. New data obtained during subsequent repair work or long-term monitoring is difficult to effectively use to correct or verify the initial assessment results, resulting in insufficient information feedback between the initial assessment and subsequent engineering practice. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a method and system for evaluating ancient wooden structures based on 3D modeling. By fusing multi-source sensing data to correct the confidence level of damage depth, and dynamically adjusting the 3D reconstruction strategy based on damage parameters, mechanical simulation is performed using a refined stiffness reduction model that considers the orthogonal anisotropy of wood and the angle of damage direction. This enables high-confidence assessment of damage to ancient wooden structures, efficient modeling prioritizing damaged areas, and scientific multi-objective repair decision-making.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] In a first aspect, embodiments of the present invention provide a method for evaluating the timber structure of ancient buildings based on three-dimensional modeling, the method comprising:

[0011] Acquire multi-source sensing data of ancient wooden structures, including surface image data, three-dimensional geometric data, and internal defect detection data;

[0012] Based on the surface image data and three-dimensional geometric data, damaged areas on the surface of the wooden structure are identified and damage parameters are extracted. The damage parameters include the location, geometric shape, and depth level of the damage. The depth level is corrected for confidence by fusing internal defect detection data.

[0013] Based on the damage parameters, the reconstruction strategy of the 3D reconstruction algorithm is dynamically adjusted between the damaged area and the undamaged area to generate a 3D digital model in which the geometric details of the damaged area are preserved to a higher degree than those of the undamaged area.

[0014] The damage parameters are used as initial defect boundary conditions and mapped to the finite element analysis model. Mechanical simulation is then performed on the damaged finite element model to output the structural safety assessment results.

[0015] On the other hand, embodiments of the present invention provide an evaluation system for ancient wooden structures based on three-dimensional modeling, used in the evaluation method for ancient wooden structures based on three-dimensional modeling described in embodiments of the present invention. The system includes:

[0016] The perception acquisition module is used to acquire multi-source perception data of the ancient wooden structure, including surface image data, three-dimensional geometric data and internal defect detection data.

[0017] The damage identification and parameter acquisition module is used to identify the damaged areas on the surface of the wood structure and extract damage parameters based on the surface image data and three-dimensional geometric data. The damage parameters include the location, geometric shape, and depth level of the damage. The depth level is corrected for confidence by fusing internal defect detection data.

[0018] The model reconstruction module is used to dynamically adjust the reconstruction strategy of the three-dimensional reconstruction algorithm between the damaged area and the non-damaged area according to the damage parameters, so as to generate a three-dimensional digital model in which the geometric details of the damaged area are preserved to a higher degree than those in the non-damaged area.

[0019] The simulation evaluation module is used to map the damage parameters as initial defect boundary conditions to the finite element analysis model, perform mechanical simulation on the damaged finite element model, and output the structural safety assessment results.

[0020] The beneficial effects of this invention include: by fusing surface image data, three-dimensional geometric data, and internal defect detection data, this invention establishes a multi-source collaborative mechanism in the damage depth level assessment stage, which includes preliminary judgment of visual / geometric features, spatial registration of internal detection data, and correction of overlap confidence. This effectively suppresses misjudgments caused by environmental noise, surface optical properties, or sensor accuracy from a single data source, and significantly improves the accuracy and robustness of damage depth level determination.

[0021] A three-dimensional damage influence field is generated based on the location, geometry, and depth level of the damage. Based on the Poisson surface reconstruction algorithm with an octree structure, the depth of the local octree is dynamically adjusted according to the weights of the damage influence field. While ensuring the integrity of the geometric details of the damaged area, the data volume of the overall model and the reconstruction computational overhead are significantly reduced, achieving the optimal balance between model accuracy and computational efficiency.

[0022] Based on the geometric shape and depth level of the damage, as well as the angle between the damage direction and the direction of the wood fibers and the principal stress direction of the component, anisotropic stiffness reduction is performed on the finite element elements traversed by the damage. This allows for a more realistic simulation of stress redistribution and residual load-bearing capacity under damage, improving the accuracy of finite element simulation in predicting the behavior of actual structures.

[0023] After completing the mechanical simulation of the damaged structure, multiple virtual repair schemes are simulated in parallel, and the recommended repair scheme with the smallest comprehensive distance is automatically selected. This overcomes the subjectivity of traditional reliance on a single indicator or engineering experience, and provides repair personnel with quantitative and traceable multi-objective decision-making basis, significantly improving the scientificity and rationality of the repair scheme.

[0024] The closed-loop mechanism enables the evaluation model to have adaptive calibration and continuous learning capabilities, effectively solving the problem that traditional single-stage performance evaluation cannot use measured data for model correction, thus improving the reusability and long-term reliability of the evaluation method. Based on the confidence correction mechanism, a machine learning proxy model based on random forest or extreme gradient boosting is constructed. The dual-model fusion strategy further improves the response speed and generalization ability of depth level determination, making it particularly suitable for rapid on-site evaluation scenarios. Attached Figure Description

[0025] The invention will now be further described with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart illustrating the evaluation method for ancient wooden structures based on three-dimensional modeling, as described in this invention. Detailed Implementation

[0027] 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.

[0028] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0029] Example 1, see Figure 1 This invention provides a method for evaluating the timber structure of ancient buildings based on three-dimensional modeling, the method comprising:

[0030] Step 1: Acquire multi-source sensing data of the ancient wooden structure, including surface image data, three-dimensional geometric data, and internal defect detection data;

[0031] Step 2: Based on the surface image data and three-dimensional geometric data, identify the damaged areas on the wooden structure surface and extract damage parameters, including the location, geometric shape, and depth level of the damage; wherein, the depth level is corrected for confidence by fusing internal defect detection data;

[0032] Step 3: Based on the damage parameters, dynamically adjust the reconstruction strategy of the 3D reconstruction algorithm between the damaged and undamaged areas to generate a 3D digital model in which the geometric details of the damaged area are preserved to a higher degree than those in the undamaged area.

[0033] Step 4: Use the damage parameters as initial defect boundary conditions, map them to the finite element analysis model, perform mechanical simulation on the damaged finite element model, and output the structural safety assessment results.

[0034] The process involves mapping damage parameters as initial defect boundary conditions to a finite element analysis model, performing mechanical simulations on the damaged finite element model, and outputting structural safety assessment results, including:

[0035] The location, geometry, and depth level in the damage parameters are converted into the initial damage state of the corresponding element in the finite element model to establish a finite element model with damage.

[0036] A preset load is applied to the damaged finite element model, and mechanical simulation is performed to calculate the stress distribution, deformation characteristics and remaining bearing capacity under the damaged state.

[0037] The structural safety assessment results are output based on the simulation results. The structural safety assessment results include at least one of the following: the ratio of the maximum stress in the damaged area to the material strength, the percentage of the remaining bearing capacity of the component to the design bearing capacity, and the impact level of the damage on the overall structural stability.

[0038] In practice, the surface image data is acquired using a high-resolution digital camera or a drone aerial photography system to record the two-dimensional morphological information of the texture, color, and visible damage (such as cracks, decay, and paint peeling) on ​​the wooden structure surface. During acquisition, multiple camera positions are set up around the exterior and key internal parts of the ancient building (such as beam-column joints, bracket sets, and mortise-and-tenon joints) to ensure coverage of all areas to be measured. Multiple exposures are performed under ring lighting or natural light conditions to obtain clear images with uniform illumination.

[0039] The three-dimensional geometric data is acquired using a three-dimensional laser scanner. Specifically, multiple control targets are set up around the ancient building, and a terrestrial laser scanner (such as a phase-type or pulse-type scanner) is used to perform a panoramic scan of the wooden structure. The scanning resolution is set according to the importance of the components (such as setting high resolution for column bases and beam-column joints, and medium resolution for general components) to acquire high-density three-dimensional point cloud data and record the spatial coordinates and geometric shape of the wooden structure.

[0040] The internal defect detection data is acquired using an infrared thermal imager or an ultrasonic detector. Infrared thermal imaging is used to detect cavities, delamination, and areas of abnormal moisture content within the wood; ultrasonic detection is used to measure crack depth and the spatial location of internal defects. During detection, measuring points are laid out along the surface of the component according to a preset grid, and the distribution range and depth information of abnormal internal areas are recorded.

[0041] When collecting the above three types of data, a unified spatial control point is set up using a total station or GPS to establish a global coordinate system, ensuring that different data sources can be fused and processed under the same spatial reference system.

[0042] Based on the surface image data and three-dimensional geometric data obtained in step one, damaged areas on the surface of the wooden structure are identified and damage parameters are extracted. The damage parameters include the location, geometry, and depth level of the damage.

[0043] Specifically, the surface image data is first preprocessed (e.g., image enhancement, denoising, histogram equalization), and then image segmentation algorithms (e.g., threshold-based segmentation, edge detection, or deep learning semantic segmentation models) are used to extract the two-dimensional contour of the damaged area, and the pixel location and geometric shape of the damage in the image are determined (e.g., crack length, width, shape factor, etc.).

[0044] Meanwhile, the three-dimensional geometric data (point cloud) is filtered, denoised, and simplified. An algorithm based on the difference of normal vectors or curvature changes is used to identify the boundary of the damaged area in the point cloud and extract the position coordinates, orientation, and spatial distribution of the damage in three-dimensional space.

[0045] The two-dimensional contour of the damage extracted from the image data is mapped to the three-dimensional point cloud coordinate system. Multi-source data fusion is achieved through spatial registration to determine the precise three-dimensional location and geometric parameters of the damaged area.

[0046] The initial value of the depth level is determined by combining visual width features (such as the average and maximum width of cracks) in the surface image data and spatial penetration features (such as the proportion of crack extension along the component thickness) in the three-dimensional geometric data. Then, the internal defect detection data obtained in step one (such as infrared thermal imaging anomaly areas and ultrasonic echo depth data) is spatially registered with the aforementioned surface damage areas. The spatial overlap between the internal anomaly areas and the surface damage areas is calculated. Based on the degree of overlap, the initial depth level is corrected for confidence, and the corrected depth level is output. This depth level is used to characterize the severity of the damage.

[0047] Based on the damage parameters extracted in step two (including the location, geometry, and depth level of the damage), the reconstruction strategy of the 3D reconstruction algorithm is dynamically adjusted between the damaged and undamaged areas to generate a 3D digital model in which the geometric details of the damaged area are preserved to a higher degree than those in the undamaged area.

[0048] Specifically, a three-dimensional damage influence field is constructed based on damage parameters, with higher weights assigned to the core damage region, lower weights assigned to the non-damage region, and continuous weight changes in the transition region.

[0049] The damage parameters extracted in step two are used as initial defect boundary conditions and mapped to the finite element analysis model. Specifically, the location, geometry, and depth level of the damage are converted into the initial damage state of the corresponding element in the finite element model, thus establishing a finite element model with damage.

[0050] On the damaged finite element model, the actual loads borne by the ancient wooden structure (such as self-weight, wind load, seismic action, etc.) are applied, and mechanical simulation is performed using linear elastic or elastoplastic finite element methods to calculate the stress distribution, deformation characteristics, and remaining bearing capacity of the structure under damaged conditions. Based on the simulation results, a structural safety assessment is output. This assessment includes, but is not limited to: the ratio of the maximum stress in the damaged area to the material strength, the percentage of the remaining bearing capacity of the component to the design bearing capacity, the impact level of the damage on the overall structural stability, and preliminary recommendations on whether repair measures are necessary.

[0051] Specifically, step four is implemented as follows:

[0052] First, the damage parameters extracted in step two (including damage location, geometry, and depth level) are mapped to the finite element analysis model as initial defect boundary conditions. Specifically, the mapping method is as follows: based on the damage's position coordinates in three-dimensional space, the finite element element numbers it passes through are determined; based on the damage's geometry (such as crack length, width, orientation, and curvature) and depth level, each passed element is assigned an initial damage state identifier. This initial damage state identifier at least includes the damage presence indicator, damage direction, and damage severity coefficient for that element. After mapping, a damaged finite element model containing the aforementioned initial damage information is established.

[0053] Secondly, loads that the ancient wooden structure might experience in actual service environments are applied to the damaged finite element model. These loads include, but are not limited to: the structure's self-weight (calculated based on timber density and component volume), wind loads (determined based on local basic wind pressure and shape coefficients), seismic forces (applied using response spectrum analysis or time history analysis based on seismic fortification intensity and site category), and shrinkage or expansion loads caused by temperature and humidity changes. The load application method is selected as static or dynamic loading depending on the analysis objective.

[0054] Then, mechanical simulations are performed using finite element analysis software (such as ANSYS, ABAQUS, or open-source computational programs). The simulation method can be selected as either linear elastic finite element method or elastoplastic finite element method, depending on the degree of damage and the importance of nonlinear behavior. Specifically, linear elastic analysis is used for conditions with minor damage and the material still within the elastic range; elastoplastic analysis is used for conditions with severe damage that may enter the plastic range or cause local buckling, and the orthotropic constitutive relation of the wood and the stiffness degradation rules of the damaged elements are defined (e.g., by reducing the elastic modulus).

[0055] Through simulation calculations, the following output results were obtained: stress distribution cloud map, displacement deformation map, and residual bearing capacity of the components under damage conditions. The residual bearing capacity was determined by gradually increasing the load until the structure reached its ultimate limit state or deformation became uncontrollable.

[0056] Finally, a structural safety assessment report is automatically generated based on the simulation results. The assessment report includes at least one or more of the following quantitative indicators:

[0057] The ratio of the maximum stress in the damaged area to the design value of the bending or shear strength of the timber used;

[0058] The percentage of the remaining bearing capacity of key components (such as beams, columns, and brackets) relative to the original design bearing capacity of the component;

[0059] The impact of damage on overall structural stability is classified into three levels (e.g., Level I (no significant impact), Level II (local impact), and Level III (decreased overall stability)).

[0060] Preliminary recommendations on whether remedial measures are needed (e.g., "observe", "local reinforcement required" or "stop use immediately and repair").

[0061] This embodiment does not limit the specific implementation method of stiffness degradation. For example, it can adopt uniform reduction, reduction based on the degree of damage, or anisotropic reduction as detailed below.

[0062] The effects of the above technical solution are as follows:

[0063] By integrating multi-source sensing data such as surface images, 3D point clouds, and internal defect detection, a confidence correction mechanism is introduced in the damage depth level determination, which improves the accuracy of damage assessment. Through an adaptive 3D reconstruction strategy that prioritizes the damaged area, a balance between high precision and high efficiency is achieved. By mapping damage parameters as initial defects to a finite element model for mechanical simulation, the impact of damage on structural safety can be objectively quantified, providing a scientific basis for the protection and restoration of ancient wooden structures.

[0064] In one possible implementation, the depth level is determined in the following way:

[0065] Visual width features of damage are extracted based on surface image data, and spatial penetration features of damage are extracted based on three-dimensional geometric data; the visual width features and spatial penetration features are input into a pre-trained regression model to output an initial depth level;

[0066] Spatial registration is performed between internal defect detection data and surface image data or three-dimensional geometric data to calculate the spatial overlap between internal abnormal areas and surface damaged areas.

[0067] The initial depth level is adjusted according to the spatial overlap: when the overlap is greater than the first threshold, the initial depth level is increased by one or two levels; when the overlap is less than the second threshold, the initial depth level is decreased by one level; otherwise, it remains unchanged.

[0068] The internal defect detection data includes at least one of infrared thermal imaging data or ultrasonic echo data, and the first threshold and the second threshold are determined by the statistical distribution of historical damage samples.

[0069] The visual width feature includes at least one of the average width of the damage, the maximum width, and the width change rate; the spatial penetration feature includes at least one of the extension depth of the damage along the thickness direction of the component and the penetration ratio.

[0070] The regression model employs multiple linear regression, support vector regression, or random forest regression algorithms, and is trained using historical damage samples. Each historical damage sample contains measured visual width features, spatial penetration features, and the corresponding nominal depth level.

[0071] The spatial registration adopts the iterative nearest point algorithm or the coordinate transformation method based on common target points. The spatial overlap is defined as the ratio of the overlapping volume of the internal abnormal region and the surface damage region to the total volume of the internal abnormal region, or as the percentage of the overlapping part to the area of ​​the surface damage region.

[0072] The first threshold and the second threshold are determined by collecting historical damage samples that have undergone destructive testing or tomography, calculating the overlap of each sample and its true depth level, and selecting the threshold that maximizes the consistency between the corrected depth level and the true depth level using receiver operating characteristic curve analysis or percentile method.

[0073] The internal defect detection data is obtained by setting up measuring points along the surface of the wooden component according to a preset grid, using an infrared thermal imager to collect thermal response differences, or using an ultrasonic detector to measure propagation time, wave velocity attenuation and echo characteristics.

[0074] In the specific implementation, the first step is to acquire surface image data and three-dimensional geometric data of the damaged area of ​​the ancient wooden structure to be assessed. The surface image data is acquired using a high-resolution digital camera, and the three-dimensional geometric data is obtained using a three-dimensional laser scanner.

[0075] The surface image data is processed for image enhancement, denoising, and edge detection to extract the two-dimensional contours of damage (such as cracks) in the image. Based on the calibration relationship between pixel scale and physical scale, the visual width features of the damage are calculated. These visual width features include, but are not limited to: the average width, maximum width, width change rate, and width distribution curve along the damage direction. For example, for a crack penetrating a beam, multiple measuring points can be selected at equal intervals along the crack length direction, the crack width at each measuring point can be calculated, and the average value can be taken as the average width, and the maximum value as the maximum width.

[0076] The three-dimensional geometric data (point cloud) is filtered, simplified, and triangulated to establish a three-dimensional surface model of the damaged area. Based on the distribution pattern of the damage in three-dimensional space, spatial penetration features are extracted. These features include: the depth of the damage along the thickness direction of the component, the ratio of the penetration depth to the total thickness of the component (i.e., the penetration ratio), and the degree of penetration in the thickness direction (e.g., no penetration, partial penetration, complete penetration). For example, for a crack at a beam-column joint, the maximum depth of the crack extending from the surface of the component into the interior can be calculated using point cloud data; dividing this depth by the total thickness of the component at that location yields the penetration ratio.

[0077] Construct a regression model whose inputs are the aforementioned visual width and spatial penetration features, and whose output is the initial depth level of the injury. The depth level is pre-divided into several levels (e.g., Level I, Level II, and Level III, representing mild, moderate, and severe injury, respectively). The regression model can employ algorithms such as multiple linear regression, support vector regression, or random forest regression.

[0078] The model training process is as follows: A large number of historical samples of damage to ancient wooden structures are collected. Each sample includes a nominal depth level (as a label) determined through on-site measurement or laboratory testing, as well as corresponding visual width and spatial penetration features. These sample data are used to train the regression model, enabling it to predict the initial depth level based on the input features. After training, the visual width and spatial penetration features of the current damage are input into the regression model to output the initial depth level.

[0079] Infrared thermal imagers or ultrasonic detectors are used to acquire defect detection data inside the wooden structure. Specifically:

[0080] Infrared thermal imaging: Under stable environmental conditions, thermal imaging scans are performed on the surface of wooden components to identify areas of abnormal temperature. Areas with internal cavities, delamination, or abnormal moisture content in wood typically exhibit different thermal response characteristics compared to the surrounding wood (such as significant temperature differences or abnormal heat conduction). Based on this, the location, shape, and extent of internal abnormal areas can be extracted.

[0081] Ultrasonic testing: Transmitting and receiving probes are arranged along a pre-defined grid on the surface of the component to measure the propagation time, wave velocity attenuation, and echo characteristics of ultrasonic waves in the wood. When ultrasonic waves encounter internal cracks, voids, or delamination, they will produce reflection, refraction, or sudden changes in wave velocity, which can be used to determine the spatial coordinates and depth of internal defects.

[0082] The aforementioned internal defect detection data are recorded in a unified spatial coordinate system (this coordinate system is registered with the coordinate systems of the surface image data and the three-dimensional geometric data through total station control points).

[0083] Spatial registration is performed between internal anomaly regions identified in internal defect detection data (such as high-temperature anomaly regions in infrared thermal imaging or strong reflection regions in ultrasonic echoes) and surface damage regions determined by surface image data or three-dimensional geometric data. Registration methods can include point cloud registration based on the iterative nearest point (ICP) algorithm or coordinate transformation based on common target points.

[0084] After registration, the overlap between the internal anomalous region and the surface damage region in three-dimensional space is calculated. The overlap is defined as the ratio of the volume (or area) of the overlap between the internal anomalous region and the surface damage region to the total volume (or area) of the internal anomalous region, or as the percentage of the overlapping portion to the area of ​​the surface damage region. The overlap reflects the spatial consistency between visible surface damage and hidden internal defects: a higher overlap indicates a stronger correlation between surface damage and internal defects; a lower overlap indicates that internal defects exist independently or that surface damage has not extended into the interior.

[0085] Based on the calculated overlap, the initial depth level is adjusted for confidence, according to the following rules:

[0086] When the overlap is greater than the first threshold (e.g., overlap > 0.7), it indicates that the surface damage and internal defects are highly consistent, and the actual severity of the damage may be greater than the degree reflected by the surface features. In this case, the initial depth level is increased by one or two levels (e.g., from level II to level III, or from level I to level II).

[0087] When the overlap is less than the second threshold (e.g., overlap < 0.3), it indicates that the surface damage is weakly associated with the internal defects, and the internal anomaly may be an independently existing defect (such as pre-existing voids or non-structural localized alteration), or the surface cracks may not have penetrated deep into the interior. In this case, the initial depth level is downgraded by one level (e.g., from level II to level I).

[0088] When the overlap is between the first and second thresholds, it indicates that the correlation between surface damage and internal defects is generally low, the confidence level of the initial depth level is high, and it remains unchanged.

[0089] The first and second thresholds are determined through the statistical distribution of a large number of historical damage samples. For example, samples that have undergone actual destructive testing or CT scans are collected, and the relationship between their overlap and the true damage level is statistically analyzed. The optimal threshold is selected through ROC curve analysis or percentile method to achieve the highest consistency between the corrected depth level and the true damage level.

[0090] After the confidence level correction described above, the final damage depth level is output. This depth level is used in subsequent structural safety assessments to determine the stiffness reduction factor, assign initial defect values ​​to the finite element model, and select repair schemes.

[0091] The effects of the above technical solution are as follows:

[0092] By extracting visual width and spatial penetration features from surface images and 3D point clouds, an initial depth level is output through a regression model. Then, the initial level is spatially consistent with and corrected using internal defect data obtained by infrared thermal imaging or ultrasound. This forms a closed-loop mechanism of multi-source data cross-validation, which effectively suppresses random errors from a single data source and significantly improves the consistency between the depth level and the actual damage level.

[0093] Infrared thermal imaging can detect cavities, delamination, and areas with abnormal moisture content inside wood, while ultrasonic testing can accurately measure crack depth and the location of internal defects. This method spatially registers these internal detection data with visible surface damage and calculates the degree of overlap. When the degree of overlap is high, the depth level is increased to promptly detect serious damage that is shallow on the surface but deep inside; when the degree of overlap is low, the level is decreased to eliminate interference from false damage caused by surface texture or paint. This effectively solves the problem of traditional methods that only look at the surface and ignore the internal structure, or where internal anomalies are disconnected from surface damage.

[0094] The system employs a correction rule of adjusting the threshold based on whether the value is greater than a first threshold, lowering it if it is less than a second threshold, and otherwise leaving it unchanged. The thresholds are quantified using ROC curves or percentile methods from historical samples, avoiding subjective arbitrariness. Visual width features (average width, maximum width, and width change rate) and spatial penetration features (depth of extension, penetration ratio) can be reliably extracted using conventional image processing and point cloud analysis techniques. The regression models (linear regression, support vector regression, or random forest) have mature training processes and high computational efficiency, and the overall process is easily integrated into ancient building detection and evaluation software systems.

[0095] The revised depth level is directly used in the finite element model to assign stiffness reduction coefficients to damaged elements, determine initial defect boundary conditions, and prioritize repair schemes. Compared to the traditional empirically assigned depth level, the depth level output by this method has higher confidence, making subsequent mechanical simulation results closer to the actual response of the structure, thereby improving the scientificity and effectiveness of the entire ancient wooden structure evaluation method.

[0096] In one possible implementation, the reconstruction strategy of the dynamically adjusted 3D reconstruction algorithm includes:

[0097] Based on the location and depth level in the damage parameters, a three-dimensional damage influence field is generated. The damage influence field has continuously transitioning weights within the damage area, and the weights are positively correlated with the depth level.

[0098] The Poisson surface reconstruction algorithm using an octree structure dynamically adjusts the local octree depth based on the weights of the damage influence field: in regions where the weights are higher than the first weight threshold, the octree depth is set to the maximum depth value; in regions where the weights are lower than the second weight threshold, the octree depth is set to the minimum depth value; and in the intermediate region, the octree depth is determined by linear interpolation.

[0099] The maximum depth value is at least three levels greater than the minimum depth value, such that the mesh side length of the damaged area is less than half the mesh side length of the undamaged area.

[0100] In one possible implementation, when generating the three-dimensional damage influence field, the method further includes:

[0101] A morphological expansion operation is performed on the initial damaged area to obtain a damage transition zone, with the expansion radius being proportional to the depth level.

[0102] Within the damage transition zone, the weights smoothly decay from the highest value in the damage core zone to zero, and the decay function uses a Gaussian kernel or an exponential kernel.

[0103] The dynamic adjustment of the octree depth is also constrained by the spatial curvature of the damaged region: when the curvature of the damaged region exceeds a preset curvature threshold, the octree depth of the corresponding region is forcibly increased to the maximum depth value.

[0104] In the specific implementation, after identifying the damaged area and extracting the damage parameters, the location coordinates, geometric contours, and depth levels of the damage are obtained. To guide the subsequent reconstruction algorithm to use different accuracies in different areas, a three-dimensional damage influence field is first constructed. This influence field is a scalar field defined in the three-dimensional space of the wooden structure. Each spatial point in the field is assigned a weight, and the magnitude of the weight reflects the degree to which the point is affected by the damage.

[0105] For example, the weighting rules are as follows: within the core area of ​​the damage determined by the damage parameters, the weight is positively correlated with the damage depth level. For instance, when the depth level is I (mild), the weight of the core area is set to 0.4; when it is II (moderate), the weight is set to 0.7; and when it is III (severe), the weight is set to 1.0. The weights can also be linearly mapped according to the depth level to ensure that areas with more severe damage receive higher weights.

[0106] To avoid abrupt weight shifts between damaged and undamaged regions that could lead to discontinuities or step-like artifacts in the reconstructed model, this embodiment also performs morphological dilation on the initial damaged region. Specifically, starting from the boundary of the core damaged region, a certain radius is extended outward to form a damage transition zone. The dilation radius is proportional to the damage depth level: the higher the depth level, the larger the dilation radius (e.g., 10 mm for level I, 20 mm for level II, and 30 mm for level III). Within the transition zone, the weights smoothly decay radially outward from the highest value in the core region to zero. A Gaussian kernel or an exponential kernel can be used as the decay function.

[0107] After obtaining the three-dimensional damage influence field, a three-dimensional digital model is generated using a Poisson surface reconstruction algorithm based on an octree structure. The Poisson surface reconstruction algorithm adaptively partitions the space by constructing an octree, with each octree node corresponding to a spatial cube region. The Poisson equation is then solved within this region to fit the surface. Traditional methods use a globally uniform maximum octree depth; however, this embodiment dynamically adjusts the octree depth for each local region based on the weights of the damage influence field.

[0108] The specific dynamic adjustment strategy is as follows:

[0109] For regions with weights higher than the first weight threshold (e.g., threshold T1=0.8), it indicates that the region is severely damaged and requires preservation of subtle geometric features (such as crack edges and localized damage to tenons and mortises). In this case, the octree depth for that region is set to the maximum depth value. The maximum depth value is typically 8 to 10 layers, corresponding to mesh edge lengths in the millimeter or sub-millimeter range.

[0110] For regions with weights below the second weight threshold (e.g., threshold T2 = 0.2), it indicates that the region is undamaged or minimally damaged, and has low requirements for geometric detail. In this case, the octree depth of this region is set to the minimum depth value. The minimum depth value is usually 5 to 6 layers, corresponding to a mesh side length in the centimeter range.

[0111] For the transition region where the weights fall between the first and second weight thresholds, the octree depth D is determined by linear interpolation based on the weight values:

[0112] D=Dmin+(Dmax-Dmin)×(q-T2) / (T1-T2)

[0113] Dmin is the minimum depth; Dmax is the maximum depth; q is the weight of the region;

[0114] With the above settings, the maximum depth value is at least three levels larger than the minimum depth value, so that the mesh side length of the damaged area is less than half the mesh side length of the undamaged area.

[0115] In the damaged area, some parts, although not of high depth, have complex spatial geometry (such as crack bifurcation, acute corners at mortise and tenon joints, or high-curvature surfaces). If weighted linear interpolation is used, sufficient resolution may not be obtained, leading to model distortion. Therefore, this embodiment also introduces a curvature constraint mechanism.

[0116] After generating the 3D damage influence field, local curvature calculations are performed on the point cloud or preliminarily reconstructed surface mesh of the damaged area. Principal component analysis (PCA) based on neighborhood point sets can be used to estimate the normal vector change of each point, and then Gaussian curvature or average curvature can be calculated. When the curvature of a point exceeds a preset curvature threshold (e.g., average curvature greater than 0.5 mm⁻¹), regardless of the weight of the damage influence field at that point, the octree depth of the local region where that point is located is forcibly increased to the maximum depth value. This forced increase operation ensures that geometrically abrupt parts such as crack tips and tenon edges are sufficiently refined, avoiding model smoothing or feature loss.

[0117] After completing the dynamic adjustment of the octree depth and the forced curvature boosting, the Poisson surface reconstruction solver is invoked to solve the implicit surface function within each octree node, extract isosurfaces, and perform mesh simplification and smoothing to finally generate a 3D digital model. This model features a fine mesh and rich geometric details in the damage core and high curvature regions, while a coarser mesh is used in the non-damage regions to reduce data volume. The mesh size changes continuously in the transition regions. The overall model ensures the accuracy required for damage assessment while controlling computational and storage costs.

[0118] The effects of the above technical solution are as follows:

[0119] By guiding the dynamic adjustment of the octree depth through a 3D damage influence field, the highest resolution is used only in severely damaged and high curvature regions, while a lower resolution is used in undamaged regions. This avoids computational redundancy caused by global high precision, improves model generation speed, and reduces file size.

[0120] The damage core area uses the maximum octree depth, and the mesh side length can reach the millimeter level or even the sub-millimeter level. It can completely preserve the sharp features of the crack edge, the local damage morphology of the tenon and mortise nodes, and the complex texture, providing high-fidelity geometric input for subsequent finite element analysis.

[0121] Morphological dilation and Gaussian / exponential kernel weight decay ensure a continuous transition of weights between damaged and undamaged areas, while octree depth linear interpolation allows the mesh size to change gradually, avoiding step-like or crack-like splicing marks on the model surface.

[0122] Curvature constraint mechanisms can force a higher resolution at geometric abrupt changes such as crack tips and tenon corners, preventing the loss of important features due to low weights and ensuring the geometric fidelity of the model.

[0123] The entire process is based on three-dimensional point cloud processing and Poisson reconstruction algorithm. The threshold parameter, depth level mapping relationship, expansion radius, and curvature threshold can all be calibrated and adjusted according to actual scanning accuracy and damage types, with good adaptability and operability.

[0124] In a possible implementation, mapping the damage parameters to a finite element analysis model as initial defect boundary conditions includes:

[0125] Determining the finite element mesh elements penetrated by the damage according to the geometric shape and depth level of the damage;

[0126] Performing non-isotropic stiffness reduction on the penetrated elements, where the reduced elastic matrix is expressed as the undamaged elastic matrix minus the stiffness degradation matrix caused by damage;

[0127] The stiffness degradation matrix adopts a wood orthotropic constitutive model, wherein the diagonal elements respectively correspond to the elastic modulus reduction coefficients in the grain direction, radial direction and chord direction, and the off-diagonal elements are reduced proportionally;

[0128] The elastic modulus reduction coefficient is jointly determined according to the included angle between the damage trend and the wood fiber direction and the included angle between the damage trend and the main stress direction of the component.

[0129] The elastic modulus reduction coefficient is determined in the following manner:

[0130] Defining the damage normal direction and damage tangential direction, and defining the main stress direction and the wood fiber direction;

[0131] Calculating a first included angle, which is the included angle between the damage trend and the main stress direction; calculating a second included angle, which is the included angle between the damage trend and the wood fiber direction;

[0132] The elastic modulus reduction coefficient in the grain direction is calculated by the following formula:

[0133] Dg=1-(w1×sin²θ1+w2×cosθ2)

[0134] wherein Dg is the elastic modulus reduction coefficient in the grain direction; θ1 is the first included angle; θ2 is the second included angle; w1 and w2 are weight coefficients, w1+w2=1, 0<w1<1, 0<w2<1 and w1>w2;

[0135] The elastic modulus reduction coefficients in the radial direction and chord direction are respectively calculated by the following formulas:

[0136] Dr=1-(u1×sin²θ1+u2×cosθ2)

[0137] Dt=1-(v1×sin²θ1+v2×cosθ2)

[0138] Wherein, u1 and u2 are the weight coefficients in the radial direction, v1 and v2 are the weight coefficients in the chord direction, which satisfy 0<u2<u1<1, u1+u2=1, and 0<v2<v1<1, v1+v2=1 respectively; and satisfy u1<w1, v1<w1;

[0139] When the first included angle is less than or equal to the first angle threshold and the second included angle is less than or equal to the second angle threshold, multiply the calculated reduction coefficients along the grain, in the radial direction and in the chord direction by the forced attenuation factor γ, where 0<γ<1.

[0140] After completing the construction of the three-dimensional digital model and the extraction of damage parameters, a finite element analysis model of the ancient building timber structure is established; first, mesh the three-dimensional model to generate a finite element mesh composed of hexahedral or tetrahedral units. According to the position coordinates and geometric morphology in the damage parameters (for example, the spatial trend and depth distribution of cracks), a spatial intersection detection algorithm (such as the ray casting method or the voxel traversal method) is used to determine which finite element units the damaged area passes through. For each crack or each damaged area, record the number of the unit it passes through and the coordinates of the entry point and exit point inside the unit, so as to facilitate subsequent stiffness reduction.

[0141] For units passed through by damage, instead of simple unit deletion or uniform stiffness reduction, non-isotropic stiffness reduction is performed. The reduced unit elasticity matrix C damaged is expressed as the undamaged elasticity matrix minus the stiffness degradation matrix caused by damage:

[0142] C damaged =C undamaged -C deg

[0143] C undamaged is the elasticity matrix when the wood is undamaged, which is constructed based on the orthotropic constitutive model, and includes three elastic moduli (along the grain, radial direction, chord direction), three shear moduli and corresponding Poisson's ratios; C deg is a stiffness degradation matrix, whose diagonal elements correspond to the reductions of the elastic moduli in the along-grain, radial and chord directions respectively, and the off-diagonal elements are reduced in the same proportion as the diagonal elements to maintain the consistency of the constitutive relationship.

[0144] The determination of the elastic modulus reduction coefficient is the core of stiffness reduction; this embodiment introduces two key angles, namely the included angle between the damage trend and the wood fiber direction, and the included angle between the damage trend and the principal stress direction of the component, and calculates the reduction coefficients along the grain, in the radial direction and in the chord direction respectively.

[0145] First, the following geometric relationships are defined:

[0146] Damage normal direction: The direction perpendicular to the damaged surface (such as the crack surface).

[0147] Damage tangent direction (damage direction): The direction in which the damaged surface extends along the surface of the component.

[0148] Wood fiber direction: parallel to the grain direction of the wood component.

[0149] Principal stress direction: The direction of the maximum principal tensile stress of a component under the main load, obtained through non-destructive finite element analysis or structural mechanics calculation.

[0150] The definitions of parallel, radial, and tangential directions follow the conventional conventions of orthogonal anisotropic constitutive models of wood:

[0151] Along the grain: parallel to the direction of wood fiber growth;

[0152] Radial direction: perpendicular to the fiber growth direction and pointing towards the pith;

[0153] Tangential direction: The direction perpendicular to the fiber growth direction and tangent to the annual rings.

[0154] Based on the above definition, calculate the two included angles:

[0155] First included angle: The angle between the damage direction and the principal stress direction (take an acute angle, ranging from 0° to 90°).

[0156] Second included angle: The angle between the direction of damage and the direction of wood fibers (take the acute angle).

[0157] Based on the aforementioned formula, calculate the reduction factor in each direction.

[0158] In certain extreme cases, the damage trajectory may simultaneously approach both the principal stress direction and the wood fiber direction (i.e., both the first and second included angles are very small), resulting in a severe weakening of the load-bearing capacity of the wood structure. For safety reasons, this embodiment introduces a forced attenuation mechanism.

[0159] Set a first angle threshold (e.g., 15°) and a second angle threshold (e.g., 15°). When the first included angle is less than or equal to the first angle threshold and the second included angle is less than or equal to the second angle threshold, multiply the calculated reduction coefficients in the longitudinal, radial, and chordal directions by a forced attenuation factor γ, where 0 < γ < 1, preferably γ = 0.7.

[0160] The aforementioned first angle threshold, second angle threshold, forced attenuation factor γ, and weighting coefficient can be determined by those skilled in the art through any of the following methods, based on the wood species, damage characteristics, and structural safety level:

[0161] The values ​​were determined by referring to the orthogonal anisotropic elastic constants and fracture parameters of the same type of wood given in the wood mechanical property database (such as the "Wood Design Manual" or relevant wood constitutive model literature).

[0162] Using new specimens of the same type of wood, the stiffness degradation law under different crack orientations was determined according to standard mechanical test methods (such as three-point bending test or compact tensile test), and the above parameters were calibrated by finite element inversion.

[0163] For commonly used timbers in traditional buildings (such as pine, fir, and nanmu), the optimal parameters verified by experiments can be directly adopted.

[0164] In practical engineering applications, if the type of wood or the type of damage changes, those skilled in the art can obtain reasonable results simply by making routine adjustments to the above parameters.

[0165] The effects of the above technical solution are as follows:

[0166] Based on the orthogonal anisotropic constitutive model of wood, non-isotropic stiffness reduction is performed, distinguishing different reduction degrees along the grain, radial, and tangential directions. This method is closer to the actual mechanical behavior of wood than traditional element deletion or uniform reduction methods.

[0167] By combining the angle between the damage direction and the principal stress direction (reflecting the influence of the stress state on damage propagation) and the angle between the damage direction and the wood fiber direction (reflecting the influence of material anisotropy on damage weakening), the reduction factor calculation formula can be more accurately quantified to determine the differentiated impact of damage with different directions on the load-bearing capacity of the component.

[0168] The forced attenuation mechanism automatically increases the reduction amplitude when both included angles are small, avoiding the assessment results being too dangerous under critical conditions. It is especially suitable for important cultural heritages with low safety redundancy, such as ancient wooden structures.

[0169] Based on existing finite element software (such as ANSYS and ABAQUS) user subroutines or secondary development interfaces, the above stiffness reduction rules can be easily implemented, and they have good engineering feasibility.

[0170] In one possible implementation, the stiffness reduction is further performed in a layered, non-uniform manner based on the distribution of damage depth along the thickness of the member section, including:

[0171] Determine the set of finite element elements that the damage passes through in each zone, and calculate the elastic modulus reduction factor in the longitudinal direction, radial direction and chord direction for each element in each zone;

[0172] For each of the surface, middle, and deep regions, the elastic modulus reduction coefficients of the units within that region in the grain direction, radial direction, and tangential direction are determined based on the angle between the damage direction and the wood fiber direction, and the angle between the damage direction and the principal stress direction of the component, respectively, and then multiplied by the layering correction factor corresponding to that region; wherein, the layering correction factor for the surface region is greater than 1, the layering correction factor for the deep region is between 0 and 1, and the layering correction factor for the middle region is equal to 1.

[0173] When the damage depth distribution curve of the unit in the middle layer shows that the damage penetrates the entire cross section, the elastic modulus reduction factor in the longitudinal direction of the unit in the middle layer is forcibly set to a preset lower limit (e.g., 0.2), and an emergency check of the bearing capacity is triggered; the elastic modulus reduction factors in the radial and chord directions remain unchanged or are processed according to the same rules.

[0174] Based on the stiffness reduction method, further differentiated stiffness reduction treatments are applied to the surface, middle, and deep layers according to the distribution of damage depth along the component cross-sectional thickness. This is to more accurately reflect the impact of damage at different depths on the load-bearing capacity of timber structures. The specific steps include:

[0175] First, the depth distribution curve of the damage along the thickness direction of the component section is obtained. This curve can be obtained by fitting the extension trajectory of the crack along the thickness direction in the three-dimensional point cloud data, or by inversion after fusing infrared thermal imaging and ultrasonic detection data.

[0176] Based on the depth distribution curve, the component cross-section is divided into three regions along the thickness direction: the surface region (close to the outer surface of the component, typically accounting for 0 to 1 / 3 of the cross-sectional thickness), the middle region (the middle of the cross-section, accounting for approximately 1 / 3 to 2 / 3), and the deep region (close to the other side of the component's surface or deep inside, accounting for approximately 2 / 3 to 1 / 3). The division ratio can be appropriately adjusted according to the actual size and damage characteristics of the component; for example, the proportion of the middle region can be appropriately increased for thin-walled components.

[0177] Then, the set of finite element elements traversed by the damage within each partition is determined. Since damage (such as cracks) may not be completely continuous in the thickness direction of the cross section, the elements traversed by damage at different depths may be different. Using a spatial intersection detection algorithm, Boolean operations are performed between the damage trajectory and the finite element mesh to create an independent list of damage elements for each partition.

[0178] For each unit within each zone (surface zone, middle zone, and deep zone), the elastic modulus reduction factors in the grain direction, radial direction, and tangential direction are calculated separately. The basic reduction factors in these three directions are determined according to the aforementioned formula based on the double angles (the angle between the damage direction and the principal stress direction, and the angle between the damage direction and the wood fiber direction), that is, the basic reduction values ​​in the three directions are calculated based on the local direction of the damage within the unit.

[0179] After obtaining the basic reduction factor, based on the partition type of the cell, the basic reduction factors in all three directions are multiplied by a stratification correction factor corresponding to that partition:

[0180] For units within the surface region, the stratification correction factor is greater than 1 (e.g., 1.2). This is because surface damage is directly exposed to the external environment and is often accompanied by wood surface aging, cracking, and moisture changes, resulting in a more significant weakening effect on local stiffness than the same damage inside the surface.

[0181] For elements in the deep region, the stratification correction factor is between 0 and 1 (e.g., 0.8). Deep damage is encased in surrounding timber, and the resulting stress concentration and stiffness loss are constrained and shared to some extent by the outer timber; therefore, the actual reduction can be appropriately lowered.

[0182] For units within the middle layer, the stratification correction factor is equal to 1, meaning the basic reduction factor remains unchanged.

[0183] It should be noted that the specific values ​​of the above-mentioned stratification correction factors (such as 1.2, 0.8) are recommended values ​​for commonly used woods such as pine and fir. Those skilled in the art can calibrate or adjust these values ​​through comparative experiments (such as static load tests on specimens with pre-existing cracks of different depths) or by referring to the correction coefficients for surface cracks and internal cracks in the Wood Mechanics Handbook.

[0184] When the damage depth distribution curve of the unit in the middle layer shows that the damage has penetrated the entire cross section (i.e., the crack extends completely from one side of the component to the other side), it indicates that the damage has completely cut off the continuous force transmission path of the component on that cross section, posing a serious threat to structural safety.

[0185] At this point, the reduction factor of the elastic modulus in the direction parallel to the grain of all elements in the middle layer is forcibly set to a preset lower limit (e.g., 0.2), meaning that the remaining elastic modulus in this direction does not exceed 20% of the undamaged value. Simultaneously, an emergency load-bearing capacity check procedure is triggered, which means that in subsequent finite element analyses, the stiffness in the direction parallel to the grain of these elements is automatically reduced to an extremely low value, and a warning message is output to prompt structural engineers to pay close attention to this area.

[0186] The reduction factors for the elastic modulus in the radial and chord directions can be chosen to remain unchanged (based on the layered correction values) or multiplied by an additional forced attenuation factor, depending on the actual situation. Generally, since through cracks mainly weaken the tensile and compressive stiffness along the grain, their impact on the radial and chord directions is relatively small. Therefore, it is recommended to keep the reduction factors in the radial and chord directions unchanged to avoid overly conservative assessments.

[0187] After assigning the elastic modulus reduction coefficient to each element in the above partition, a new orthogonal anisotropic elastic matrix is ​​constructed for each damaged element (i.e., the original undamaged elastic matrix minus the corresponding stiffness degradation matrix). Then, all element stiffness matrices are assembled into an overall stiffness matrix, and boundary conditions and external loads are applied for finite element solution to obtain the stress, deformation and residual bearing capacity of the structure under layered non-uniform damage state.

[0188] The effects of the above technical solution are as follows:

[0189] It more realistically reflects the actual impact of damage along the depth direction, enabling more precise differentiation of the varying effects of damage at different depths on structural performance. A clear safety baseline is set for through-section damage; when through-cracks appear in the middle layer, the reduction factor along the grain is forcibly reduced to an extremely low value, triggering an emergency check. This avoids assessment results being biased towards danger and ensures timely and reliable safety warnings for severely damaged components. Implementation is simple and parameters are adjustable. The layer division ratio and layer correction factor can be adjusted according to component size, wood type, and damage morphology using conventional engineering methods. It is highly compatible with existing finite element software (such as ABAQUS and ANSYS) user subroutine interfaces, facilitating engineering implementation. Layered processing adds a correction layer on top of the basic reduction factor; the two form a complementary relationship. The double-angle rule reflects the weakening mechanism of stiffness due to damage direction, while the layer correction reflects the degree of influence of damage depth on stiffness, jointly improving the simulation accuracy of the finite element model for complex damage states.

[0190] In one possible implementation, when the same finite element element is traversed by multiple damaged regions, the elastic modulus reduction factors in each direction are superimposed according to the following rules:

[0191] Determine the reduction coefficients of elastic modulus along the grain, radial elastic modulus, and chordal elastic modulus under the individual action of each type of damage;

[0192] For each of the longitudinal, radial, and chordal directions, the combined elastic modulus reduction factor in that direction is equal to the product of a factor minus the complement of each individual damage reduction factor in that direction, where the complement is a factor minus the individual damage reduction factor in that direction.

[0193] When at least one through-section damage exists among multiple damages, for each of the parallel, radial, and chordal directions, the combined elastic modulus reduction factor in that direction is corrected by multiplying the correction factor in that direction by the minimum value among all single-damage elastic modulus reduction factors in that direction; wherein, the correction factor in each direction is greater than 0 and less than 1, and the correction factor in the parallel direction is not greater than the minimum value among the correction factors in the radial and chordal directions.

[0194] Preferably, the correction factor along the grain is 0.9, the radial correction factor is 0.95, and the chordal correction factor is 0.95.

[0195] In the specific implementation,

[0196] For multiple damages passing through the same finite element element, the elastic modulus reduction coefficients of the element in the grain direction, radial direction, and chord direction are first determined for each damage under individual action. The single damage reduction coefficients in these three directions can be calculated according to the aforementioned formula based on the double angles (the angle between the damage direction and the principal stress direction, and the angle between the damage direction and the wood fiber direction), and are denoted as: the grain direction reduction coefficient, radial direction reduction coefficient, and chord direction reduction coefficient corresponding to the i-th damage.

[0197] For example, if a unit is simultaneously traversed by a main crack and a branch crack, the reduction coefficients in the three directions when the main crack exists alone, and the reduction coefficients in the three directions when the branch crack exists alone, are calculated separately.

[0198] The reduction factors for the elastic modulus in each direction after superposition are:

[0199]

[0200]

[0201]

[0202] , , These are the elastic modulus reduction factors in the longitudinal, radial, and chordal directions after superposition, respectively; the product symbol ∏ indicates a product from i=1 to n; Let be the reduction factor of the elastic modulus of wood along the grain under the action of the i-th damage alone; Let be the reduction factor of the elastic modulus of wood in the radial direction under the action of the i-th damage alone; is the reduction factor of the elastic modulus of wood in the tangential direction under the action of the i-th damage alone; n is the total number of damage regions passing through the same finite element element.

[0203] The physical significance of the superposition formula is: when multiple damages coexist, the weakening effect on stiffness is calculated according to the probability superposition method of independent events, which avoids the unreasonable situation that the reduction coefficient exceeds 1 caused by simple linear superposition. When there is a serious damage penetrating through the cross-section, the superposition result is replaced by the product of the minimum single-damage reduction coefficient and a correction factor, so as to reflect the controlling influence of the penetrating damage on the overall stiffness.

[0204] When at least one damage penetrating through the cross-section exists among multiple damages (that is, a crack extends completely from one side of the component to the other side, or a void completely penetrates the thickness of the cross-section), the above probability superposition result needs to be forcibly corrected.

[0205] The correction rule is: for each direction among the grain direction, radial direction and chordal direction, the final comprehensive reduction coefficient in this direction no longer adopts the probability superposition result, but is directly obtained by multiplying the correction factor corresponding to this direction by the minimum value of all single-damage reduction coefficients in this direction.

[0206] The comprehensive elastic modulus reduction coefficient in the grain direction is corrected as: =k_g×min(D_g i ); wherein, 0<k_g<1;

[0207] The comprehensive elastic modulus reduction coefficient in the radial direction is corrected as: =k_r×min(D_r i ); wherein, 0<k_r<1;

[0208] The comprehensive elastic modulus reduction coefficient in the chordal direction is corrected as: =k_t×min(D_r i ); wherein, 0<k_t<1;

[0209] k_g<min(k_r,k_t); k_g is the grain direction reduction coefficient correction factor for the damage penetrating through the cross-section; k_r is the radial direction reduction coefficient correction factor for the damage penetrating through the cross-section; k_t is the chordal direction reduction coefficient correction factor for the damage penetrating through the cross-section;

[0210] In practical engineering, for conventional ancient building wood (such as pine and fir), the following preferred parameters are recommended: the correction factor in the grain direction is 0.9, the correction factor in the radial direction is 0.95, and the correction factor in the chordal direction is 0.95. If higher safety redundancy is required, the correction factor in the grain direction can be further reduced (e.g., to 0.8).

[0211] The effect of the above technical solution is:

[0212] To avoid the physical inconsistencies of linear superposition, a probabilistic superposition rule (product of complements) is adopted to calculate the comprehensive reduction coefficient, ensuring that the reduction result always lies between 0 and 1 when multiple damages act together, which conforms to the basic constraints of mechanics of materials and avoids the non-positive stiffness problem that may occur with simple linear summation.

[0213] This approach highlights the controlling effect of through-section damage. When severe through-section cracks exist in the damage, replacing the probability superposition result with the minimum value multiplied by a correction factor emphasizes the weakest link effect—the most severe damage determines the residual stiffness of the element, and the correction factor further provides safety redundancy. This approach is more in line with the weakest link principle in structural damage engineering.

[0214] The anisotropic reduction is clearly distinguished. The longitudinal, radial, and tangential directions are calculated independently, and the correction factor for the longitudinal direction is no greater than that for the radial and tangential directions when correcting for through-sections, which fully reflects the characteristic that the longitudinal direction is most sensitive to through-damage under orthotropic wood.

[0215] The parameters are adjustable, making it highly adaptable to engineering projects. The correction factor can be routinely adjusted according to the type of timber, the structural importance level, and safety requirements. For general historical buildings, the recommended values ​​are sufficient to obtain a reasonable assessment; for particularly precious cultural relics, the grain-parallel correction factor can be appropriately reduced to increase the safety margin.

[0216] It only involves basic arithmetic operations and minimum value operations, which are easy to implement in user subroutines or secondary development scripts of finite element software. It has low computational overhead and is suitable for rapid evaluation of large-scale ancient building mesh models.

[0217] In one possible implementation, the mechanical simulation includes:

[0218] Simulate at least two different virtual repair operations on a damaged finite element model;

[0219] Calculate the remaining bearing capacity and stress redistribution cloud map after each repair operation;

[0220] The structural safety assessment results include the percentage of remaining bearing capacity corresponding to each repair operation and the recommended repair scheme; wherein, the recommended repair scheme is selected through Pareto multi-objective optimization, and the optimization objectives include maximizing the recovery rate of remaining bearing capacity, minimizing the repair intervention volume, and minimizing the construction period; the Pareto front is solved using a multi-objective evolutionary algorithm, and the solution closest to the utopia point is selected from the Pareto front as the recommended repair scheme, wherein the utopia point is composed of the optimal values ​​obtained by optimizing each optimization objective individually.

[0221] Based on the aforementioned mechanical simulation of the damaged finite element model, a specific implementation method is provided for automatically selecting recommended repair schemes through Pareto multi-objective optimization. This method simulates multiple different virtual repair operations for the same damaged component, using the remaining bearing capacity recovery rate, repair intervention volume, and construction period as optimization objectives. A multi-objective evolutionary algorithm is used to solve for the Pareto front, and the optimal compromise repair scheme is automatically output based on the principle of minimizing the distance to the utopian point. The specific steps include:

[0222] For the established damaged finite element model, structural engineers or cultural relic conservation experts will design at least two different virtual repair operations based on the type, location, and depth of the damage. These repair operations include, but are not limited to: surface grouting (sealing only the surface of the crack), pressure grouting (injecting structural adhesive into the crack), wood strip patching (removing rotten parts and embedding the same type of wood, then bonding it), fiber-reinforced composite material wrapping (attaching carbon fiber or fiberglass cloth to the outside of the component), adding iron hoops or steel tie rods, and partial replacement of wooden components.

[0223] For each repair operation, the structural state after repair is simulated in the finite element model by modifying the properties of the damaged elements (such as restoring part of the elastic modulus, changing boundary conditions, and adding reinforcement layer elements). For example, for pressure injection, the elastic modulus of the element through which the crack passes can be restored from the damage reduction value to 80% to 90% of the undamaged value; for adding iron hoops, shell elements or beam elements are added around the component to simulate the constraint effect of the hoops.

[0224] Apply the same loads (self-weight, wind load, seismic action, etc.) as the unrepaired state to the finite element model after each virtual repair operation, perform static or dynamic analysis, and calculate the following output results:

[0225] The remaining load-bearing capacity of the repaired structure is the ultimate load value obtained by gradually loading the structure until it fails or deforms out of control. This value is compared with the original design load-bearing capacity of the undamaged component to obtain the remaining load-bearing capacity recovery rate (for example, if the original load-bearing capacity is 100kN and it is 85kN after repair, then the recovery rate is 85%).

[0226] The stress redistribution cloud map of the repaired structure focuses on whether the stress peak value in and around the original damaged area exceeds the material strength and whether there is new stress concentration.

[0227] At the same time, two engineering indicators were extracted from the construction plan:

[0228] Restoration intervention volume: refers to the volume of materials involved in the restoration operation or the volume of wood that needs to be removed or replaced. For glue injection, the intervention volume is the volume of the injected glue; for patching, it is the volume of the removed decayed parts; for wrapping, it is the surface area covered by the fiber cloth multiplied by the thickness. The smaller the intervention volume, the less impact it has on the authenticity of the artifact.

[0229] Construction period: The number of working days required from site preparation to completion of repair, including material preparation, construction operations, and maintenance time.

[0230] The above three virtual repair operations are considered as candidate solutions (if the number of solutions is small, more solutions can be generated by changing parameters, such as changing the injection depth, the number of fiber wrapping layers, etc., to form 10-20 candidate solutions). Three optimization objectives are defined:

[0231] Objective 1 (maximization): Remaining bearing capacity recovery rate, denoted as R, with a value range of 0~100%, and the larger the value, the better.

[0232] Objective 2 (Minimization): Repair intervention volume, denoted as V, usually in cubic centimeters or cubic meters, the smaller the value, the better.

[0233] Objective 3 (Minimization): Construction period, denoted as T, usually in days, the smaller the value, the better.

[0234] Since the three objectives mentioned above are conflicting (for example, improving the load-bearing capacity recovery rate often requires more intervention volume and a longer construction period), there is no single solution that simultaneously optimizes all objectives. Therefore, a multi-objective optimization method is used to solve for the Pareto optimal solution set (i.e., the Pareto front).

[0235] Candidate schemes are non-dominated by using multi-objective evolutionary algorithms (such as the non-dominated sorting genetic algorithm NSGA-II, the decomposition-based multi-objective evolutionary algorithm MOEA / D, or the intensity Pareto evolutionary algorithm SPEA2).

[0236] The specific operating steps are as follows:

[0237] Each virtual repair scheme is encoded as an individual in the algorithm, and each individual contains the repair type and key parameters (such as injection depth, patch length, number of fiber layers, etc.).

[0238] Set the population size (e.g., 50 individuals) and the number of generations (e.g., 100 generations), and generate a new generation of schemes through selection, crossover, and mutation operations.

[0239] The individuals in the population are non-dominated and ordered according to the three objective function values, and the crowding distance is calculated.

[0240] After multiple generations of evolution, a set of non-dominant Pareto optimal solutions is obtained. Within this set, it is impossible to increase one objective value without decreasing at least one objective value.

[0241] To automatically select a recommended repair solution from the Pareto front, the concept of a utopian point is introduced. A utopian point is a virtual point formed by optimizing each objective individually to achieve its optimal value.

[0242] The three optimization objectives are optimized separately: First, without considering the intervention volume and construction period, the optimal value of the objective is obtained by maximizing the recovery rate of the remaining bearing capacity; second, without considering the bearing capacity and construction period, the optimal value of the objective is obtained by minimizing the repair intervention volume; and finally, without considering the bearing capacity and intervention volume, the optimal value of the objective is obtained by minimizing the construction period. These three optimal values ​​together constitute an ideal reference point, called the Utopia Point.

[0243] For each feasible remediation scheme on the Pareto front (i.e., the scheme where no objective can be improved without reducing other objectives), calculate the differences between its remaining bearing capacity recovery rate and the optimal bearing capacity value at the Utopia point, the difference between its remediation intervention volume and the optimal volume value at the Utopia point, and the difference between its construction period and the optimal construction period value at the Utopia point. Sum the squares of these three differences and take the square root to obtain the comprehensive distance between the scheme and the Utopia point; select the scheme with the smallest comprehensive distance as the recommended remediation scheme.

[0244] If the dimensions of different targets differ significantly (for example, the intervention volume may be 10³ cm³, and the construction period may only be a few tens of days), it is necessary to normalize each target before calculating the distance. Normalization can be achieved using the extreme value method: mapping the maximum and minimum values ​​of each target in the Pareto front to the interval [0,1] respectively.

[0245] The final structural safety assessment report includes:

[0246] The percentage of remaining load-bearing capacity and stress redistribution cloud map corresponding to each virtual repair operation;

[0247] Pareto front plot (3D scatter plot or 2D projection plot) shows the advantages and disadvantages of each option;

[0248] The automatically recommended repair plan includes repair type, key parameters, expected remaining bearing capacity recovery rate, intervention volume, and construction period.

[0249] The effects of the above technical solution are as follows:

[0250] There is an inherent contradiction among the remaining bearing capacity recovery rate, the volume of repair intervention, and the construction period. Traditional methods rely on engineering experience for subjective trade-offs, while this method uses Pareto optimization to intuitively present the non-dominant relationships of all feasible solutions, avoiding arbitrariness and one-sidedness in decision-making.

[0251] By introducing the principle of minimum distance to utopian points, recommended solutions can be automatically selected without requiring decision-makers to provide additional preference information, making it easy to apply quickly to cultural relic protection projects. Furthermore, this decision-making rule has a clear geometric meaning, making it easy to understand and accept.

[0252] This embodiment uses NSGA-II as an example for illustration, but in actual implementation, any multi-objective evolutionary algorithm capable of solving the Pareto front (such as MOEA / D, SPEA2, NSGA-III, etc.) can be used. Technicians can freely choose according to the model size and computing resources without affecting the scope of protection.

[0253] The output Pareto front plot and the index values ​​of each scheme make the advantages and disadvantages of different restoration schemes clear at a glance, providing a transparent and reproducible basis for expert review and cultural relic protection decisions.

[0254] The simulation of virtual repair operations is based on the secondary development function of conventional finite element software (such as ANSYS and ABAQUS). The intervention volume and construction period data can be obtained from construction quotas or historical projects. The algorithm has a small computational load and is suitable for promotion and application in actual protection projects.

[0255] In one possible implementation, the method further includes a closed-loop feedback step:

[0256] During and after the actual repair work, deformation and stress data of the repaired area are collected by on-site sensors.

[0257] The collected measured deformation data is compared with the deformation data predicted by the finite element model at the same measuring point, and the displacement residual vector is calculated. Each component of the displacement residual vector is the measured displacement value minus the predicted displacement value at the same measuring point in the same direction.

[0258] The collected measured stress data is compared with the predicted stress data, and the stress residual vector is calculated.

[0259] When the magnitude of the displacement residual vector or the magnitude of the stress residual vector exceeds a preset threshold, a model update is triggered: using the Bayesian parameter inversion method, with measured data as constraints, the damage parameters and material constitutive parameters in the finite element model are updated.

[0260] The updated parameters are stored in the knowledge base as prior information for subsequent evaluation tasks of similar ancient wooden structures.

[0261] A closed-loop feedback method is used to reverse-correct the finite element model using measured data collected by on-site sensors during and after the actual restoration work. By comparing the differences between the measured and predicted values, the Bayesian inversion update of the model parameters is triggered, and the updated parameters are stored in a knowledge base for subsequent evaluation tasks of similar ancient wooden structures. The specific steps include:

[0262] Before the repair work begins, various types of field sensors are deployed on the damaged component and its surrounding critical areas. Sensor types include, but are not limited to:

[0263] Displacement sensors: used to measure the deformation of components during the repair process, such as crack opening displacement, beam mid-span deflection, and column top lateral displacement. Linear variable differential transformer displacement gauges or laser displacement gauges can be used.

[0264] Strain sensor: Used to measure linear strain on the surface of wood. It can be a resistance strain gauge or a fiber optic strain sensor, arranged along the grain and across the grain respectively.

[0265] Stress sensors: used to indirectly estimate the stress state inside a component. They can be obtained by using a vibrating wire rebar stress gauge (embedded type) or by multiplying the strain by the elastic modulus.

[0266] Temperature and humidity sensors: used to monitor environmental conditions and serve as auxiliary parameters to correct measurement data.

[0267] The sensor measurement points should correspond to the nodes or element integration points used to output predicted values ​​in the finite element analysis model. The measurement point layout plan should be designed before construction, and the same virtual measurement point numbers should be established in the finite element model.

[0268] Data collection is conducted throughout the entire repair process and continues for a period of time (e.g., 30 days) after repair completion to monitor the stability of the repair effect. The collection frequency is adjusted according to the construction stage: during intensive construction stages (such as grouting, patching, and reinforcement), it can be set to once per minute; during the curing and post-construction monitoring stages, it can be set to once per hour or once per day.

[0269] In the finite element model, predicted deformation and stress values ​​are extracted from the same locations and directions as those from the field sensors. These predicted values ​​are theoretical values ​​calculated based on initial damage parameters and the material constitutive model before repair.

[0270] Subtract the predicted value from the measured value in the same direction at the same measuring point to obtain the displacement residual (for deformation data) or stress residual (for stress data) at that point; the displacement residuals at all measuring points constitute a multidimensional vector (displacement residual vector), with each component corresponding to the difference in one direction at a measuring point; similarly, all stress residuals constitute a stress residual vector.

[0271] Pre-set threshold values ​​for the magnitudes of the displacement residual vector and the stress residual vector. The criteria for setting these threshold values ​​include:

[0272] The measurement accuracy of the sensor (for example, if the accuracy of a displacement sensor is 0.01mm, then the threshold can be set to 0.02~0.05mm).

[0273] Permissible errors in engineering (e.g., the permissible deviations specified in the standard for controlling deformation of ancient wooden structures);

[0274] Historical experience values ​​or determined through trial calculations.

[0275] When the magnitude of the calculated displacement residual vector or the magnitude of the stress residual vector exceeds the corresponding preset threshold, it indicates that there is a significant deviation between the damage parameters or material constitutive parameters of the current finite element model and the actual structural behavior, and the model update process needs to be initiated.

[0276] If the residual modulus values ​​are all within the threshold, it means that the model prediction accuracy meets the requirements, no update is needed, and normal monitoring can continue.

[0277] Once a model update is triggered, the uncertain parameters in the finite element model are corrected using the Bayesian parameter inversion method. The parameters to be updated include:

[0278] Damage parameters: such as the actual reduction factor of each damaged element, damage depth level, penetration degree, etc.

[0279] Constitutive parameters of materials: such as the modulus of elasticity along the grain, the modulus of elasticity in the radial and tangential directions, the shear modulus, Poisson's ratio, and strength parameters of wood.

[0280] The basic idea of ​​Bayesian parameter inversion is to combine prior information (i.e., the parameter values ​​and their uncertainty range used in the initial modeling) with field measured data, and calculate the posterior probability distribution of the parameters using Bayes' formula, thereby obtaining the parameter values ​​that are most likely to conform to the actual structure.

[0281] In practice, the following procedures can be followed:

[0282] Determine the prior distribution: Based on the type of wood, historical experience, or laboratory test data, set a prior probability distribution (e.g., normal or uniform distribution) for each parameter to be updated, and give its mean and variance.

[0283] Constructing the likelihood function: The predicted output of the finite element model (such as the displacement of a certain measuring point) is expressed as a function of the parameters to be updated. The difference between the measured data and the predicted value follows a Gaussian distribution, and the likelihood function is established accordingly.

[0284] Sampling and posterior estimation: Markov chain Monte Carlo sampling methods (such as the Metropolis-Hastings algorithm or Hamiltonian Monte Carlo) are used to draw samples from the posterior distribution, and the posterior mean, median or maximum posterior probability estimate of the parameters is obtained through statistics.

[0285] Update the finite element model: Reassign the obtained posterior parameter values ​​to the corresponding attributes in the finite element model, perform the mechanical simulation again, and compare the residuals between the predicted and measured values ​​once more. If the residuals still exceed the limits, iterate the inversion steps until convergence.

[0286] To reduce computational costs, approximate Bayesian computational methods or surrogate models based on neural networks can be used in practical engineering to replace complete finite element reanalysis.

[0287] After the model update is complete, the following information will be stored in the knowledge base:

[0288] Basic information about ancient buildings (name, era, structural type, type of wood);

[0289] Location, geometry, initial depth level, and final corrected depth level of the damaged area;

[0290] Updated material constitutive parameters (such as measured inversion values ​​of parallel elastic modulus);

[0291] Summary of the repair plan and measured data during the construction process;

[0292] Comparison of residuals and convergence before and after model update.

[0293] The knowledge base can be built using relational databases or graph databases, and supports retrieval based on conditions such as wood species, damage type, and geographical region.

[0294] When conducting subsequent damage assessments on similar ancient wooden structures, prior parameters for similar cases (e.g., the same type of wood, similar damage patterns) can be retrieved from the knowledge base and used as initial values ​​for finite element modeling of the new project. This can significantly improve the initial accuracy of the new model, reduce the number of inversion iterations, and even allow direct use of parameters from the library without recalibration.

[0295] The effects of the above technical solution are as follows:

[0296] By triggering Bayesian inversion with residual thresholds, the finite element model can automatically adjust its parameters according to the actual structural response, continuously approximating the true state, and the reliability of the evaluation results improves with data accumulation.

[0297] Reduce reliance on prior parameters; even if the constitutive parameters of the wood used in the initial model come from manuals or empirical values ​​and deviate from the actual wood in the field, they can be quickly corrected through inversion of field monitoring data, avoiding errors in the evaluation conclusions due to inaccurate parameters.

[0298] The measured data and inversion parameters from each restoration project are systematically stored. As the number of cases increases, the knowledge base gradually covers various types of wood, various types of damage, and various restoration techniques. Subsequent projects can directly reuse these data, significantly reducing the workload of preliminary modeling and parameter calibration.

[0299] Supports long-term health monitoring and early warning. After repair, the sensors can continue to operate (e.g., collecting data monthly) to continuously compare measured values ​​with model predictions. If a residual suddenly increases, it indicates that the structure may have suffered new damage or that the original repair has failed. The system can automatically issue an early warning signal to achieve early detection of damage.

[0300] In one possible implementation, the depth level correction is further combined with a pre-trained machine learning agent model:

[0301] The textural features of surface image data, the local curvature features of three-dimensional geometric data, and the abnormal region volume of internal defect detection data are used as input features;

[0302] Using the depth level labels of on-site measurements or historical damage cases as output, train random forest or extreme gradient boosting models;

[0303] During online evaluation, the surrogate model outputs a depth level prediction value in real time and performs a weighted fusion with the result of the regression model. The weighting coefficient is adaptively adjusted according to the quality of the input data.

[0304] Building upon the aforementioned depth level determination method based on visual width features, spatial penetration features, and confidence correction of internal defect detection data, a pre-trained machine learning surrogate model is further introduced to output real-time depth level predictions. This prediction is then weighted and fused with the results of the aforementioned regression model (i.e., a regression model based on visual width and spatial penetration features) to improve the response speed and robustness of depth level determination. Furthermore, the fusion weights can be adaptively adjusted according to the input data quality, ensuring relatively reliable evaluation results even in cases of missing data or sensor anomalies. The specific steps include:

[0305] The input features of the surrogate model come from three types of perceptual data:

[0306] Texture features of surface image data: Extracted from images acquired by high-resolution digital cameras. Specifically, after grayscale conversion and histogram equalization of local images of damaged areas (such as around cracks), the gray-level co-occurrence matrix is ​​calculated, and texture feature parameters such as contrast, correlation, energy, and homogeneity are extracted. These parameters can reflect the texture changes on the damaged surface caused by wood fiber breakage, decay, or surface cracks, and can serve as apparent features to distinguish different depth levels.

[0307] Local curvature features of 3D geometric data: extracted from 3D point cloud data. For the point cloud within the damaged region and its neighborhood, an algorithm based on neighborhood principal component analysis is used to estimate the normal vector of each point, and then the Gaussian curvature and mean curvature are calculated. The distribution characteristics of statistical curvature (such as mean curvature, maximum curvature, skewness, etc.) are used as a quantitative indicator of the geometric complexity of the damaged surface. Generally, the greater the depth and the more severe the damage, the more drastic the changes in local curvature.

[0308] The volume of anomalous regions in internal defect detection data is extracted from infrared thermal imaging or ultrasonic testing data. First, temperature anomalies (infrared thermal imaging) or echo anomalies (ultrasonic imaging) are identified through image segmentation or thresholding. Then, combined with spatial calibration and 3D reconstruction, the total volume of these internal anomaly regions is estimated. Larger anomaly volumes indicate larger internal cavities, delamination, or cracks, and a stronger need for upward correction at the depth level.

[0309] The above three types of features need to be normalized after extraction to eliminate the influence of different units and ensure that each feature makes a balanced contribution to model training.

[0310] Construct a training dataset in which each sample contains three components: first, the three types of features (texture features, curvature features, and abnormal volume) extracted from a historical damage case; second, the true depth level label (e.g., Level I, Level II, Level III) determined by the case through on-site measurement (such as excavation verification, core sampling, or 3D X-ray scanning) or expert evaluation.

[0311] Collect a sufficient number of samples (e.g., 500-1000) covering different wood species, damage types, and depth levels. Then train the model using either a random forest algorithm or an extreme gradient boosting algorithm. Both algorithms are ensemble learning models, exhibiting good resistance to overfitting and fitting of nonlinear relationships, and can output a ranking of feature importance.

[0312] During training, the samples are randomly divided into training and validation sets in a 7:3 ratio. Cross-validation is used to adjust hyperparameters such as the number of trees, maximum depth, and learning rate. The classification accuracy or mean absolute error on the validation set is used as the model performance index.

[0313] For the damaged area of ​​the wooden ancient building to be assessed, the texture features, local curvature features, and internal abnormal area volume of the area are extracted in real time and input into the pre-trained surrogate model. The surrogate model outputs a depth level prediction value (e.g., predicted as Level II, along with the confidence probability of the prediction).

[0314] Since the input features of the surrogate model can be automatically calculated directly from imagery, point cloud, and interior detection data without the need for manual annotation or complex regression model fitting, its prediction speed is very fast (usually in the millisecond range), making it suitable for scenarios that require rapid screening of a large number of damaged areas.

[0315] The regression model in the aforementioned embodiments was trained based on visual width and spatial penetration features, and its output is the initial depth level. This step weightedly fuses the output of the regression model with the output of the surrogate model to obtain the final depth level.

[0316] The core of weighted fusion is to dynamically determine the weight coefficients of the two models, allowing the weights to adaptively adjust based on the quality of the current input data. The specific rules are as follows:

[0317] Evaluating the data quality of the regression model: The regression model relies on surface imagery data and 3D geometric data. If the imagery data suffers from severe illumination inconsistencies, blurring, or occlusion (e.g., taken in shadow or confined spaces), or if the point cloud data contains large areas of holes (due to scanning blind spots or low wood reflectivity), the reliability of the regression model will decrease. Objective indicators such as the imagery's signal-to-noise ratio, point cloud density, and effective coverage ratio can be used to quantify data quality; higher quality data receives greater weight.

[0318] Evaluate the data quality of the surrogate model: The surrogate model relies on texture features, curvature features, and anomalous region volumes. The reliability of the surrogate model is reduced if internal defect detection data is missing (e.g., no infrared or ultrasonic scanning was performed), or if texture features are distorted due to surface paint or dust, or if curvature calculations are unstable due to sparse point clouds.

[0319] Adaptive weight allocation principle: Set the sum of the weights of the regression model and the surrogate model to 1. Initially, each can be set to 0.5. When the data quality index of the regression model is higher than the preset threshold (e.g., point cloud density greater than 100 points / square centimeter), the regression model weight is appropriately increased, up to a maximum of 0.9; conversely, if the data quality of the regression model is severely insufficient (e.g., point cloud holes exceed 30%), the regression model weight is decreased, down to a minimum of 0.1. Similarly, the surrogate model weight is dynamically adjusted based on the completeness of each input feature.

[0320] The final depth level is determined by weighted majority voting or weighted average. For level classification problems (levels I, II, and III), the levels output by the two models can be mapped to numerical values ​​(I=1, II=2, III=3), the weighted average is calculated, and then rounded to the nearest integer to obtain the final level. For example, if the regression model outputs level II (2) and the surrogate model outputs level III (3), and the weighted average is 2.3, then level II is taken.

[0321] As repair work and monitoring data accumulate, newly acquired damage cases (containing complete three types of features and the true depth level determined through field verification) are periodically added to the training dataset, or after each closed-loop feedback update, to incrementally train or periodically retrain the surrogate model, continuously improving its predictive accuracy. Simultaneously, the parameters of the regression model can be fine-tuned based on the consistency after fusion.

[0322] The effects of the above technical solution are as follows:

[0323] Once the surrogate model is trained, online prediction takes only milliseconds, which is several orders of magnitude faster than the point-by-point calculation of traditional finite element inversion or regression models. It is particularly suitable for rapid damage screening of hundreds of components in large ancient buildings.

[0324] When the quality of a certain type of data is poor (e.g., missing data for internal defect detection), adaptive weighted fusion can automatically reduce its reliance on that data source and rely more on another model, thus avoiding evaluation failure due to missing data. This fault tolerance is particularly important in practical engineering.

[0325] Make full use of multimodal information. The regression model mainly utilizes visual width and spatial penetration features (2D + 3D geometry), while the surrogate model introduces complementary features such as texture, curvature, and anomalous volume. The fusion of the two has higher information coverage and discrimination ability than a single model, and can reduce false positives and false negatives.

[0326] Example 2: This example provides a three-dimensional modeling-based evaluation system for ancient wooden structures, used in the three-dimensional modeling-based evaluation method for ancient wooden structures described in Example 1. The system includes:

[0327] The perception acquisition module is used to acquire multi-source perception data of the ancient wooden structure, including surface image data, three-dimensional geometric data and internal defect detection data.

[0328] The damage identification and parameter acquisition module is used to identify the damaged areas on the surface of the wood structure and extract damage parameters based on the surface image data and three-dimensional geometric data. The damage parameters include the location, geometric shape, and depth level of the damage. The depth level is corrected for confidence by fusing internal defect detection data.

[0329] The model reconstruction module is used to dynamically adjust the reconstruction strategy of the three-dimensional reconstruction algorithm between the damaged area and the non-damaged area according to the damage parameters, so as to generate a three-dimensional digital model in which the geometric details of the damaged area are preserved to a higher degree than those in the non-damaged area.

[0330] The simulation evaluation module is used to map the damage parameters as initial defect boundary conditions to the finite element analysis model, perform mechanical simulation on the damaged finite element model, and output the structural safety assessment results.

[0331] In one possible implementation, the damage identification and parameter acquisition module is further configured as follows:

[0332] Visual width features of damage are extracted based on surface image data, and spatial penetration features of damage are extracted based on three-dimensional geometric data; the visual width features and spatial penetration features are input into a pre-trained regression model to output an initial depth level;

[0333] Spatial registration is performed between internal defect detection data and surface image data or three-dimensional geometric data to calculate the spatial overlap between internal abnormal areas and surface damaged areas.

[0334] The initial depth level is adjusted according to the spatial overlap: when the overlap is greater than the first threshold, the initial depth level is increased by one or two levels; when the overlap is less than the second threshold, the initial depth level is decreased by one level; otherwise, it remains unchanged.

[0335] The internal defect detection data includes at least one of infrared thermal imaging data or ultrasonic echo data, and the first threshold and the second threshold are determined by the statistical distribution of historical damage samples.

[0336] In one possible implementation, the model reconstruction module is configured as follows:

[0337] Based on the location and depth level in the damage parameters, a three-dimensional damage influence field is generated. The damage influence field has continuously transitioning weights within the damage area, and the weights are positively correlated with the depth level.

[0338] The Poisson surface reconstruction algorithm using an octree structure dynamically adjusts the local octree depth based on the weights of the damage influence field: in regions where the weights are higher than the first weight threshold, the octree depth is set to the maximum depth value; in regions where the weights are lower than the second weight threshold, the octree depth is set to the minimum depth value; and in the intermediate region, the octree depth is determined by linear interpolation.

[0339] The maximum depth value is at least three levels greater than the minimum depth value, such that the mesh side length of the damaged area is less than half the mesh side length of the undamaged area.

[0340] In a possible implementation, the model reconstruction module is further configured to:

[0341] When generating the three-dimensional damage influence field, a morphological dilation operation is performed on the initial damage area to obtain a damage transition zone, wherein the dilation radius is proportional to the depth grade; in the damage transition zone, the weight value smoothly decays from the maximum value in the damage core area to zero, and a Gaussian kernel or an exponential kernel is adopted as the attenuation function;

[0342] and dynamically adjusting the octree depth according to the constraint of the spatial curvature of the damage area: when the curvature of the damage area exceeds a preset curvature threshold, forcibly increasing the octree depth of the corresponding area to the maximum depth value.

[0343] The simulation evaluation module is configured to:

[0344] determining the finite element mesh cells penetrated by the damage according to the geometric morphology and depth grade of the damage;

[0345] performing non-isotropic stiffness reduction on the penetrated cells, wherein the reduced elasticity matrix is expressed as an undamaged elasticity matrix minus a stiffness degradation matrix caused by damage; the stiffness degradation matrix adopts a wood orthogonal anisotropic constitutive model, the diagonal elements of which respectively correspond to the elastic modulus reduction coefficients in the grain direction, radial direction and chord direction, and the non-diagonal elements are reduced in proportion;

[0346] wherein the elastic modulus reduction coefficient is jointly determined according to the included angle between the damage trend and the wood fiber direction and the included angle between the damage trend and the main stress direction of the component.

[0347] In a possible implementation, the elastic modulus reduction coefficient is determined in the following manner:

[0348] defining a damage normal direction and a damage tangent direction, and defining a main stress direction and a wood fiber direction;

[0349] calculating a first included angle, which is the included angle between the damage trend and the main stress direction; calculating a second included angle, which is the included angle between the damage trend and the wood fiber direction;

[0350] the elastic modulus reduction coefficient in the grain direction is calculated by the following formula:

[0351] Dg=1-(w1×sin²θ1+w2×cosθ2)

[0352] wherein Dg is the elastic modulus reduction coefficient in the grain direction; θ1 is the first included angle; θ2 is the second included angle; w1 and w2 are weight coefficients, w1+w2=1, 0<w1<1, 0<w2<1 and w1>w2;

[0353] the elastic modulus reduction coefficients in the radial direction and the chord direction are respectively calculated by the following formulas:

[0354] Dr=1-(u1×sin²θ1+u2×cosθ2)

[0355] Dt=1-(v1×sin²θ1+v2×cosθ2)

[0356] wherein u1 and u2 are radial direction weight coefficients, v1 and v2 are chordwise direction weight coefficients, which respectively satisfy 0<u2<u1<1, u1+u2=1, and 0<v2<v1<1, v1+v2=1; and satisfy u1<w1, v1<w1;

[0357] when the first included angle is less than or equal to a first angle threshold and the second included angle is less than or equal to a second angle threshold, multiply the calculated reduction coefficients in the parallel-to-grain direction, radial direction and chordwise direction all by a forced attenuation factor γ, wherein 0<γ<1.

[0358] In a possible implementation, the stiffness reduction is further subjected to layered non-uniform treatment according to the distribution of damage depth along the thickness of the member section, comprising:

[0359] respectively determining a set of finite element elements penetrated by damage in each partition, and calculating the elastic modulus reduction coefficients in the parallel-to-grain direction, radial direction and chordwise direction for the elements in each partition respectively;

[0360] for each of the surface layer area, the middle layer area and the deep layer area, after the elastic modulus reduction coefficients in the parallel-to-grain direction, radial direction and chordwise direction of the elements in the partition are respectively determined based on the included angle between the damage trend and the wood fiber direction and the included angle between the damage trend and the main stress direction of the member, they are uniformly multiplied by the layered correction factor corresponding to the partition; wherein the layered correction factor of the surface layer area is greater than 1, the layered correction factor of the deep layer area is between 0 and 1, and the layered correction factor of the middle layer area is equal to 1;

[0361] when the damage depth distribution curve of the elements in the middle layer area shows that the damage penetrates the entire section, forcibly set the elastic modulus reduction coefficient in the parallel-to-grain direction of the elements in the middle layer area to a preset lower limit value (e.g., 0.2), and trigger an emergency check of bearing capacity; the elastic modulus reduction coefficients in the radial direction and chordwise direction remain unchanged or are processed according to the same rule.

[0362] In a possible implementation, when the same finite element is penetrated by a plurality of damage areas together, the elastic modulus reduction coefficients in all directions are superposed according to the following rules:

[0363] respectively determining the elastic modulus reduction coefficient in the parallel-to-grain direction, radial elastic modulus reduction coefficient and chordwise elastic modulus reduction coefficient under the individual action of each damage;

[0364] For each of the longitudinal, radial, and chordal directions, the combined elastic modulus reduction factor in that direction is equal to the product of a factor minus the complement of each individual damage reduction factor in that direction, where the complement is a factor minus the individual damage reduction factor in that direction.

[0365] When at least one through-section damage exists among multiple damages, for each of the parallel, radial, and chordal directions, the combined elastic modulus reduction factor in that direction is corrected by multiplying the correction factor in that direction by the minimum value among all single-damage elastic modulus reduction factors in that direction; wherein, the correction factor in each direction is greater than 0 and less than 1, and the correction factor in the parallel direction is not greater than the minimum value among the correction factors in the radial and chordal directions.

[0366] In one possible implementation, the simulation evaluation module is configured as follows:

[0367] Simulate at least two different virtual repair operations on a damaged finite element model;

[0368] Calculate the remaining bearing capacity and stress redistribution cloud map after each repair operation;

[0369] The structural safety assessment results include the percentage of remaining load-bearing capacity corresponding to each repair operation and the recommended repair scheme;

[0370] The recommended repair scheme is selected through Pareto multi-objective optimization. The optimization objectives include maximizing the remaining bearing capacity recovery rate, minimizing the repair intervention volume, and minimizing the construction period. The Pareto front is solved using a multi-objective evolutionary algorithm, and the solution closest to the utopia point is selected from the Pareto front as the recommended repair scheme. The utopia point is composed of the optimal values ​​obtained by optimizing each optimization objective individually.

[0371] In one possible implementation, the system further includes a closed-loop feedback module, configured as follows:

[0372] During and after the actual repair work, deformation and stress data of the repaired area are collected by on-site sensors.

[0373] The collected measured deformation data is compared with the deformation data predicted by the finite element model at the same measuring point, and the displacement residual vector is calculated. Each component of the displacement residual vector is the measured displacement value minus the predicted displacement value at the same measuring point in the same direction.

[0374] The collected measured stress data is compared with the predicted stress data, and the stress residual vector is calculated.

[0375] When the magnitude of the displacement residual vector or the magnitude of the stress residual vector exceeds a preset threshold, a model update is triggered: using the Bayesian parameter inversion method, with measured data as constraints, the damage parameters and material constitutive parameters in the finite element model are updated.

[0376] The updated parameters are stored in the knowledge base as prior information for subsequent evaluation tasks of similar ancient wooden structures.

[0377] In one possible implementation, the damage identification and parameter acquisition module is further configured as follows:

[0378] The depth level is corrected by combining a pre-trained machine learning agent model: the texture features of surface image data, the local curvature features of 3D geometric data, and the abnormal region volume of internal defect detection data are used as input features; the depth level labels of field measurements or historical damage cases are used as outputs to train random forest or extreme gradient boosting models.

[0379] During online evaluation, the surrogate model outputs a depth level prediction value in real time and performs a weighted fusion with the result of the regression model. The weighting coefficient is adaptively adjusted according to the quality of the input data.

[0380] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for evaluating the timber structure of ancient buildings based on three-dimensional modeling, characterized in that, The method includes: Acquire multi-source sensing data of ancient wooden structures, including surface image data, three-dimensional geometric data, and internal defect detection data; Based on the surface image data and three-dimensional geometric data, damaged areas on the surface of the wooden structure are identified and damage parameters are extracted. The damage parameters include the location, geometric shape, and depth level of the damage. The depth level is corrected for confidence by fusing internal defect detection data. Based on the damage parameters, the reconstruction strategy of the 3D reconstruction algorithm is dynamically adjusted between the damaged area and the undamaged area to generate a 3D digital model in which the geometric details of the damaged area are preserved to a higher degree than those of the undamaged area. The damage parameters are used as initial defect boundary conditions and mapped to the finite element analysis model. Mechanical simulation is then performed on the damaged finite element model to output the structural safety assessment results.

2. The method according to claim 1, characterized in that, The depth level is determined in the following way: Visual width features of damage are extracted based on surface image data, and spatial penetration features of damage are extracted based on three-dimensional geometric data; the visual width features and spatial penetration features are input into a pre-trained regression model to output an initial depth level; Spatial registration is performed between internal defect detection data and surface image data or three-dimensional geometric data to calculate the spatial overlap between internal abnormal areas and surface damaged areas. The initial depth level is adjusted according to the spatial overlap: when the overlap is greater than the first threshold, the initial depth level is increased by one or two levels; when the overlap is less than the second threshold, the initial depth level is decreased by one level; otherwise, it remains unchanged.

3. The method according to claim 1, characterized in that, The reconstruction strategy of the dynamically adjusted 3D reconstruction algorithm includes: Based on the location and depth level in the damage parameters, a three-dimensional damage influence field is generated. The damage influence field has continuously transitioning weights within the damage area, and the weights are positively correlated with the depth level. The Poisson surface reconstruction algorithm using an octree structure dynamically adjusts the local octree depth based on the weights of the damage influence field: in regions where the weights are higher than the first weight threshold, the octree depth is set to the maximum depth value; in regions where the weights are lower than the second weight threshold, the octree depth is set to the minimum depth value; and in the intermediate region, the octree depth is determined by linear interpolation. The maximum depth value is at least three levels greater than the minimum depth value, such that the mesh side length of the damaged area is less than half the mesh side length of the undamaged area.

4. The method according to claim 3, characterized in that, When generating the three-dimensional damage influence field, the method further includes: A morphological expansion operation is performed on the initial damaged area to obtain a damage transition zone, with the expansion radius being proportional to the depth level. Within the damage transition zone, the weights smoothly decay from the highest value in the damage core zone to zero, and the decay function uses a Gaussian kernel or an exponential kernel. The dynamic adjustment of the octree depth is also constrained by the spatial curvature of the damaged region: when the curvature of the damaged region exceeds a preset curvature threshold, the octree depth of the corresponding region is forcibly increased to the maximum depth value.

5. The method according to claim 1, characterized in that, The step of mapping the damage parameters as initial defect boundary conditions to the finite element analysis model includes: Based on the geometry and depth level of the damage, determine the finite element mesh elements that the damage passes through. Performing non-isotropic stiffness reduction on the penetrated unit, where the reduced elastic matrix is expressed as the undamaged elastic matrix minus the stiffness degradation matrix caused by damage; The stiffness degradation matrix adopts the wood orthotropic constitutive model, whose diagonal elements respectively correspond to the elastic modulus reduction coefficients in the grain direction, radial direction and chord direction, and the off-diagonal elements are reduced proportionally; The elastic modulus reduction coefficient is jointly determined according to the included angle between the damage trend and the wood fiber direction and the included angle between the damage trend and the principal stress direction of the component.

6. The method according to claim 5, characterized in that, The elastic modulus reduction coefficient is determined in the following manner: Defining the normal direction of damage and the tangential direction of damage, and defining the principal stress direction and the wood fiber direction; Calculating a first included angle and a second included angle, wherein the first included angle is the included angle between the damage trend and the principal stress direction; the second included angle is the included angle between the damage trend and the wood fiber direction; The elastic modulus reduction coefficient in the grain direction is calculated by the following formula: Dg=1-(w1×sin²θ1+w2×cosθ2); wherein, Dg is the elastic modulus reduction coefficient in the grain direction; θ1 is the first included angle; θ2 is the second included angle; w1 and w2 are weight coefficients, w1+w2=1, 0<w1<1, 0<w2<1 and w1>w2; The elastic modulus reduction coefficients in the radial direction and the chord direction are respectively calculated by the following formulas: Dr=1-(u1×sin²θ1+u2×cosθ2); Dt=1-(v1×sin²θ1+v2×cosθ2); wherein, u1 and u2 are weight coefficients for the radial direction, v1 and v2 are weight coefficients for the chord direction, which respectively satisfy 0<u2<u1<1, u1+u2=1, and 0<v2<v1<1, v1+v2=1; and satisfy u1<w1, v1<w1; When the first included angle is less than or equal to the first angle threshold and the second included angle is less than or equal to the second angle threshold, multiply the calculated reduction coefficients of the grain direction, radial direction and chord direction all by a forced attenuation factor γ, wherein 0<γ<1.

7. The method according to claim 1, characterized in that, The mechanical simulation includes: Simulating at least two different virtual repair operations respectively on the finite element model with damage; Calculating the residual bearing capacity and stress redistribution cloud map after each repair operation; The structural safety assessment result includes the residual bearing capacity percentage corresponding to each repair operation and a recommended repair scheme; wherein the recommended repair scheme is selected through Pareto multi-objective optimization, the optimization objectives include maximizing the residual bearing capacity recovery rate, minimizing the repair intervention volume and minimizing the construction period; a multi-objective evolutionary algorithm is used to solve the Pareto front, and the solution with the shortest distance to the utopia point is selected from the Pareto front as the recommended repair scheme, and the utopia point is composed of the optimal values obtained by individual optimization of each optimization objective.

8. The method according to claim 1, characterized in that, The method further comprises a closed-loop feedback step: During the actual repair construction process and after the repair is completed, collecting the deformation data and stress data of the repaired area through on-site sensors; Comparing the collected measured deformation data with the predicted deformation data of the finite element model at the same measuring points, and calculating a displacement residual vector, wherein each component in the displacement residual vector is the measured displacement value minus the predicted displacement value in the same direction at the same measuring point; The collected measured stress data is compared with the predicted stress data, and the stress residual vector is calculated. When the magnitude of the displacement residual vector or the magnitude of the stress residual vector exceeds a preset threshold, a model update is triggered: using the Bayesian parameter inversion method, with measured data as constraints, the damage parameters and material constitutive parameters in the finite element model are updated. The updated parameters are stored in the knowledge base as prior information for subsequent evaluation tasks of similar ancient wooden structures.

9. The method according to claim 1, characterized in that, The depth level correction also incorporates a pre-trained machine learning agent model: The textural features of surface image data, the local curvature features of three-dimensional geometric data, and the abnormal region volume of internal defect detection data are used as input features; Using the depth level labels of on-site measurements or historical damage cases as output, train random forest or extreme gradient boosting models; During online evaluation, the surrogate model outputs a depth level prediction value in real time and performs a weighted fusion with the result of the regression model. The weighting coefficient is adaptively adjusted according to the quality of the input data.

10. An evaluation system for ancient wooden structures based on 3D modeling, characterized in that, The system includes: The perception acquisition module is used to acquire multi-source perception data of the ancient wooden structure, including surface image data, three-dimensional geometric data and internal defect detection data. The damage identification and parameter acquisition module is used to identify the damaged areas on the surface of the wood structure and extract damage parameters based on the surface image data and three-dimensional geometric data. The damage parameters include the location, geometric shape, and depth level of the damage. The depth level is corrected for confidence by fusing internal defect detection data. The model reconstruction module is used to dynamically adjust the reconstruction strategy of the three-dimensional reconstruction algorithm between the damaged area and the non-damaged area according to the damage parameters, so as to generate a three-dimensional digital model in which the geometric details of the damaged area are preserved to a higher degree than those in the non-damaged area. The simulation evaluation module is used to map the damage parameters as initial defect boundary conditions to the finite element analysis model, perform mechanical simulation on the damaged finite element model, and output the structural safety assessment results.

Images