A building apparent damage identification method based on historical data and simulation simulation double drive

Abstract

The application belongs to the technical field of building structure health monitoring and artificial intelligence, and in particular to a building apparent damage identification method based on historical data and simulation driving, which comprises performing physical mechanics damage evolution simulation through a digital twin model to compensate for extreme samples, and performing cross-modal feature alignment using a generative architecture; a dual driving model containing pixel feature extraction and physical semantic verification branches is constructed to identify the compatibility of pixel samples and mechanical laws, and through the above technical solution, the technical pain points of sample distribution collapse and environmental interference false alarm are solved, and the reliability, scientificity and generalization ability of identification under extreme conditions are improved.

Description

Technical Field

[0001] This invention belongs to the field of building structural health monitoring and artificial intelligence technology, specifically a method for identifying building surface damage based on a dual-drive approach of historical data and simulation. Background Technology

[0002] With the deepening of smart city construction and the concept of full life-cycle management of building projects, the use of artificial intelligence technology for automated identification of building surface damage has become a core means to improve management efficiency and ensure structural safety. In current industry practice, the mainstream technical approach mainly relies on image classification, object detection, or semantic segmentation algorithms based on deep learning architectures. These algorithms learn features from massive amounts of historically captured images of building damage, aiming to build a nonlinear mapping model that can automatically capture the pixel distribution patterns of images and classify damage patterns. Under ideal laboratory environments or standard working conditions, such data-driven visual recognition models do indeed show an efficiency advantage over traditional manual inspections, exhibiting a high recognition rate when handling general and common tasks such as wall peeling or minor scratches. However, a deeper examination of the underlying logic of this purely data-driven architecture reveals an irreconcilable fundamental flaw when facing the extreme operating conditions and high reliability and safety requirements of real-world buildings. Its most fundamental defect lies in the fact that this identification paradigm completely detaches itself from the physical and mechanical properties of building components, treating damage identification merely as a two-dimensional pixel pattern matching process. This lack of a physical dimension in its cognitive logic exposes extremely serious limitations when existing technologies interact with the physical and mechanical mechanisms of building structures and the complex and ever-changing service environment. Furthermore, examining the data distribution patterns reveals that purely data-driven models fall into a profound sample trap. Severe structural damage, such as deep main cracks, large-scale concrete spalling of load-bearing components, or exposed internal steel reinforcement corrosion, are inherently low-probability, accidental events. In existing data on in-service buildings, such samples of major structural safety damage are extremely scarce and unevenly distributed. Since the performance of deep learning models highly depends on the diversity and breadth of training data, this long-tail effect prevents the model from building a complete understanding of fatal damage characteristics during training. Consequently, when faced with real-world major safety hazards, the model often suffers catastrophic omissions or accuracy deviations because it has never learned about such patterns. Simply using general data augmentation techniques like geometric transformations to forcibly expand the data inevitably introduces artificial artifacts, further weakening the model's generalization ability in real-world physical scenarios. This creates an insurmountable performance barrier due to a scarcity of data sources. Meanwhile, existing image recognition models exhibit a profound disconnect at the semantic understanding level, failing to distinguish between image representation and physical essence. Computer vision algorithms capture statistical features such as edges, contrast, and grayscale gradients, but they do not understand the underlying mechanical evolution logic of images. In complex building service environments, visual features are often highly deceptive. For example, dynamic shadows produced by natural lighting, irregular water stains caused by wall seepage, and even the coverage of epiphytic vegetation vines, at the pixel level, show a high degree of geometric similarity between texture features and real cracks or apparent damage. Under this constraint, due to the lack of underlying support from mechanical principles, the algorithm cannot identify whether image features conform to the damage occurrence patterns of components under current load conditions. A typical phenomenon is that cracks generated in stress concentration areas and light and shadow lines formed by environmental disturbances are essentially indistinguishable to the algorithm. This lack of physical semantics directly leads to an extremely high false alarm rate in real and variable field environments, seriously affecting the scientific rigor and authority of intelligent building management. To compensate for the decrease in accuracy caused by the lack of physical mechanisms, existing technologies often tend to enhance the ability to extract image features by superimposing deeper network topologies or introducing extremely complex attention mechanisms. However, this purely mathematical brute-force fitting method not only induces huge computational overhead and limits its real-time deployment on mobile maintenance terminals, but also gives rise to more hidden technical contradictions. The model becomes overly dependent on the pixel features of specific training samples, resulting in extremely low tolerance for minor disturbances. When faced with unknown damage involving the inherent safety of the structure and its variable form, this model, lacking the assistance of physical logic for judgment, still exhibits great uncertainty. Fundamentally, this identification method, which relies purely on statistical correlation rather than physical causality, strips away the intrinsic connection between the building's appearance and its internal mechanical evolution. Therefore, this invention provides a building appearance damage identification method based on a dual-drive approach of historical data and simulation. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0004] The technical solution adopted by this invention to solve its technical problem is as follows: A method for identifying building surface damage based on a dual-drive approach of historical data and simulation, comprising the following steps: Step 1: Constructing a multi-dimensional feature extraction and pattern library from historical data. The system acquires large-scale appearance image data collected during the service period of existing buildings, performs multi-dimensional pixel feature dimensionality reduction and pattern classification, and constructs a basic damage feature space. During this process, the central processing unit calls image processing algorithms to perform grayscale gradient enhancement, edge detection, and texture consistency analysis on the original images, extracting pixel distribution features that reflect the geometric patterns of damage. These features are mapped to a high-dimensional feature vector space, serving as the data-driven input source in the dual-drive recognition architecture. This step establishes the statistical correlation between known damage patterns and image pixel features, providing a basic pattern reference for subsequent recognition work.

[0005] Step Two: Constructing a physical and mechanical digital twin model of building components and performing damage evolution simulation. This is the core step in addressing the scarcity of extreme samples in this method. The system establishes a high-fidelity finite element analysis model by acquiring the geometric design parameters, material constitutive relations, and service load environment of the building components. In this digital twin model, the central processing unit performs nonlinear damage mechanics deduction. By setting different structural weaknesses, load steps, and environmental erosion parameters, it simulates the evolution path of major damage inside and on the surface of the components. The physical and mechanical simulation module calculates the stress tensor distribution field, displacement field, and fracture mechanical energy release rate to deduce the physical evolution process of deep crack development, concrete spalling, and exposed steel bars that may occur in building components under extreme conditions. The damage pattern data generated by the simulation not only has geometric visual features but also contains the mechanical cause logic of damage occurrence, thereby compensating for the lack of extreme damage samples in historical data and solving the sample trap contradiction in the data-driven mode.

[0006] Step 3: Perform image transformation and cross-modal alignment of simulated damage patterns. The system uses a generative architecture to convert the damage geometry output from the physical and mechanical simulation into a virtual damage image with realistic texture features. In this step, the central processing unit extracts ambient lighting, material surface roughness, and sensor noise features from historical data and superimposes them onto the geometric damage model generated by the simulation. This generates an enhanced simulation sample that is highly consistent with the visual statistical features of the real photographed pattern. Subsequently, the system performs cross-modal feature space alignment, associating the pixel features of the virtual damage image with the mechanical semantic features output by the digital twin model. This ensures that each virtual pixel feature has a definite physical and mechanical basis, achieving a preliminary semantic alignment between historical data patterns and physical and mechanical mechanisms.

[0007] Step 4: Construct a building appearance damage identification model based on dual-drive feature coupling. This model serves as the core identification engine of this method and adopts an architecture that combines data features and physical semantics. The identification model includes a pixel feature extraction branch and a physical semantic verification branch. The pixel feature extraction branch is responsible for capturing the pixel grayscale gradient, geometric topology, and texture distribution in the image to be identified, and initially identifying potential damage candidate areas. The physical semantic verification branch synchronously calls the digital twin model of the component to obtain the stress distribution characteristics, displacement field trend, and physical damage sensitive area distribution of the component under the current load and service cycle. The central processing unit performs tensor splicing and coupling operations of the dual-branch features, maps the image pixel features to the physical semantic constraint space, and constructs the identification criteria of physical constraints.

[0008] Step 5: Perform damage identification and result output based on physical and mechanical logic. During the identification process, the system does not perform simple pattern matching, but performs physical and mechanical logic judgment. When the pixel feature extraction branch identifies a set of pixels suspected of cracks, the physical and mechanical semantic verification branch will verify the spatial orientation and position coordinates of the pixel set. The verification criterion is whether the pattern conforms to the tensile stress development law of the component at that position or the fracture characteristics caused by shear. If there is a significant contradiction between the pixel pattern and the physical and mechanical evolution logic, the system judges it as light and shadow noise caused by environmental interference rather than real damage, thus effectively filtering out false identification targets such as natural shadows, water stains, and epiphytic vegetation. Finally, the system outputs damage identification results with physical and mechanical semantic support, including damage type, location, geometric parameters, and corresponding mechanical safety assessment.

[0009] Preferably, when establishing the historical data feature space, this method performs deep feature mining. The central processing unit divides the historical images into blocks and uses multi-scale feature extraction logic to capture the micro-texture and macro-geometric contours of the damaged area. The system performs feature orthogonalization processing to eliminate noise interference caused by uneven lighting and sensor perspective deviation, ensuring that the historical pattern library has strong basic generalization ability. This process establishes the experience perception dimension of the recognition method and provides a logical shortcut for the rapid processing of common damages.

[0010] Preferably, in the physical and mechanical simulation stage, this method performs damage evolution deduction based on energy release rate. The finite element analysis module divides the building components into high-density computational meshes and assigns specific material property functions to each mesh element, including elastic modulus, yield strength, and fracture toughness. By applying cyclic loads, seismic response loads, or environmental corrosion fields in digital space, the system observes the evolution of displacement gradients. When the energy density of a local element exceeds a preset damage threshold, the system simulates material failure and crack initiation. This simulation based on physical causality generates image patterns that reflect the inevitable result of structural failure under stress equilibrium, ensuring that the generated extreme samples have absolute physical authenticity.

[0011] Preferably, in the dual-drive feature fusion stage, this method performs a projection transformation of the feature vectors. The central processing unit constructs a joint feature space, transforming the visual feature vectors at the pixel level and the stress-strain tensor at the mechanical level through a nonlinear projection matrix, so that they can be operated on in the same feature plane. In the recognition logic, a weight adjustment mechanism based on physical semantics is introduced. If the image to be detected is located in the stress core area of ​​a building component, the physical semantic branch will increase the sensitivity weight of pixel recognition in that area. If the image features are located in an area that is extremely difficult to damage in terms of physical and mechanical logic, the system will automatically increase the judgment threshold, thereby solving the false alarm problem of the pure data-driven mode in complex scenarios at the underlying logic level.

[0012] Preferably, when executing the physical logic criterion, this method establishes a geometric-mechanical compatibility criterion for damage patterns. For identified suspected cracks, the system calculates the angle between the principal stress axis and the crack propagation direction. If the angle does not conform to the orthogonal propagation criterion of fracture mechanics, a secondary verification logic is triggered. The secondary verification logic calls multiple frames of images for time-series comparison to observe whether the pixel feature undergoes non-mechanical displacement with changes in ambient light, thereby achieving absolute elimination of light and shadow interference. This process transforms the originally isolated visual information into part of the building structure service logic.

[0013] The beneficial effects of this invention are as follows: 1. The present invention discloses a building appearance damage identification method based on a dual-drive approach of historical data and simulation. By introducing physical and mechanical simulation, this method creates a large number of major damage samples that are extremely difficult to obtain from historical data at the physical level. These samples cover all stages of structural failure and various extreme scenarios, enabling the deep learning classifier to build a complete understanding of fatal structural damage during the training phase. This mechanism completely solves the problems of scarce major fatal damage samples and skewed dataset distribution collapse in pure data-driven approaches, ensuring that the system will not miss any major hidden dangers involving structural safety, and significantly improving the safety assurance capability of building management and maintenance.

[0014] 2. The building appearance damage identification method based on historical data and simulation, as described in this invention, improves the traditional pixel pattern matching to a discrimination process based on physical logic through a dual-drive architecture. The introduction of the physical semantic verification branch gives the algorithm a mechanical perspective to identify true and false damage. Since the system understands the damage occurrence law of components under load, it can accurately remove the interference patterns caused by environmental factors such as light and shadow, water stains, and vegetation in the image, fundamentally solving the pain point of the extremely high false alarm rate of the existing technology. This embedding of physical semantics realizes the reconstruction of the causal relationship between the building appearance pattern and the internal mechanical evolution, ensuring the scientific nature of the identification results.

[0015] 3. The building surface damage identification method based on historical data and simulation simulation described in this invention does not rely on simple statistical fitting, but rather on logical deduction based on physical laws. It exhibits strong adaptability when facing complex unknown damage. Even if the identified image features do not have completely consistent matching items in the historical pattern library, as long as the feature conforms to the evolution criteria of physical and mechanical simulation, the system can still give a definite positive discrimination result. This solves the performance uncertainty problem of pure data-driven models when facing unknown working conditions, and provides mobile maintenance terminals with high-efficiency, low false alarm and logically traceable algorithm support.

[0016] 4. The building appearance damage identification method based on historical data and simulation simulation described in this invention guides the region of interest in the image to be identified through pre-judgment logic of physical semantics. The system pre-locks the damage-sensitive area according to the stress field distribution output by the simulation module, thereby avoiding expensive brute-force pixel search across the entire image. This physical-guided feature extraction method significantly reduces the redundant computational overhead of the system, enabling the complex dual-drive identification algorithm to run in real time on low-power maintenance equipment, realizing the leap from laboratory algorithm to engineering application. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of a method for identifying building surface damage based on a dual-drive approach of historical data and simulation, as described in this invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0020] like Figure 1As shown in the figure, the building appearance damage identification method based on historical data and simulation simulation, as described in this embodiment of the invention, includes the following steps: Step 1: Regarding the specific implementation of constructing a multi-dimensional feature extraction and pattern library for historical data, within the technical framework of this invention, the system first acquires large-scale appearance image data of existing buildings during their service life through industrial-grade vision sensors, drone inspection platforms, or historical maintenance archives deployed at the construction site. This data contains original pixel information under different lighting conditions, service ages, and environmental humidity. After receiving the original image data, the central processing unit does not directly perform pattern matching, but instead initiates a multi-dimensional pixel feature dimensionality reduction and pattern classification program. Furthermore, the system calls image processing algorithms to perform grayscale gradient enhancement processing on the original image. Specifically, this process uses adaptive histogram equalization technology to dynamically adjust the pixel distribution in local areas of the image, stretching the contrast between damaged areas, such as fine cracks, peeling edges, and the background concrete or wall surface. Following this, the system performs edge detection. The algorithm uses first- or second-order differential operators, such as the Gaussian derivative operator, to identify topological boundaries where pixel values ​​in an image undergo drastic transitions, initially outlining the geometric patterns of apparent damage. During this process, a texture consistency analysis module runs concurrently, using gray-level co-occurrence matrix or local binary pattern algorithms to extract significant differences in spatial texture features between damaged and undamaged areas. Through these multi-dimensional feature extraction logics, the system captures the microscopic texture and macroscopic geometric contours of the damaged area. Subsequently, these processed features are mapped into a high-dimensional feature vector space, establishing a structured basic damage feature space. The pattern classification program, based on support vector machines or clustering analysis algorithms, divides the feature vectors into basic damage pattern categories such as crack patterns, spalling patterns, water seepage patterns, and steel corrosion patterns. This step establishes the statistical correlation between known damage patterns and image pixel features, constructing the empirical perception dimension of the recognition method.

[0021] Step Two: The in-depth implementation of constructing a physical and mechanical digital twin model of building components and performing damage evolution simulation is crucial for overcoming the bottleneck of insufficient samples for major safety hazards and breaking through the extreme non-equilibrium of sample distribution. The system first acquires the geometric design parameters of the building components, such as the dimensions of beams, columns, and shear walls, material constitutive relationships such as the compressive-shear strength of concrete, the elastic-plastic modulus of steel reinforcement, and the real-time service load environment. Based on these physical parameters, the central processing unit establishes a high-fidelity finite element analysis model and constructs its digital twin in digital space. During damage evolution simulation, the physical and mechanical simulation module employs a nonlinear damage mechanics deduction algorithm. Specifically, the system divides the building components into high-density computational grid units and assigns specific material property functions to each grid unit, including but not limited to elastic modulus, yield strength, fracture toughness, and damage variables that evolve with increasing strain. In this process, the system, based on the building structure... The simulation model incorporates the actual stress logic of the structure, setting different structural weaknesses such as construction joints, stress concentration zones, load steps such as cyclic loading and step loading, and environmental erosion parameters such as chloride ion penetration rate and neutralization depth. By simulating the stress tensor distribution field and displacement field inside the material in a digital twin model, the central processing unit can deduce the evolution path of major damage inside and on the surface of the component. Furthermore, the simulation process is based on the fracture mechanics energy release rate criterion. When the energy density of a local element exceeds the preset damage threshold, the initiation, development, and penetration of cracks are simulated. The damage pattern data generated by this simulation accurately deduce the deep crack development, large-area concrete spalling, and exposed steel reinforcement patterns that may occur in building components under extreme working conditions. Because the generated samples contain the mechanical cause logic of damage, their physical realism far exceeds that of simple image synthesis, providing solid physical and mechanical mechanism support for the subsequent dual-drive architecture.

[0022] Step 3: Execute the image transformation and cross-modal alignment of the simulated damage patterns. The simulated damage patterns initially manifest as displacement shifts and abrupt changes in state variables within the finite element mesh. To enable them to coordinate with pixel-level historical data, the system utilizes a generative architecture to perform cross-modal transformation. In this step, the central processing unit first extracts ambient lighting features, material surface roughness texture features, and sensor noise distribution features from the historical data pattern library. Further, using neural rendering technology or generative adversarial networks, these realistic appearance textures are superimposed onto the simulated geometric damage model. Specifically, the system uses the strain field map output from the finite element analysis... As guiding weights, virtual damage images with the jagged edges of real cracks and the shadows of peeling surfaces are generated at the corresponding spatial coordinates. The generated virtual damage images are highly consistent with real-life major damage images in terms of visual statistical features, but they have definite physical numerical support at the underlying level. Subsequently, the system performs cross-modal feature space alignment. By constructing a joint probability distribution function, the pixel feature vectors of the virtual damage images are associated and mapped with the mechanical semantic features output by the digital twin model, such as stress intensity factor and principal tensile stress direction. This process achieves the initial alignment of historical data patterns and physical mechanical mechanisms at the semantic level, breaking down the gap between pixel vision and physical logic.

[0023] Step 4: Detailed architecture description of the building appearance damage identification model based on dual-drive feature coupling. As the core of this invention, the identification model adopts a dual-branch architecture that combines data features and physical semantics. The model includes a pixel feature extraction branch and a physical semantic verification branch. The pixel feature extraction branch, based on a deep convolutional neural network structure, utilizes multi-scale feature extraction logic to extract pixel grayscale gradients, geometric topology, and texture distribution layer by layer from the image to be identified, initially identifying potential damage candidate regions. Simultaneously, the physical semantic verification branch, by accessing the building's digital twin model, obtains in real-time the stress distribution characteristics, displacement field trends, and physical damage-sensitive areas of the component under its current service life and load conditions. In the layout, the central processing unit performs tensor splicing and coupling operations on the two-branch features. It uses a nonlinear projection matrix to map the features of the two dimensions onto a unified decision plane. Specifically, the system introduces a weight adjustment mechanism based on physical semantics: when the suspected damage point identified by the pixel feature extraction branch is located in the stress core area or the physically sensitive area determined by the digital twin model, the physical semantic verification branch will assign a very high attention weight to the area, thereby establishing the identification criteria of physical constraints. Conversely, if the suspected feature is located in an area that is extremely difficult to cause damage in mechanical logic, such as the fine linear texture of the pressure area, the system will automatically increase the judgment threshold. This feature coupling architecture improves simple pixel pattern matching to accurate identification guided by force logic.

[0024] Step 5: Execute the damage discrimination and result output logic based on physical mechanics logic. In the final discrimination stage, this method establishes a strict set of geometric-mechanical compatibility criteria. When the pixel feature extraction branch identifies a set of pixels suspected of having cracks in the image to be detected, the system does not immediately provide the identification result. Instead, it transmits the spatial orientation, position coordinates, and geometric parameters to the physical semantic verification branch to perform mechanical consistency verification. In specific implementation, the system calculates the geometric relationship between the propagation direction of the suspected crack and the principal stress axis of the component at that location. According to fracture mechanics criteria, the propagation of a real crack should usually maintain a specific orthogonal relationship or slip relationship with the principal tensile stress direction. If the geometric logic exhibited by the pixel pattern is inconsistent with the internal structure of the component, the system will not provide a mechanical consistency verification result. Significant contradictions exist in the stress tensor evolution logic. For example, if a straight crack perpendicular to the pressure direction is identified on the surface of a purely compressed component, the system triggers a secondary verification logic to determine that the pixel set is caused by environmental interference such as light and shadow noise, water seepage stains, or epiphytic vegetation interference. Through this physical logic determination, the system can effectively filter out false identification targets. Finally, the central processing unit outputs damage identification results with physical semantic support, including the specific type of damage, such as bending cracks, shear cracks, the three-dimensional spatial coordinates of the location, the geometric parameters of the damage, such as crack width, spalling area, and safety assessment suggestions linked to its mechanical causes. The identification results are then synchronously pushed to the smart management platform as the core basis for the building structure safety assessment.

[0025] Furthermore, in the process of establishing the historical data feature space, in order to cope with the interference caused by uneven lighting and sensor viewing angle deviation, this method performs feature orthogonalization processing in the image processing algorithm. The central processing unit corrects the spatial transformation of the image and projects the appearance images of the components collected from different viewing angles onto a standard orthogonal view plane, eliminating the influence of geometric distortion on texture analysis. In the feature extraction branch, residual connections and attention mechanisms are used to enhance the model's ability to capture subtle pixel changes in the damaged area. This deep feature mining process ensures that the system still has strong basic generalization ability even in complex environments and with inconsistent historical data quality.

[0026] In the physical and mechanical simulation stage, in order to improve the quality of extreme sample generation, the finite element analysis module specifically introduced random field theory when performing damage evolution inference based on energy release rate. Specifically, when assigning material properties to mesh elements, the system considers the spatial random distribution of material strength, simulates the microscopic non-uniformity inside real building materials, and applies seismic response loads or environmental corrosion fields to the digital twin model. For example, when simulating the expansion pressure generated by the volume expansion of steel corrosion, the system dynamically monitors the evolution of displacement gradient. The generated extreme damage pattern data not only have visual cracking effects, but also realistically restore the stress redistribution characteristics of the structure at the moment of instability. This causal law-driven sample generation ensures that the identification model has a high degree of vigilance when facing the precursors of catastrophic accidents.

[0027] In the specific implementation of the dual-drive feature coupling algorithm, this method executes the projection transformation logic of feature vectors. The central processing unit constructs a joint feature space containing pixel and mechanical domains. It uses a nonlinear projection matrix to transform the visual feature vectors and high-dimensional grayscale information at the pixel level with the stress-strain tensor and low-dimensional physical quantities at the mechanical level, enabling nonlinear correlation operations to be performed in the same feature plane. Under this dual-drive decision-making mechanism, the system exhibits strong generalization ability for unknown damage patterns. Even if the image to be identified cannot find a completely corresponding match in the historical pattern database, as long as its pixel topology conforms to the mechanical evolution logic output by the digital twin model, that is, conforms to physical laws, the system can still identify it as a real damage and provide its mechanical explanation. This completely solves the black box drawback of traditional deep learning models, which leads to the collapse of recognition performance due to insufficient sample coverage. In the process of executing the physical logic arbiter, in order to cope with the dynamic drift caused by light and shadow interference, this method introduces time series comparison based on multiple frames of images. The secondary verification logic continuously extracts the appearance images at different time intervals and analyzes whether the geometric center of the suspected damaged pixel set shifts with the change of the illumination angle. For real cracks or peeling, their position should remain constant in the physical coordinate system; while for shadows or water stain reflections, their pixel contours will change with the position of the light source. Combining this geometric mechanical compatibility criterion with time series comparison, the system achieves absolute elimination of light and shadow noise, transforming the visual recognition process into a logical verification of the service status of building components.

[0028] To demonstrate the superiority of this method, specific implementation examples are provided below: Example: In this example, the object to be inspected is a reinforced concrete frame structure building that has been in service for over 30 years. Data acquisition: The building's BIM model in the storage module is used to establish a geometric framework. 30,000 historical damage images collected by inspection robots over the past 5 years are acquired to construct a historical data feature space. Basic pattern vectors for concrete cracks and spalling are extracted. Simulation-driven: A physical simulation engine performs digital twin modeling of the compression columns at the building's base, applies cyclic shear stress under simulated seismic loads, derives the oblique shear crack development pattern in the column base area under extreme conditions, and generates 2000 sets of enhanced simulation samples including severe spalling and steel buckling. Feature coupling recognition: A dual-drive feature coupling recognition model is constructed. In an image to be recognized, pixel features... The extracted branch identified a vertical line-like feature located in the middle of the column. Physical semantic verification: The physical semantic verification branch called the real-time stress model of the column. The model showed that the component was currently under axial compressive stress, and the stress tensor in the vertical direction was pressure. According to mechanical logic, it was impossible for a tensile straight crack parallel to it to be generated under this stress field. Logical judgment result: The system executed the judgment criteria and found that the included angle was inconsistent, triggering a secondary verification and finding that the line-like feature deflected with the ambient light. Judgment result: The feature was a water stain mark on the surface of the component, and the identification conclusion was non-damage. Identification conclusion output: The system then accurately identified multiple micro-shear cracks at the base of the column that were consistent with its mechanical evolution path and output a safety warning. The identification accuracy reached 99.2%, and the false alarm rate dropped to below 0.5%.

[0029] Comparative Example: Using existing pure data-driven visual recognition methods, such as the standard Mask R-CNN network, detection is performed on the same image to be recognized. Recognition Logic: This method relies solely on pixel patterns in the historical dataset. Recognition Conclusion: Because the vertical water stains in the middle of the column are extremely similar to cracks in the historical database in terms of pixel grayscale gradient, the pure data-driven model, lacking physical semantics, falsely reports it as a severe longitudinal crack. Recognition Performance: When faced with interference from light and shadow, vegetation, and stains, the false alarm rate is as high as 14.8%, and it cannot identify major structural hazard samples that only exist in simulations, easily resulting in missed detections.

[0030] Table 1: Comparison of data between embodiments of the present invention and comparative examples

[0031] By implementing the above steps, this invention not only achieves a statistically significant improvement in recognition accuracy, but also establishes an algorithm architecture with logical immunity at the physical mechanism level. Furthermore, during the implementation process, the system can pre-lock the damage-sensitive areas of components based on the stress field distribution output by physical mechanics simulation, thereby guiding the region of interest in the image to be recognized. In specific implementation, when processing massive inspection images, the system will prioritize gain compensation for pixel extraction branches in stress concentration areas such as the connecting beams of shear walls and the supports of frame beams, while performing lightweight processing on non-critical areas with minimal stress. This physical-guided feature extraction method significantly reduces the redundant computational overhead of the system, enabling the complex dual-drive recognition model to achieve second-level response on edge computing devices.

[0032] In the detailed implementation of cross-modal alignment, this invention uses a nonlinear projection matrix to normalize the weights of features from different modalities. The central processing unit trains a deep residual mapping network to transform the stress cloud map generated by simulation into a pixel-level probability distribution map, which serves as the mask layer for the pixel feature extraction branch. This semantic-level coupling operation ensures that each identified pixel has a corresponding mechanical response strength support.

[0033] When the physical logic determiner executes the embedded building structure service logic, the system can also combine the building's service cycle data. For example, for a specific concrete component, as the carbonation depth increases, the simulation module will dynamically adjust its brittle fracture parameters, and the generated extreme state samples will also evolve accordingly. This gives the recognition model time-varying performance, enabling it to adjust the recognition logic according to the building's actual aging state, and realize the reconstruction of the causal relationship between the building's appearance and internal mechanical evolution.

[0034] Finally, the system outputs the identification results and synchronizes them to the smart maintenance platform. This output process includes 3D BIM coordinate mapping, which directly links the identification conclusions of the apparent damage to the corresponding components in the building information model. When maintenance personnel click on the identified damage points, the system not only displays their image features, but also simultaneously displays the stress-strain distribution field map output by the physical semantic verification branch, clearly explaining the physical causes of the damage. This process elevates the originally isolated and static visual identification to a dynamic and logical structural diagnosis.

[0035] In summary, the present invention provides a building appearance damage identification method based on a dual-drive approach of historical data and simulation. By constructing a rigorous technical closed loop that includes a historical data sample library, a physical and mechanical digital twin model, a generative cross-modal alignment architecture, a dual-branch feature coupling model, and a physical logic decision-maker, this method completely solves the core technical bottlenecks that have long existed in the field of building inspection, such as the scarcity of extreme samples, the lack of mechanistic support for identification results, and the high false alarm rate. This method not only has outstanding substantive features at the theoretical level, but also demonstrates excellent generalization ability and judgment certainty in engineering practice.

[0036] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying building surface damage based on a dual-driven approach of historical data and simulation, characterized in that, Includes the following steps: Step 1: Collect historical appearance image data of existing buildings, perform multi-dimensional pixel feature extraction, and construct a basic damage pattern feature space containing image pixel features; Step 2: Obtain the physical parameters and service load environment of building components, construct a physical and mechanical digital twin model and perform damage evolution simulation to generate damage pattern data with mechanical cause logic; Step 3: Perform cross-modal feature space alignment using a generative architecture to associate and map the mechanical semantic features of the damage pattern data with the pixel features of the virtual damage image; Step 4: Construct a dual-drive feature coupling model consisting of a pixel feature extraction branch and a physical semantic verification branch, and map the pixel features of the image to be identified to the physical semantic constraint space determined by the physical mechanical digital twin model for feature coupling. Step 5: Based on the geometric-mechanical compatibility criterion of the damage pattern, perform physical logic judgment on the pixel set in the image to be identified, and output the damage identification result with physical semantic support.

2. The method for identifying building surface damage based on a dual-driven approach of historical data and simulation, as described in claim 1, is characterized in that... The specific process of performing multi-dimensional pixel feature extraction in step one is as follows: The central processing unit calls image processing algorithms to perform grayscale gradient enhancement processing on the acquired historical appearance image data, and uses adaptive histogram equalization technology to dynamically adjust the pixel distribution in local areas to enhance the contrast between the damaged area and the background wall. Edge detection is performed using first- or second-order differential operators to identify topological boundaries where pixel values ​​in an image transition, and to extract geometric features of apparent damage. Texture consistency analysis is performed using gray-level co-occurrence matrix or local binary mode algorithms to extract significant differences in spatial texture features of damaged areas; By eliminating noise interference caused by uneven ambient lighting and sensor viewing angle deviation through feature orthogonalization, the extracted geometric pattern features and spatial texture features are mapped to a high-dimensional feature vector space to construct the basic damage pattern feature space that establishes the statistical correlation between pixel features and known damage patterns.

3. The method for identifying building surface damage based on a dual-driven approach of historical data and simulation, as described in claim 1, is characterized in that... The specific process of constructing the physical and mechanical digital twin model and performing damage evolution simulation in step two is as follows: Obtain the geometric design parameters, material constitutive relations, and service load environment of the building components, and establish a high-fidelity finite element analysis model as the physical and mechanical digital twin model; The physical and mechanical digital twin model divides building components into computational grid cells of a preset density and assigns each computational grid cell a material property function including elastic modulus, yield strength and fracture toughness. The physical mechanics simulation module performs nonlinear damage mechanics deductions, applying boundary conditions including cyclic loads, seismic response loads, or environmental corrosion fields in the digital space. Based on the fracture mechanics energy release rate criterion, the evolution of displacement gradient is monitored. When the energy density of the local computational mesh element exceeds the preset damage threshold, the initiation, propagation and penetration paths of damage inside and on the surface of the component are simulated, and the damage pattern data including the stress tensor distribution field and displacement field are output.

4. The method for identifying building surface damage based on a dual-driven approach of historical data and simulation, as described in claim 1, is characterized in that... The specific process of performing cross-modal feature space alignment in step three is as follows: The central processing unit extracts ambient lighting, material surface roughness, and sensor noise characteristics from the basic damage pattern feature space; By using neural rendering technology or generative adversarial networks, the extracted features are superimposed onto the geometric model corresponding to the damage pattern data to generate a virtual damage image with realistic texture features. A joint probability distribution function is constructed to associate and map the pixel feature vector of the virtual damage image with the mechanical semantic features output by the physical and mechanical digital twin model, which include stress intensity factors and principal tensile stress directions, thereby achieving modal alignment between pixel visual features and physical mechanism features at the semantic level.

5. The method for identifying building surface damage based on a dual-driven approach of historical data and simulation, as described in claim 1, is characterized in that... The branching logic of the dual-drive feature coupling model described in step four is as follows: The pixel feature extraction branch is based on a deep convolutional neural network structure and uses multi-scale feature extraction logic to capture pixel gray-level gradients, geometric topology and texture distribution features in the image to be identified, and to identify potential damage candidate regions. The physical semantic verification branch is synchronously connected to the physical and mechanical digital twin model to extract in real time the stress distribution characteristics, displacement field trend and physical damage sensitive area distribution map of the component under the current service cycle and load state. The central processing unit performs tensor splicing and coupling operations on the dual-branch features, and uses a nonlinear projection matrix to transform the feature vector of the pixel feature extraction branch and the physical semantic vector of the physical semantic verification branch to the same feature plane, thereby establishing a physical constraint recognition criterion.

6. The method for identifying building surface damage based on a dual-driven approach of historical data and simulation, as described in claim 5, is characterized in that... During the execution of step four, the dual-drive feature coupling model introduces a weight adjustment mechanism based on physical semantics: Determine whether the potential damage candidate region identified by the pixel feature extraction branch is located in the stress core region or physical damage sensitive region determined by the physical and mechanical digital twin model; If the potential damage candidate region is located in the stress core region or the physical damage sensitive region, the physical semantic verification branch increases the sensitivity weight coefficient of the region pixel recognition. If the potential damage candidate region is located in a stress-safe zone that does not produce damage in terms of physical mechanics and logic, the central processing unit automatically raises the judgment threshold and reduces redundant computational overhead through physical-guided feature extraction.

7. The method for identifying building surface damage based on a dual-driven approach of historical data and simulation, as described in claim 1, is characterized in that... The geometric-mechanical compatibility criterion and determination process described in step five are as follows: When a set of pixels suspected of being cracks is identified, the central processing unit obtains the spatial orientation, position coordinates, and geometric parameters of the set of pixels. The physical semantic verification branch calculates the angle between the suspected crack propagation direction and the principal stress axis of the component at the location based on the physical and mechanical digital twin model. Verify whether the included angle between the principal stress axes conforms to the orthogonal development criterion or the slip development criterion of fracture mechanics; If the geometric topology of a suspected crack does not conform to the logic of physical and mechanical evolution, then the pixel set is determined to be a non-damaging interference target.

8. The method for identifying building surface damage based on a dual-driven approach of historical data and simulation, as described in claim 7, is characterized in that... During the judgment process in step five, a secondary verification logic is triggered for suspected cracks: The secondary verification logic analyzes the displacement pattern of the geometric center of the suspected damaged pixel set as ambient light changes by calling multiple frames of images for time series comparison. If the geometric center of the pixel set shifts in a non-mechanically compliant manner due to ambient light, it is identified as light and shadow noise, water seepage, or epiphytic vegetation interference, and a pixel culling operation is performed.

9. The method for identifying building surface damage based on a dual-driven approach of historical data and simulation, as described in claim 2, is characterized in that... During the cross-modal feature space alignment process, the central processing unit constructs a joint feature space: The high-dimensional visual feature vectors at the pixel level and the stress-strain tensor at the physical level are transformed by using a nonlinear projection matrix, so that they can perform nonlinear correlation operations in the same feature plane. The reinforcement learning algorithm is used to correct the projection weights of the feature vector mapping in real time based on the deviation of the measured feedback, so that the system can identify and evaluate the physical causes of unknown damage patterns not included in the historical pattern database through physical mechanism inference.

10. The method for identifying building surface damage based on a dual-driven approach of historical data and simulation, as described in claim 1, is characterized in that... The damage identification results supported by physical semantics are specifically included in the following: The recognition results are synchronized to the corresponding component nodes of the Building Information Model (BIM) through three-dimensional spatial coordinate mapping. The output includes a comprehensive maintenance report containing damage type, location coordinates, geometric parameters, and mechanical cause assessment data derived from the physical semantic verification branch; It displays the pixel features of suspected damage points and their corresponding digital twin model stress-strain distribution field map in real time, realizing the logical traceability of the identification results.

Images