Concrete structure earthquake damage assessment method and system based on multi-source data feature fusion

Abstract

The application discloses a kind of concrete structure earthquake damage assessment method and system based on multi-source data feature fusion, belong to civil engineering structure earthquake damage assessment technical field, for the problem that single source data cannot accurately reflect concrete structure earthquake damage state, the present application extracts apparent damage features using computer vision, extracts geometric deformation features using laser radar, extracts overall stiffness features using acceleration sensor;By weighted fusion algorithm to construct holographic damage feature vector, and using the vector to automatically correct the constitutive parameters of nonlinear finite element model, finally the residual bearing capacity of the structure is obtained by numerical simulation calculation. It realizes the objective, rapid identification of post-earthquake structure damage, and can output specific structure safety reserve quantitative index.

Description

Technical Field

[0001] This invention relates to the field of seismic damage assessment technology for civil engineering structures. More specifically, this invention relates to a method and system for seismic damage assessment of concrete structures based on multi-source data feature fusion. Background Technology

[0002] After an earthquake, the rapid and accurate assessment of the damage to concrete structures (such as bridges, high-rise buildings, dams, etc.) is crucial for post-disaster emergency rescue and structural repair decisions. Currently, the following technical solutions are mainly used in structural seismic damage assessment at home and abroad: (1) Traditional assessment method based on manual visual inspection The most traditional assessment method mainly relies on professional engineers to carry tools such as tape measures and crack observation instruments to conduct visual inspections on site. This method is highly subjective and has a large degree of dispersion: the assessment results are highly dependent on the experience of the inspectors, and different people may have large deviations in their judgment of the same crack. Moreover, the structure is usually in an unstable state after the earthquake, and close contact with the human body poses serious safety hazards. In addition, large-scale inspection takes too long and cannot meet the timeliness of emergency response. Furthermore, this method is difficult to directly deduce the quantitative mechanical indicators such as the "residual bearing capacity" of the structure through visual observation. (2) Single image recognition technology based on computer vision (Deep Learning) With the development of deep learning, the use of convolutional neural networks (CNN) to identify cracks in structural surface images has become a hot topic. This method can only obtain the apparent two-dimensional features and cannot accurately measure the crack depth or the three-dimensional deformation of the component (such as shear displacement and tilting), resulting in the omission of severe structural damage (such as core concrete crushing). Moreover, it is extremely dependent on lighting conditions, and the dust, shadows and obstructions at the earthquake site can easily cause false alarms in the algorithm. At the same time, simply identifying "there is a crack" does not mean knowing "the structure will collapse", and the image cannot be directly associated with the stiffness degradation and modal changes of the structure. (3) Modal recognition technology based on vibration sensors (accelerometers) Using sensors to collect structural dynamic response data to invert the structural state is another mainstream method. However, structural dynamic response data such as frequency and other overall modal parameters mainly reflect the global stiffness of the structure, and are often insensitive to early and local concrete cracking or steel yielding, and there is a blind spot of "cannot be measured". At the same time, it is difficult to accurately locate the specific damage location (such as which column is damaged) based on sensor signals alone, and it is difficult to guide specific reinforcement and repair. (4) Simple multi-source data overlay technology In order to solve the defects of a single data source, some existing technologies attempt to combine image and sensor data. However, most existing multi-source solutions are simply listing or visually overlaying data, lacking a deep fusion mechanism for heterogeneous data at the algorithm level. For example, they fail to utilize geometric features extracted from images to correct the mechanical model of sensors. Most existing data-driven methods are "black box" models, failing to automatically map identified damage features to finite element models (FEMs) for mechanical verification, resulting in evaluation results lacking physical meaning (such as the inability to provide specific safety factors), making them difficult for the engineering community to fully trust. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for assessing seismic damage to concrete structures based on multi-source data feature fusion, which mainly solves the following technical problems: traditional manual inspection is highly subjective, inefficient, and cannot directly obtain mechanical indicators; single computer vision technology lacks three-dimensional depth information and is difficult to reflect the internal state; single vibration sensing technology is insensitive to local early damage and is difficult to locate; existing data fusion schemes lack physical model support and cannot perform mechanical quantitative verification.

[0004] To achieve these objectives and other advantages according to the present invention, a method for assessing seismic damage to concrete structures based on multi-source data feature fusion is provided, comprising the following steps:

[0005] S1. Collect multi-source heterogeneous data of the concrete structure to be evaluated, including image data, point cloud data and vibration data;

[0006] S2. Perform semantic segmentation on the image data to extract the apparent damage features of the concrete structure to be evaluated; extract the geometric deformation features of the concrete structure to be evaluated based on the point cloud data; extract the dynamic modal features of the concrete structure to be evaluated based on the vibration data.

[0007] S3. Weighted fusion of the apparent damage features, geometric deformation features and dynamic modal features to construct a comprehensive damage feature vector;

[0008] S4. Based on the comprehensive damage feature vector, the stiffness parameters of the preset parametric finite element model of the concrete structure to be evaluated are corrected, and the stiffness parameters of each component are optimized and corrected by a genetic algorithm.

[0009] S5. Apply pushover load or incremental dynamic analysis to the modified finite element model, calculate the ultimate bearing capacity of the concrete structure to be evaluated, and calculate the ratio with the required load of the concrete structure under the design earthquake to obtain the safety reserve index of the structure.

[0010] Preferably, S1 also includes a data preprocessing step, including: distortion correction of image data, denoising and registration of point cloud data, and bandpass filtering denoising of vibration data.

[0011] Preferably, S1 further includes a step of removing unstructured interference from the point cloud data, including:

[0012] A method combining temporal multi-frame difference and octree raster mapping is used to automatically remove point cloud noise from moving objects;

[0013] A semantic segmentation network is used to identify unstructured components in image data, and then the point cloud regions of the identified unstructured components are marked as invalid regions in the point cloud data.

[0014] Geometric fitting is performed only on the point cloud of structural components.

[0015] Preferably, in S2, the apparent damage characteristics include the crack density index D. img The calculation formula is as follows:

[0016]

[0017] in, Let be the average width of the i-th crack. l i Let A be the length of the i-th crack. surface The surface area of ​​the structural member where the crack is located;

[0018] Geometric deformation characteristics include the inter-story residual displacement angle of beam and column members. θ r The calculation formula is as follows:

[0019]

[0020] in, and These are the horizontal displacements of the top and bottom of the beam / column members relative to the reference plane, respectively, where H is the story height;

[0021] Dynamic modal characteristics include the first-order natural frequencies of the structure.

[0022] Preferably, in S3, a normalized weighted fusion algorithm is used to construct a comprehensive damage feature vector, and the weighting coefficients are dynamically adjusted based on the information entropy of each data source.

[0023] Preferably, the comprehensive damage feature vector is represented as follows:

[0024] in, The normalized crack density index, The normalized inter-layer residual displacement angle. and The values ​​are the measured first-order natural frequencies of the structure after the earthquake and the initial natural frequencies before the earthquake, respectively. , and The weighting coefficients correspond to the image, point cloud, and vibration data sources, respectively.

[0025] The weighting coefficients for each data source are calculated using the information entropy weighting method, specifically including:

[0026] Calculate the information entropy of each data source. H :

[0027]

[0028] For image data,x i This represents the grayscale distribution of the crack pixels; for point cloud data, x i To fit the residual distribution of the plane, for vibration data, x i The time history distribution of the vibration signal, p ( x i ) represents a feature x i The probability of occurrence;

[0029] For the t-th data source, its weight coefficient w t The calculation formula is:

[0030]

[0031] in, m Indicates the total number of data sources. H t Represents the information entropy of the t-th data source, 1- H t This represents the information redundancy of the t-th data source. This represents the sum of redundancy across all data sources.

[0032] Preferably, in S4, an objective function is established for the parametric finite element model of the concrete structure to be evaluated. J ( k The objective function is iteratively optimized using an optimization algorithm. The process stops when the error is less than a preset threshold, and the corrected stiffness parameters and the corresponding corrected finite element model are output. J ( k )for:

[0033]

[0034] in, k Here are the stiffness reduction factors for each structural component. λ The regularization coefficient is . This is the measured comprehensive damage feature vector. This is the comprehensive damage feature vector calculated from the current finite element model.

[0035] Preferably, S4 employs a substructure hierarchical correction strategy based on sensitivity analysis, including:

[0036] Calculate the sensitivity of the stiffness reduction factor of each structural component to the comprehensive damage characteristic vector;

[0037] Calculate the sensitivity of the stiffness reduction factor of each structural component to each feature in the comprehensive damage feature vector;

[0038] First, correct the stiffness reduction factor for dynamic modal features that are more sensitive to the first threshold but less sensitive to other features that are less sensitive to the second threshold. Then, correct the stiffness reduction factor for geometric deformation features and / or apparent damage features that are more sensitive to the second threshold.

[0039] Preferably, the safety reserve index calculated in S5 is:

[0040]

[0041] in, C residual The current ultimate bearing capacity of the structure, S demand for The required load under the design earthquake is taken from the elastoplastic shear force requirement in the seismic design code document;

[0042] The current ultimate bearing capacity of the structure is obtained by applying a pushover load or incremental dynamic analysis to the modified finite element model to obtain the base shear force-apex displacement curve of the structure, and then reading the current ultimate bearing capacity of the structure from the base shear force-apex displacement curve.

[0043] This invention also provides a seismic damage assessment system for concrete structures based on multi-source data feature fusion, comprising:

[0044] The data acquisition module is used to collect multi-source heterogeneous data of the concrete structure to be evaluated, including image data, point cloud data, and vibration data.

[0045] The feature extraction module is used to perform semantic segmentation on the image data to extract the apparent damage features of the concrete structure to be evaluated; extract the geometric deformation features of the concrete structure to be evaluated based on the point cloud data; and extract the dynamic modal features of the concrete structure to be evaluated based on the vibration data.

[0046] The feature fusion module is used to weight and fuse the apparent damage features, geometric deformation features and dynamic modal features to construct a comprehensive damage feature vector.

[0047] The model correction module is used to correct the stiffness parameters of the preset parametric finite element model of the concrete structure to be evaluated based on the comprehensive damage feature vector.

[0048] The bearing capacity assessment module is used to apply push-over loads or incremental dynamic analysis to the modified finite element model and calculate the safety reserve index of the concrete structure to be assessed.

[0049] The present invention has at least the following beneficial effects:

[0050] To address the problems of low efficiency and extremely high risks associated with manual in-home inspections, particularly in dire buildings experiencing ongoing aftershocks, this invention constructs a non-contact data acquisition front-end using drones, long-range laser scanners, and wireless sensor networks. Combined with automated feature extraction and model correction algorithms, it replaces the traditional process of manual sketching and calculation. Inspection personnel do not need to enter the damaged building, completely avoiding the risk of collapse and injury from aftershocks. The system significantly reduces the time required for detailed assessment of individual buildings (including data acquisition and calculation time) compared to traditional manual methods. Following a major regional earthquake, this system can quickly generate regional damage distribution maps, providing crucial intelligence support for resource allocation during the critical rescue period.

[0051] To address the limitations of single-sensor computer vision (lacking depth information, thus failing to distinguish between surface spalling and structural deformation) and single-sensor technology (failing to capture geometric features), this invention employs a heterogeneous data weighted fusion algorithm. This algorithm utilizes the interlayer residual displacement angle (θ) extracted from laser point clouds. r The system supplements the missing depth information in visual images; the frequency attenuation characteristics of the accelerometer are used to verify whether the visual cracks cause substantial stiffness degradation. Through point cloud geometric feature verification, the system can effectively distinguish between "non-structural damage" (such as only a large area of ​​decorative layer peeling off, but the column is not tilted) and "structural damage" (the column undergoes shear displacement), significantly reducing the false alarm rate compared to a single vision solution. Through a dynamic adjustment mechanism of weighting coefficients (α,β,γ), the system can still maintain basic evaluation functions based on vibration signals even at night or in smoky environments (image weights decrease, sensor weights increase), solving the pain point of traditional optical detection being severely limited by ambient light.

[0052] To address the limitations of existing single vibration sensing technologies, which can only acquire overall modal parameters and are insensitive to local damage and difficult to pinpoint precisely (i.e., damage at different locations may cause the same frequency drop), this invention innovatively introduces a visually recognized "crack density index (D)" into the objective function J(k). img The visual feature vector directly constrains the correction direction of specific elements in the stiffness matrix during the optimization algorithm iteration. This fusion strategy of "visual local constraint + vibration global calibration" reduces the solution space of the model correction inverse problem (depending on the complexity of the structure). This enables the system to accurately distinguish damage to components in symmetrical positions, improving the damage location accuracy from the traditional "story-level" to the specific "component-level", effectively avoiding the omission of severe local damage due to insignificant overall frequency changes.

[0053] To address the problem that existing technologies primarily rely on expert experience scoring or only identify surface cracks, failing to directly reflect the residual bearing capacity within the structure and thus lacking clear physical meaning in the assessment results, this invention utilizes a modified "digital twin" finite element model for nonlinear dynamic time history analysis or static pushover analysis. Unlike traditional methods that only output "moderate / severe damage" labels, this invention can calculate a specific Safety Reserve Index (SRI). For example, "the current ultimate inter-story drift reserve of the structure is 1 / 150" or "the residual shear bearing capacity is 75% of the design value." This quantitative data provides a direct and reliable basis for decision-making regarding whether "immediate demolition," "load-limited use," or "local reinforcement" is necessary after an earthquake, eliminating the problem of "large dispersion" (typically, the difference rate in manual assessments can reach 20%-30%), and ensuring the objectivity and consistency of the assessment results.

[0054] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0055] Figure 1 The flowchart is a process for assessing seismic damage to concrete structures based on multi-source data feature fusion as described in this invention.

[0056] Figure 2 System architecture diagram for implementing the concrete structure seismic damage assessment method described in this invention;

[0057] Figure 3 This is a logic block diagram of the extraction process of apparent damage features, geometric deformation features and dynamic modal features described in this invention;

[0058] Figure 4 This is a closed-loop control diagram of the model correction process described in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0060] like Figure 1 As shown, this invention provides a method for assessing seismic damage to concrete structures based on multi-source data feature fusion, comprising the following steps:

[0061] S1. Collect multi-source heterogeneous data of the concrete structure to be evaluated, including image data, point cloud data and vibration data;

[0062] Specifically, such as Figure 2 As shown, image data can be obtained using high-resolution industrial cameras or drones by acquiring multi-view RGB images of key structural components (columns, beams, nodes).

[0063] Point cloud data can be obtained using a terrestrial 3D laser scanner (LiDAR) to acquire high-precision point clouds of the structure. To ensure the accuracy of subsequent geometric deformation feature extraction, the acquired point cloud density should meet the minimum resolution requirement. Specifically, for overall tilt or large deformation measurements, an effective point cloud density of no less than 1000 points / m is recommended. 2 (The corresponding average point spacing is approximately 31 mm); if further point cloud-assisted identification of local peeling or subtle displacement is required, the preferred point cloud density should be no less than 10,000 points / m². 2 (The corresponding average point spacing is approximately 10mm). During actual data acquisition, adjustments can be made based on the performance parameters of the on-site equipment, as long as the surface of the structure to be evaluated is covered. In a low-cost approach, dense point clouds generated by oblique photography can be used instead of lidar point clouds (although the accuracy is slightly lower, the principle is the same).

[0064] Vibration data can be acceleration time-history signals, acquired by deploying triaxial accelerometers on the top floor and key floors of the structure, with a sampling frequency set above 100Hz, under post-earthquake environmental excitation. The sensors can also be replaced with fiber optic grating (FBG) sensors or dynamic vision sensors (DVS).

[0065] Even better, the image data, point cloud data, and vibration data can be preprocessed to improve the quality and usability of the collected multi-source heterogeneous data, laying the foundation for accurate feature extraction in the future. Preprocessing includes distortion correction of the image data, denoising and registration of the point cloud data, and bandpass filtering denoising of the vibration data.

[0066] For image data, distortion correction is necessary. Optical distortion in the lenses of drones or industrial cameras used to acquire images can distort the geometry of structures in the images, affecting the accuracy of measurements of apparent damage features such as crack length and width. Correction can be achieved using methods such as Zhang Zhengyou's calibration. This involves pre-capturing images of a calibration plate of known dimensions, calculating the camera's intrinsic parameter matrix and distortion coefficients, and then correcting the post-earthquake images of the structure to eliminate radial and tangential distortion, thus obtaining orthophotos with accurate geometric relationships.

[0067] For point cloud data, denoising and registration operations are required. Raw laser-scanned point clouds often contain discrete noise points introduced by environmental dust or scanner errors. Denoising can be achieved through methods such as statistical filtering, for example, calculating the average distance and standard deviation between each point and its neighbors, and discarding points whose distance exceeds a certain multiple of the standard deviation of the mean as outliers. Registration is the process of unifying local point cloud data obtained from multiple, multi-station scans into a single global coordinate system. Iterative nearest-point algorithms are typically used, continuously iterating to find the optimal spatial transformation matrix between point clouds to achieve precise alignment, thereby constructing a complete and seamless 3D point cloud model of the structure to be evaluated.

[0068] For vibration data, bandpass filtering is required for noise reduction. The raw vibration signals acquired by accelerometers contain not only useful components reflecting the structure's dynamic characteristics, but also high-frequency electronic noise, low-frequency wind loads, or ground micro-vibration interference. The purpose of bandpass filtering is to retain the frequency band containing the structure's main modal frequencies while filtering out noise outside this band. For example, based on the structure type, its fundamental frequency range can be estimated, and a bandpass filter covering a passband frequency range of 0.1Hz to 50Hz can be designed. Processing the raw acceleration time history signal through this filter can effectively suppress high-frequency noise and low-frequency drift, obtaining a relatively pure structural acceleration response signal, which facilitates subsequent accurate identification of the structure's first-order natural frequencies and other dynamic modal characteristics.

[0069] In practice, image distortion correction can rely on pre-calibrated camera parameter files to batch process on-site images; point cloud denoising and registration can be completed using professional point cloud processing software or by writing specific algorithm scripts; and vibration data filtering can be achieved using digital filter design tools in the signal processing toolbox. These preprocessing operations work synergistically to significantly improve the signal-to-noise ratio and geometric fidelity of images, point clouds, and vibration data, making the apparent damage features, geometric deformation features, and dynamic modal features extracted from these data sources more reliable. This ensures the accuracy of subsequent feature fusion and model correction, and is a crucial preliminary step for the entire evaluation method to obtain objective and reliable results.

[0070] Even better, the point cloud data collected at the earthquake site inevitably contains a large amount of non-structural interference in addition to information about the target concrete structure itself. This interference mainly comes from two categories: first, moving objects at the site, such as rescue workers, vehicles, and flying debris; and second, non-load-bearing components attached to or near the structure, such as temporary scaffolding, hanging billboards, decorative curtain walls, and accumulated rubble. If these interfering point clouds are not removed, they will introduce serious errors in subsequent geometric deformation feature extraction. For example, points from moving trucks may be misidentified as local structural bulges, or points from suspended objects may be included in the planar fitting calculation of beams and columns, leading to distortion of key indicators such as residual displacement angles. Therefore, implementing a dedicated non-structural interference removal step is crucial.

[0071] To address dynamic point cloud noise generated by moving objects (such as rescue workers, vehicles, and dust), this method employs a combination of temporal multi-frame differencing and octree raster mapping for automatic noise removal. The specific implementation steps are as follows:

[0072] Step A: Multi-frame temporal point cloud acquisition

[0073] During structural scanning, keep the 3D laser scanner (LiDAR) station station fixed and enable "continuous scanning mode" or "time series mode". Set the acquisition duration to 30 to 60 seconds and the sampling frequency to 0.5 Hz to 1 Hz (i.e., acquire one complete low-resolution point cloud or local point cloud frame every 1-2 seconds), thereby obtaining a temporal point cloud set {P1, P2, ..., P...} containing N frames (N≥5) of the same scene. N Within this short period, the probability of displacement of the main structure (such as beams and columns) is extremely low, and it can be regarded as a static background, while personnel and machinery are the dynamic foreground.

[0074] Step B: Octree Space Partitioning and Parameter Setting

[0075] Construct an axis-aligned bounding box (AABB) that encloses the entire scanned scene, and establish an octree index structure. Key parameter settings are as follows: Set the minimum leaf node resolution (voxel side length) of the octree to [value missing]. d leaf In this embodiment, it is recommended to... d leaf The resolution is set between 0.05m and 0.15m. This size is slightly larger than the width of a typical human limb or the diameter of a small drone, but much smaller than the size of major structural components (such as columns, which are typically wider than 0.4m). If the resolution is too high (e.g., 1mm), minute vibrations and noises will be misidentified as moving objects; if the resolution is too low (e.g., 0.5m), noise from people close to the wall cannot be accurately filtered out.

[0076] Step C: Voxel status determination and dynamic elimination

[0077] Each frame of point cloud P t Mapping to an octree grid, calculate the occupancy status of each leaf node voxel.

[0078] 1. State definition: For the i-th voxel V i If in the first t If a laser dot falls within the voxel in the frame, record the state S. i,t =1 (occupied); otherwise S i,t = 0 (idle).

[0079] 2. Change detection criterion: Traverse all voxels and calculate their occupancy probability in the N-frame time series. , V i Let i represent the i-th voxel in the octree.

[0080] 3. Classification threshold: T noise To determine the threshold for point cloud noise of moving objects; T static The threshold for determining static backgrounds;

[0081] like Prob ( V i )> T static The voxel is determined to be a static structure (always existing, such as a wall).

[0082] like T noise < Prob ( V i )< T static (For example T static =0.9, T noise =0.1), the voxel is determined to be dynamic noise (sometimes present, such as a person walking), and all point clouds within the voxel are marked as noise and removed;

[0083] like Prob ( V i )< T noise These are considered random measurement errors or airborne dust and are therefore discarded. T static It can be set to 0.9. T noiseIt can be set to 0.1, but in other cases, it can be set according to specific needs.

[0084] The above process effectively filters out transient noise introduced by on-site activities.

[0085] For non-structural components fixed on-site but not belonging to the load-bearing system, this method utilizes a semantic segmentation network for identification and removal. First, a deep learning semantic segmentation model trained on a large number of architectural images is used to analyze synchronously acquired high-resolution structural images. This model can identify and segment various objects in the images at the pixel level, such as classifying pixels into "concrete beams," "concrete columns," "brick walls," "glass windows," "temporary fencing," and "mechanical equipment." After identifying non-structural components such as "temporary fencing" and "mechanical equipment," the regions containing these non-structural components in the two-dimensional images are mapped to a three-dimensional point cloud space based on the pre-calibrated coordinate transformation relationship between the camera and the laser scanner. The corresponding three-dimensional point sets are then marked as invalid regions on the point cloud data.

[0086] After removing the two types of interference mentioned above, the points retained in the point cloud data mainly belong to key structural components such as concrete beams, columns, and walls. Based on this, accurate geometric fitting (such as RANSAC fitting) can be performed to extract deformation features. For example, for a column, a cylindrical or cuboid surface can be fitted from its point cloud using a random sampling consensus algorithm; for a wall, a plane can be fitted. By comparing the fitted ideal geometric elements with the pre-earthquake design model or the undamaged baseline model, the geometric deformation parameters such as tilt and deflection of the components can be calculated. This series of removal steps significantly improves the purity of the point cloud data, ensuring that the subsequently extracted geometric deformation features truly reflect the damage status of the main structure, avoiding interference from irrelevant factors in the evaluation results, and providing a reliable three-dimensional geometric information foundation for fusion evaluation.

[0087] S2. Perform semantic segmentation on the image data to extract the apparent damage features of the concrete structure to be evaluated; extract the geometric deformation features of the concrete structure to be evaluated based on the point cloud data; extract the dynamic modal features of the concrete structure to be evaluated based on the vibration data.

[0088] Specifically, such as Figure 3As shown, semantic segmentation of image data is a crucial step in extracting apparent damage features from concrete structures. The core of this method lies in utilizing a pre-trained deep learning semantic segmentation network to automatically identify and classify the acquired structural surface images at the pixel level. Specifically, the segmentation network can accurately classify each pixel in the image into different semantic categories such as "background," "intact concrete surface," and "cracks." For example, when given an image of the side of a concrete beam containing cracks, stains, and shadows, the network can effectively distinguish between genuine crack pixels and visually similar interfering features, ultimately outputting a binary segmentation mask image that highlights only all crack areas.

[0089] To achieve high-precision segmentation, convolutional neural network models trained on large-scale building damage image datasets are typically used. This embodiment employs U-Net, but DeepLabV3+, Mask R-CNN, or the SegFormer network based on the Transformer architecture can also be used. These models have an encoder-decoder structure, capable of simultaneously capturing local details and global contextual information of the image, thus ensuring good recognition capabilities even for small, discontinuous cracks. In application, simply inputting the structural component image obtained on-site into the model will automatically yield the corresponding crack segmentation results.

[0090] After obtaining a precise segmentation mask of the cracks, quantitative surface damage features can be extracted. One of the most important features is the crack density index. The extraction process first identifies each independent crack region based on the segmentation mask through image morphology operations and contour tracking algorithms, and represents it as a continuous set of pixels. For each crack, its length can be obtained by calculating the sum of the Euclidean distances of its skeleton pixels; its average width can be estimated by dividing the area of ​​the crack region by its approximate skeleton length. Simultaneously, the surface area of ​​the structural component containing the crack needs to be determined, which can usually be obtained by converting the 3D dimensions of the component based on the calibration information at the time of image capture or by combining point cloud data. Finally, the weighted sum of the products of the width and length of all cracks, divided by the surface area of ​​the component, yields the crack density index D, which quantifies the severity of surface cracking on the component. img The calculation formula is as follows:

[0091]

[0092] in, Let be the average width of the i-th crack. l i Let A be the length of the i-th crack. surface This represents the surface area of ​​the structural member where the crack is located.

[0093] Extracting the geometric deformation features of the concrete structure to be evaluated based on point cloud data is the core of this approach. It utilizes the massive spatial point coordinates of the structural surface obtained by high-precision 3D laser scanning, and through a series of data processing and geometric analyses, quantifies the spatial position and shape changes of the structure after an earthquake. This process first relies on a clean structural point cloud that has been preprocessed and interference removed. This point cloud data accurately records the three-dimensional morphology of the surfaces of major load-bearing components such as beams, columns, and walls.

[0094] A key step in extracting geometric deformation features is fitting geometric elements to the point clouds of each structural component. Different mathematical models are used for fitting different types of components. For example, for a concrete column, its point cloud is typically fitted to an ideal cylinder or cuboid; for a shear wall or floor slab, its point cloud can be fitted to a spatial plane. This embodiment uses the RANSAC (Random Sample Consensus) algorithm for this fitting process; however, optimization algorithms such as ICP (Iterative Closest Point) can also be used to find the geometric model parameters that best represent the spatial distribution of the point cloud. By comparing the post-earthquake measured point cloud with the pre-earthquake undamaged design model or benchmark point cloud, the deviation between the actual and theoretical positions of the component can be calculated.

[0095] Among numerous geometric deformation features, the inter-story residual displacement angle is a crucial indicator, directly reflecting the degree of lateral deformation of the main structural frame. Extracting this feature requires examining the beam-column members of each frame. Taking a column as an example, firstly, the three-dimensional coordinates of the center points of two representative sections at the top and bottom are determined from its fitted geometric model. Next, a stable reference plane is selected, which can be fitted from the point cloud of the bottom layer that has not moved. The horizontal displacement components of the column's top and bottom center points relative to this reference plane are calculated separately; the difference between the two is the inter-story relative horizontal displacement of the column. Finally, this displacement is subtracted from the story height of the floor where the column is located to obtain the inter-story residual displacement angle at that member. θ r The calculation formula is as follows:

[0096]

[0097] in, and These represent the horizontal displacements of the top and bottom of the beam and column members relative to the reference plane, respectively, with H representing the story height. This indicator effectively reveals the irreversible plastic deformation of the structure caused by seismic forces and is key geometric evidence for assessing the degree of structural damage and its resistance to collapse.

[0098] The geometric deformation features extracted from point clouds using the aforementioned methods provide precise three-dimensional spatial information that cannot be directly obtained from image data. It can not only quantitatively describe the overall tilt or misalignment of a structure but also pinpoint the extent of deformation in specific components. This technology effectively overcomes the limitations of single image recognition techniques, which are limited to two-dimensional appearance information, enabling a comprehensive perception of structural damage from its "surface" to its "shape." The obtained quantitative geometric features, along with appearance damage features and dynamic modal features, corroborate and complement each other, forming a solid data foundation for subsequent feature fusion and model correction, resulting in a more comprehensive, objective, and reliable seismic damage assessment.

[0099] Extracting the dynamic modal characteristics of a concrete structure based on vibration data primarily involves analyzing the environmental vibration response signals recorded by accelerometers installed on the structure to identify its inherent dynamic characteristics. The first-order natural frequency is the most crucial and easily obtainable overall modal parameter. Vibration data is typically obtained in the form of acceleration time-history signals; that is, under environmental excitation, sensors record the acceleration values ​​of each measuring point on the structure over time at a fixed sampling frequency.

[0100] Before feature extraction, the raw vibration data typically requires preprocessing, such as bandpass filtering to remove high-frequency noise and low-frequency drift interference, retaining the frequency bands reflecting the main vibration characteristics of the structure. The preprocessed multi-channel acceleration time history data can then be used for modal identification. A common and effective method is frequency domain analysis, which involves converting the vibration signal in the time domain to the frequency domain for observation. Specifically, the acceleration time history signals at each measuring point are subjected to a fast Fourier transform to calculate their power spectral density function, and then the average regularized power spectrum is obtained through cross-spectral analysis of all measuring point data. This power spectrum will show several distinct peaks; the frequencies corresponding to these peaks are considered the natural frequencies of the structure, with the lowest frequency being the first-order natural frequency. For example, for an undamaged six-story concrete frame structure, the first-order natural frequency identified through environmental vibration testing might be 2.0 Hz; however, after an earthquake causes cracks or stiffness degradation in some beam-column joints, the frequency obtained from subsequent testing might drop to 1.6 Hz.

[0101] The first-order natural frequency of a structure serves as a key dynamic modal characteristic because it is directly proportional to the square root of the overall stiffness and inversely proportional to the square root of the mass. Since earthquake damage primarily leads to a decrease in the stiffness of structural members while the mass remains essentially unchanged, a decrease in the first-order natural frequency can directly and sensitively reflect the degree of attenuation in the overall stiffness of the structure. This characteristic is a global indicator that can effectively capture changes in the macroscopic mechanical properties of a structure caused by cumulative damage.

[0102] The automatic and accurate extraction of the first-order natural frequency, a dynamic modal feature, from vibration data provides a basis for quantifying damage from a structural dynamics perspective in this assessment method. It overcomes the limitations of images and point cloud data, which primarily reflect localized, static damage, by detecting internal microscopic damage or overall performance degradation that has not yet formed obvious surface cracks or geometric deformations. This feature, along with apparent damage and geometric deformation features, is physically interconnected yet each has its own emphasis. Their fusion provides a more comprehensive characterization of the structural damage state, offering crucial overall constraint information for subsequent physical parameter correction in the finite element model, thus significantly improving the systematic nature and reliability of seismic damage assessment.

[0103] S3. Weighted fusion of the apparent damage features, geometric deformation features and dynamic modal features to construct a comprehensive damage feature vector;

[0104] This study employs a weighted fusion of apparent damage features, geometric deformation features, and dynamic modal features to overcome the limitations of assessments based on a single data source. Through deep fusion at the algorithmic level, it constructs a comprehensive and balanced index that characterizes the structural damage state. Since these three types of features originate from images, point clouds, and vibration sensors, respectively, their physical meanings, dimensions, and numerical ranges are drastically different. For example, the crack density index is a dimensionless ratio, the interlayer residual displacement angle is measured in radians, and frequency variation is a physical quantity measured in Hertz. Therefore, simply splicing these features together is meaningless; normalization and weighted fusion are necessary to form a consistent and comparable comprehensive damage feature vector.

[0105] The first step in the fusion process is to normalize the original features to eliminate the influence of dimensions and map them to similar numerical ranges. Normalization is usually based on the possible value range of the feature or historical statistical data. For example, the crack density index can be divided by its empirical upper limit under severe damage conditions; the interlayer residual displacement angle can be divided by the limit displacement angle allowed by the design code; and the frequency change rate is itself a relative ratio between 0 and 1. After processing, all features are transformed into dimensionless values ​​between 0 and 1, with larger values ​​indicating more significant damage in that dimension.

[0106] The core weighted fusion step employs an adaptive weight allocation method based on information entropy. Information entropy is used to measure the uncertainty or information content of the features extracted from each data source. Specifically, the information entropy of image data (crack grayscale distribution), point cloud data (plane fitting residual distribution), and vibration data (signal time history distribution) is calculated separately. H :

[0107]

[0108] For image data, xi This represents the grayscale distribution of the crack pixels; for point cloud data, x i To fit the residual distribution of the plane, for vibration data, x i This represents the time history distribution of the vibration signal.

[0109] The more chaotic the distribution and the higher the uncertainty of a data source, the greater its information entropy and the lower the quality of the information it provides. Therefore, it should be given a lower weight in the fusion process. Conversely, data sources with more concentrated distribution and clearer information receive a higher weight.

[0110] For the t-th data source, its weight coefficient w t The calculation formula is:

[0111]

[0112] in, m Indicates the total number of data sources. H t Represents the information entropy of the t-th data source, 1- H t This represents the information redundancy of the t-th data source. This represents the sum of redundancy across all data sources. Here, there are three data sources: images, point clouds, and vibration data, hence m=3. This allows the weight allocation to dynamically adapt to the quality fluctuations of each data source under different earthquake damage scenarios, achieving intelligent fusion.

[0113] use , and These represent the weighting coefficients for the image, point cloud, and vibration data sources, respectively. Finally, the comprehensive damage feature vector is formed by multiplying the normalized crack density index, the normalized interlayer residual displacement angle, and the normalized rate of change of frequency by their corresponding dynamic weighting coefficients, resulting in a three-dimensional column vector.

[0114]

[0115] in, The normalized crack density index, The normalized inter-layer residual displacement angle. and These represent the post-earthquake measured first-order natural frequencies and the pre-earthquake initial frequencies of the structure, respectively. This vector integrates information from local apparent damage, overall geometric deformation, and macroscopic dynamic performance degradation. Its advantage lies in its ability to dynamically adjust the contribution of each component based on the quality of the data itself, thus stably generating a more reliable and comprehensive digital representation of the damage state even under complex field conditions with uneven data quality. This vector provides a unique, clear, and multi-dimensional target basis for subsequent finite element correction based on the physical model, and is a crucial link in bridging the gap between multi-source data and mechanical model parameters.

[0116] S4. Based on the comprehensive damage feature vector, the stiffness parameters of the preset parametric finite element model of the concrete structure to be evaluated are corrected, and the stiffness parameters of each component are optimized and corrected by a genetic algorithm.

[0117] Correcting the stiffness parameters of the pre-set parametric finite element model of the concrete structure to be evaluated based on the comprehensive damage feature vector is a core step in achieving quantitative inversion from multi-source observation data to the physical and mechanical properties of the structure. Its fundamental purpose is to use the comprehensive feature vector obtained by fusing field measurements, which fully reflects the damage state, to reverse-calibrate and update the original ideal finite element model established based on design drawings, making its calculated response match the measured characteristics as closely as possible, thereby obtaining a digital twin model that can realistically simulate the current mechanical state of the structure after an earthquake. Here, "parametric" means that the key parameters to be corrected in the model, mainly the stiffness reduction coefficients of each structural component, are defined as adjustable variables. Changes in these coefficients directly reflect the stiffness degradation of the components caused by damage such as concrete cracking and steel yielding.

[0118] The correction process is accomplished by deriving a mathematical optimization problem. Specifically, such as... Figure 4 As shown, a parametric finite element model of the structure is established in OpenSees or Abaqus, and a vector is formed by the stiffness reduction coefficients of each component in the structure to be corrected. k To optimize the variables, an objective function is constructed. This objective function typically contains two main parts: the first part is the error norm between the measured comprehensive damage feature vector and the vector calculated by the current finite element model on the same features, which measures the degree of agreement between the model prediction and the measured data; the second part is a regularization term, used to prevent overfitting during the correction process, i.e., avoiding physically unreasonable extreme parameter values ​​obtained by forcibly matching the data. The objective function is expressed as:

[0119]

[0120] in, k Here are the stiffness reduction factors for each structural component. λ The regularization coefficient is . This is the measured comprehensive damage feature vector. This is the comprehensive damage feature vector calculated from the current finite element model.

[0121] By using a genetic algorithm (GA) (and other optimization algorithms such as particle swarm optimization (PSO), Bayesian model updating, or simulated annealing), the stiffness reduction coefficients of each component in the model are continuously adjusted, and finite element analysis is iteratively run to calculate the updated eigenvectors, ultimately minimizing the objective function value. The iteration stops when the error is less than a preset threshold, and the set of stiffness parameters output at this point is considered to be the physical parameters that best match the current structural damage state.

[0122] For example, in a frame structure with severe cracking in a beam-column joint area, the measured feature vector will show a large crack density and local deformation in that area, while the overall frequency will decrease. During the optimization process, the algorithm will automatically attempt to reduce the stiffness reduction factor of the beam and column members corresponding to that area, so that the corrected model can also produce large deformations at the corresponding locations and reproduce the decrease in overall frequency in the simulation analysis, until the overall difference between the model output features and the measured features is minimized.

[0123] To improve correction efficiency and ensure the physical rationality of the solution, this method typically employs a substructure-level correction strategy based on sensitivity analysis. This strategy first calculates the stiffness parameters of each component. k i For target feature vector V damage Sensitivity coefficient , V j Represents the comprehensive damage feature vector V damageThe j-th component is analyzed to assess the sensitivity of changes in the stiffness parameters of each component to each feature in the comprehensive damage feature vector. A sensitivity threshold (e.g., 0.05) is set. For stiffness parameters with sensitivity below the threshold, it indicates that their changes have little impact on the overall features, and they can be treated as known quantities or subjected to overall coarse adjustment during correction. For stiffness parameters with high sensitivity, fine correction is required. In practice, stiffness reduction coefficients with sensitivity to dynamic modal features higher than the first threshold but sensitivity to other features lower than the second threshold are corrected first. Then, stiffness reduction coefficients with sensitivity to geometric deformation features and / or apparent damage features higher than the second threshold are corrected. The first round of correction is performed using stiffness parameters sensitive to global features such as dynamic modal features to quickly grasp the overall stiffness degradation trend of the structure. Then, the second round of correction is performed using parameters sensitive to local deformation features such as geometric deformation features and / or apparent damage features to accurately locate and quantify severely damaged local areas. This hierarchical and step-by-step strategy effectively reduces the complexity of the optimization problem, avoids mutual interference between parameters, and makes the correction process faster and the results more stable.

[0124] Through the above modifications, what is ultimately obtained is not an abstract data label, but a finite element model with updated physical parameters that reflects the true damage distribution. This model explains the observed damage phenomenon from a mechanical perspective, laying a reliable model foundation for subsequent nonlinear pushover analysis or dynamic time history analysis, thereby quantitatively calculating the residual bearing capacity of the structure. This fundamentally solves the "black box" problem of most existing data-driven methods lacking physical model support and making it difficult to directly use the evaluation results for mechanical verification.

[0125] S5. Apply pushover load or incremental dynamic analysis to the modified finite element model, calculate the ultimate bearing capacity of the concrete structure to be evaluated, and calculate the ratio with the required load of the concrete structure under the design earthquake to obtain the safety reserve index of the structure.

[0126] The core objective of step S5 is to use the corrected finite element model obtained in step S4, which reflects the actual damage state, to predict the ultimate resistance and safety margin of the structure under the current damage conditions through numerical simulation. The final output is a quantitative indicator with clear physical meaning and engineering value—the safety reserve index. This step is a crucial step in transforming the results of multi-source data fusion and model correction into quantitative assessment conclusions that can directly serve post-earthquake emergency decision-making and repair design.

[0127] To obtain the current ultimate bearing capacity of a structure, nonlinear static pushover analysis or incremental dynamic analysis is required on the modified finite element model. Pushover analysis is a static method that simulates the overall effect of earthquake action by gradually applying increasing horizontal lateral loads along the model's height in a specific pattern. During loading, the base shear force and vertex displacement of the model are recorded until the structure collapses or the displacement exceeds the limit value, thus plotting a complete base shear force-vertex displacement curve. The base shear force value corresponding to the peak point on this curve is the ultimate bearing capacity of the structure under the current damage state. For example, in a pushover analysis of a model of a span frame beam with a plastic hinge after an earthquake, the base shear force may begin to decrease after reaching a certain peak due to the exhaustion of the plastic hinge's rotational capacity; this peak value is then extracted as the current ultimate bearing capacity.

[0128] Incremental dynamic analysis is a more refined dynamic method. It uses a series of earthquake motion records with increasing intensity as input to perform multiple nonlinear dynamic time-history analyses on the corrected model. For each earthquake motion, the maximum response of the structure under this excitation is recorded, potentially yielding a curve reflecting the relationship between earthquake intensity and structural damage parameters. By analyzing this data, the collapse resistance of the structure under earthquakes of different intensities can be assessed more statistically, and a current bearing capacity level with specific probabilistic significance can be determined. Regardless of the analysis method used, the foundation is the finite element model that has been corrected with multi-source data and whose stiffness distribution closely approximates the actual damage state. Therefore, the obtained ultimate bearing capacity can more reliably reflect the actual residual resistance of the structure after the earthquake.

[0129] The calculation of the safety reserve index essentially compares the assessed structural "residual capacity" with the potential future "earthquake demand." The demand load is typically calculated based on the design response spectrum determined by seismic design codes, using the seismic intensity of the site. It represents the seismic force the structure is expected to withstand under design earthquake conditions. The safety reserve index is the ratio of the current ultimate bearing capacity to the design earthquake demand load, and its calculation formula is as follows:

[0130]

[0131] in, C residual The current ultimate bearing capacity of the structure, S demandThe required load under design earthquake conditions is taken from the elastoplastic shear force requirement in seismic design documents (such as GB50011). This index clearly quantifies the degree of surplus bearing capacity of the structure relative to the standard design requirements under the current state. For example, if the calculated current ultimate bearing capacity of a damaged frame is 15000 kN, while the design load required according to the code is 10000 kN, then its safety reserve index is 1.5. An index greater than 1 indicates that the structure can still meet the design requirements under the current damage; an index equal to or slightly less than 1 indicates that the structure is in a critical state and requires attention; an index much less than 1 clearly warns of serious safety hazards in the structure.

[0132] The technological advantage of this step lies in its application of cutting-edge multi-source sensing, data fusion, and model correction technologies to the concepts of load-bearing capacity and safety factor, which are most familiar and trusted in the engineering field. Through the safety reserve index, decision-makers can intuitively and quantitatively assess the safety, urgency, and repair priority of damaged structures, thereby formulating scientific risk management and reinforcement plans. This fundamentally solves the limitation of traditional methods that can only qualitatively describe damage without providing clear mechanical indicators, and also overcomes the shortcomings of purely data-driven models that lack a physical core and whose results are difficult to directly apply to engineering design, achieving a complete closed loop from intelligent sensing to mechanical assessment.

[0133] The present invention will be further described in detail below with reference to specific embodiments. It should be noted that the following embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0134] Example 1: Seismic damage assessment of a 6-story reinforced concrete frame structure

[0135] In this embodiment, a six-story reinforced concrete frame structure model on a simulated earthquake test bench is selected as the evaluation object, and the method described in this invention is applied to conduct rapid post-earthquake evaluation.

[0136] 1. Hardware system configuration and data acquisition;

[0137] Image acquisition was performed using a DJI Matrice M300 RTK drone equipped with a Zenmuse P1 full-frame camera. The acquisition strategy was as follows: the flight path was set to circle the building, the shooting distance was maintained at 5m-10m, and the image overlap rate was set to 70% horizontally and 60% vertically. The data volume included 240 high-resolution images of structural beams, columns, and nodes, with a resolution of 8192×5460 pixels.

[0138] Geometric scanning was performed using a Leica ScanStation P50 3D laser scanner. The station layout consisted of four stations positioned around the building, with each station scanning for approximately 3 minutes. Point cloud density parameters were set to ensure that the point spacing at 10m intervals was less than 3mm. The equivalent point cloud density calculated using these parameters was approximately 1.1 × 10⁻⁶. 5 point / m 2 It is far higher than the aforementioned minimum density requirement, and can restore the geometry of the concrete column surface with extremely high precision, ensuring the accuracy of the calculation of interlayer residual displacement angle.

[0139] The dynamic sensor uses a 941B type ultra-low frequency vibration pickup (accelerometer), which is placed on the top floor (6th floor), the middle floor (3rd floor), and the ground floor of the structure. The acquisition parameters are: the sampling frequency is set to 200Hz, and the acquisition time is 10 minutes of environmental pulsation signal (i.e., using wind load or ground vibration as excitation, without artificial vibration).

[0140] The computing terminal is a high-performance workstation equipped with an Intel i9 processor, 64GB of memory, and an NVIDIA RTX 3090 graphics card.

[0141] 2. Feature extraction and processing algorithms;

[0142] 1) Crack feature extraction (based on PyTorch framework):

[0143] Network Model: A U-Net semantic segmentation model based on ResNet-50 as the backbone network was constructed. Pre-training was performed using a dataset containing 2000 labeled concrete crack images.

[0144] Processing procedure: Input the on-site image into the model and output a binarized mask. Use a skeletonization algorithm to convert the pixel values ​​of the crack width into the physical width (mm).

[0145] Quantitative results: X-type shear cracks were identified in the second-layer corner column (C2-1), and the normalized crack density index Dimg was calculated to be 0.42 (indicating moderate cracking).

[0146] 2) Residual deformation extraction (based on point cloud processing):

[0147] Point cloud preprocessing: Noise is removed using a statistical outlier removal (SOR) filter.

[0148] Deformation calculation: The point cloud of column C2-1 was sliced, and the central axis of the column was fitted using the RANSAC algorithm. Compared with the pre-earthquake BIM model or gravity direction benchmark, the horizontal residual displacement of the column top relative to the column bottom was measured to be 15 mm, and the story height was 3000 mm.

[0149] Quantification results: Calculation of residual displacement angle θ r =15 / 3000=1 / 200, normalized index Modal parameter identification (based on MATLAB signal processing box):

[0150] 3) Extraction of the first-order natural frequency;

[0151] Algorithm selection: Covariance-driven random subspace method (SSI-COV) is adopted.

[0152] Quantitative results: The first-order natural frequency f of the structure after the earthquake was identified. current =1.85Hz. Referring to the design documents, the initial frequency f of this structure is... initial =2.50Hz. The frequency drop rate is 26%.

[0153] 3. Weighted fusion of multi-source data

[0154] Based on the current data acquisition environment (daytime, good lighting, high sensor signal-to-noise ratio), the weighting coefficients are set as follows:

[0155] Image feature weight α = 0.4 (visual clarity, higher weight);

[0156] Geometric feature weight β = 0.3;

[0157] The dynamic characteristic weight γ = 0.3;

[0158] Construct a damage feature vector V for column C2-1 and the floor it is located on. damage :

[0159] V damage =[0.4×0.42,0.3×0.50,0.3×0.26]T=[0.168,0.150,0.078] T .

[0160] 4. Automatic correction of finite element model

[0161] Initial model: A nonlinear fiber beam-column element model of the framework is created in OpenSees software.

[0162] Optimize settings:

[0163] Optimization variables: the elastic modulus reduction factors k1, k2, ..., k6 for each column in layers 1-6. The initial values ​​are all 1.0.

[0164] Objective function: Minimize the Euclidean distance between the measured feature vector and the simulated feature vector.

[0165] Algorithm: Genetic algorithm (GA) is used. The population size is set to 50, the maximum number of iterations is 100 generations, and the mutation probability is 0.05.

[0166] Correction process:

[0167] The algorithm converged after 45 iterations. It automatically adjusted the stiffness reduction factor k2 of the second floor to 0.65 (i.e., the stiffness was degraded by 35%), while the k value of other floors remained above 0.9.

[0168] Note: Because the feature vector contains visual crack information and displacement angle information unique to the second layer, the model successfully attributed the overall frequency decrease to the damage in the second layer, achieving accurate localization.

[0169] 5. Quantitative assessment of structural safety reserves;

[0170] Numerical simulation: Standard pushover analysis was applied to the modified model (the stiffness of the second layer was reduced to 0.65).

[0171] Calculation results: The base shear force-apex displacement curve was obtained, and the current ultimate bearing capacity C of the structure was calculated. residual =1200kN.

[0172] Based on the local seismic fortification intensity (8 degrees, 0.2g), calculate the required shear force S for the fortified earthquake. demand =1000kN.

[0173] Final conclusion: The calculated Safety Reserve Index (SRI) is 1200 / 1000 = 1.2.

[0174] Assessment report generated: Although there are obvious cracks and residual deformation on the second floor, the overall structure still has a 20% safety margin. It is recommended to "suspend use and resume use after local reinforcement" rather than "demolishing immediately".

[0175] Example 2: Adaptive Adjustment under Special Environments

[0176] Based on Example 1, in order to verify the environmental adaptability of the present invention, a nighttime post-earthquake assessment scenario was simulated.

[0177] Scene changes: extremely poor lighting, blurry images captured by drone, and low signal-to-noise ratio.

[0178] Weight Adaptive: If the image sharpness is detected to be below the threshold, the weights are adaptively adjusted to: α=0.1 (for reference only), β=0.4, γ=0.5.

[0179] Results: Although most of the texture details are lost, the system mainly relies on macroscopic deformation (column tilt) captured by the LiDAR and frequency changes of the sensor.

[0180] The final model correction result shows k2=0.68, which is only 4.6% different from the daytime condition (0.65), proving that the system can still maintain high evaluation reliability when visual data fails.

[0181] Based on the same inventive concept, this invention also provides a seismic damage assessment system for concrete structures based on multi-source data feature fusion, comprising:

[0182] The data acquisition module is used to collect multi-source heterogeneous data of the concrete structure to be evaluated, including image data, point cloud data, and vibration data.

[0183] The feature extraction module is used to perform semantic segmentation on the image data to extract the apparent damage features of the concrete structure to be evaluated; extract the geometric deformation features of the concrete structure to be evaluated based on the point cloud data; and extract the dynamic modal features of the concrete structure to be evaluated based on the vibration data.

[0184] The feature fusion module is used to weight and fuse the apparent damage features, geometric deformation features and dynamic modal features to construct a comprehensive damage feature vector.

[0185] The model correction module is used to correct the stiffness parameters of the preset parametric finite element model of the concrete structure to be evaluated based on the comprehensive damage feature vector.

[0186] The bearing capacity assessment module is used to apply push-over loads or incremental dynamic analysis to the modified finite element model and calculate the safety reserve index of the concrete structure to be assessed.

[0187] All relevant content of each step involved in the aforementioned embodiments of the concrete structure seismic damage assessment method based on multi-source data feature fusion can be referenced to the functional description of the corresponding functional module of the system in the embodiments of this application, and will not be repeated here.

[0188] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for assessing seismic damage to concrete structures based on multi-source data feature fusion, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous data of the concrete structure to be evaluated, including image data, point cloud data and vibration data; S2. Perform semantic segmentation on the image data to extract the apparent damage features of the concrete structure to be evaluated; extract the geometric deformation features of the concrete structure to be evaluated based on the point cloud data; extract the dynamic modal features of the concrete structure to be evaluated based on the vibration data. S3. Weighted fusion of the apparent damage features, geometric deformation features and dynamic modal features to construct a comprehensive damage feature vector; S4. Based on the comprehensive damage feature vector, the stiffness parameters of the preset parametric finite element model of the concrete structure to be evaluated are corrected, and the stiffness parameters of each component are optimized and corrected by a genetic algorithm. S5. Apply pushover load or incremental dynamic analysis to the modified finite element model, calculate the ultimate bearing capacity of the concrete structure to be evaluated, and calculate the ratio with the required load of the concrete structure under the design earthquake to obtain the safety reserve index of the structure.

2. The method for assessing seismic damage to concrete structures as described in claim 1, characterized in that, S1 also includes data preprocessing steps, including: distortion correction of image data, denoising and registration of point cloud data, and bandpass filtering denoising of vibration data.

3. The method for assessing seismic damage to concrete structures as described in claim 1 or 2, characterized in that, S1 also includes a step of removing unstructured interference from the point cloud data, including: A method combining temporal multi-frame difference and octree raster mapping is used to automatically remove point cloud noise from moving objects; A semantic segmentation network is used to identify unstructured components in image data, and then the point cloud regions of the identified unstructured components are marked as invalid regions in the point cloud data. Geometric fitting is performed only on the point cloud of structural components.

4. The method for assessing seismic damage to concrete structures as described in claim 1, characterized in that, In S2, the apparent damage characteristics include the crack density index D. img The calculation formula is as follows: in, Let be the average width of the i-th crack. l i Let A be the length of the i-th crack. surface The surface area of ​​the structural member where the crack is located; Geometric deformation characteristics include the inter-story residual displacement angle of beam and column members. θ r The calculation formula is as follows: in, and These are the horizontal displacements of the top and bottom of the beam / column members relative to the reference plane, respectively, where H is the story height; Dynamic modal characteristics include the first-order natural frequencies of the structure.

5. The method for assessing seismic damage to concrete structures as described in claim 1, characterized in that, In S3, a normalized weighted fusion algorithm is used to construct a comprehensive damage feature vector, and the weighting coefficients are dynamically adjusted based on the information entropy of each data source.

6. The method for assessing seismic damage to concrete structures as described in claim 5, characterized in that, The comprehensive damage feature vector is represented as follows: in, The normalized crack density index, The normalized inter-layer residual displacement angle, and The values ​​are the measured first-order natural frequencies of the structure after the earthquake and the initial natural frequencies before the earthquake, respectively. , and The weighting coefficients correspond to the image, point cloud, and vibration data sources, respectively. The weighting coefficients for each data source are calculated using the information entropy weighting method, specifically including: Calculate the information entropy of each data source. H : For image data, x i This represents the grayscale distribution of the crack pixels; for point cloud data, x i To fit the residual distribution of the plane, for vibration data, x i The time history distribution of the vibration signal, p ( x i ) represents a feature x i The probability of occurrence; For the t-th data source, its weight coefficient w t The calculation formula is: in, m Indicates the total number of data sources. H t Represents the information entropy of the t-th data source, 1- H t This represents the information redundancy of the t-th data source. This represents the sum of redundancy across all data sources.

7. The method for assessing seismic damage to concrete structures as described in claim 1, characterized in that, In S4, an objective function is established for the parametric finite element model of the concrete structure to be evaluated. J ( k The objective function is iteratively optimized using an optimization algorithm. The process stops when the error is less than a preset threshold, and the corrected stiffness parameters and the corresponding corrected finite element model are output. J ( k )for: in, k Here are the stiffness reduction factors for each structural component. λ The regularization coefficient is . This is the measured comprehensive damage feature vector. This is the comprehensive damage feature vector calculated from the current finite element model.

8. The method for assessing seismic damage to concrete structures as described in claim 7, characterized in that, S4 employs a substructure hierarchical correction strategy based on sensitivity analysis, including: Calculate the sensitivity of the stiffness reduction factor of each structural component to each feature in the comprehensive damage feature vector; First, correct the stiffness reduction factor for dynamic modal features that are more sensitive to the first threshold but less sensitive to other features that are less sensitive to the second threshold. Then, correct the stiffness reduction factor for geometric deformation features and / or apparent damage features that are more sensitive to the second threshold.

9. The method for assessing seismic damage to concrete structures as described in claim 1, characterized in that, The safety reserve index calculated in S5 is: in, C residual The current ultimate bearing capacity of the structure, S demand The required load under the design earthquake is taken from the elastoplastic shear force requirement in the seismic design code document; The current ultimate bearing capacity of the structure is obtained by applying a pushover load or incremental dynamic analysis to the modified finite element model to obtain the base shear force-apex displacement curve of the structure, and then reading the current ultimate bearing capacity of the structure from the base shear force-apex displacement curve.

10. A seismic damage assessment system for concrete structures based on multi-source data feature fusion, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous data of the concrete structure to be evaluated, including image data, point cloud data, and vibration data. The feature extraction module is used to perform semantic segmentation on the image data to extract the apparent damage features of the concrete structure to be evaluated; extract the geometric deformation features of the concrete structure to be evaluated based on the point cloud data; and extract the dynamic modal features of the concrete structure to be evaluated based on the vibration data. The feature fusion module is used to weight and fuse the apparent damage features, geometric deformation features and dynamic modal features to construct a comprehensive damage feature vector. The model correction module is used to correct the stiffness parameters of the preset parametric finite element model of the concrete structure to be evaluated based on the comprehensive damage feature vector. The bearing capacity assessment module is used to apply push-over loads or incremental dynamic analysis to the modified finite element model and calculate the safety reserve index of the concrete structure to be assessed.

Images