A method and system for monitoring building cracks based on machine learning

Abstract

This invention discloses a machine learning-based method and system for monitoring building cracks, belonging to the field of building inspection technology. The method includes the following steps: S1, multi-band image data enhancement and preprocessing; S2, adaptive crack region segmentation; S3, vectorized representation of crack features; S4, crack feature learning based on self-supervised learning; S5, crack generation based on Generative Adversarial Network (GAN); S6, crack morphological dynamics modeling; and S7, crack topology analysis based on graph structures. This machine learning-based method and system for monitoring building cracks effectively improves the accuracy and robustness of crack monitoring through adaptive segmentation, self-supervised learning, and crack morphological dynamics modeling. It solves the problems of false detection and missed detection in traditional methods. The system can adaptively learn the crack behavior under different environments and materials, accurately extract crack features, and predict their propagation paths.

Description

Technical Field

[0001] The present invention relates to the technical field of building detection, and in particular to a method and system for monitoring building cracks based on machine learning. Background Art

[0002] Crack monitoring of buildings plays a vital role in structural safety maintenance. As buildings age, external environmental factors such as wind, rain, temperature changes, and earthquakes may cause cracks on the surface of the building, thus affecting its overall structural safety. Traditional crack monitoring methods rely on manual inspection or simple edge detection and threshold segmentation techniques. However, these methods often have great limitations when dealing with complex crack shapes, lighting and material changes in changing environments, and are unable to fully and accurately detect and predict crack expansion.

[0003] Although there has been progress in the existing crack detection technology based on machine learning, it is mainly limited to the insufficient accuracy of crack area segmentation: conventional image processing or segmentation algorithms based on fixed thresholds are difficult to cope with the differences in crack performance under different building materials and lighting conditions, resulting in inaccurate crack area segmentation results and serious false detection and missed detection. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a building crack monitoring method and system based on machine learning to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a building crack monitoring method based on machine learning, comprising the following steps:

[0006] S1, multi-band image data enhancement and preprocessing;

[0007] By collecting images of buildings from different angles, multi-band data enhancement is performed, and images of different bands of ultraviolet, infrared and visible light are used to enrich crack feature information. By fusing images, multi-layer image data containing rich information is formed.

[0008] S2, adaptive segmentation of crack area;

[0009] Adopting an adaptive segmentation model based on deep learning to adaptively segment multi-layer image data, the adaptive segmentation model is a fully convolutional network FCN or U-Net, which uses a convolutional neural network CNN to automatically learn the features of the crack area. During model training, based on the building crack data set, the multi-scale convolution feature hierarchy is used to capture cracks of different scales.

[0010] S3, quantized representation of crack features;

[0011] After detecting cracks of different scales, the morphological information of the cracks is converted into feature vectors for machine learning models;

[0012] S4, crack feature learning based on self-supervised learning;

[0013] Based on the self-supervised learning model, the transformed feature vectors are pre-trained to allow the model to learn the structure and manifestation of cracks, use unlabeled data to improve the understanding of crack characteristics, and enhance the generalization performance in different scenarios;

[0014] S5. Crack generation based on image generation adversarial network GAN;

[0015] Generative adversarial networks (GANs) are used to process multi-layer image data to synthesize crack images. GANs generate realistic crack samples. By generating crack images and comparing them with real crack images, the model learns more abstract crack features.

[0016] S6, crack morphology dynamics modeling;

[0017] Through the formula model of physics and mechanics, the expansion trend of cracks in buildings under different external stress conditions is simulated, and the physical model is combined with the machine learning model to construct a dynamic evolution model of cracks;

[0018] S7, Crack topology analysis based on graph structure;

[0019] By constructing the topological structure of the cracks, the correlation between the cracks is analyzed.

[0020] To further optimize the technical solution, in step S2, the adaptive segmentation model of the crack area is used for accurate crack area segmentation by combining feature extraction and image processing technology of deep learning;

[0021] In the adaptive segmentation model, it is assumed that the input image is , represents a two-dimensional image, and are the coordinates of the image, and the goal of the model is to find a set of functions , used to automatically segment crack regions from images;

[0022] The adaptive segmentation model includes a multi-scale feature extraction layer, an adaptive weight assignment mechanism, a crack contour enhancement mechanism, and a crack area prediction output.

[0023] To further optimize the technical solution, in the adaptive segmentation model:

[0024] Multi-scale feature extraction layer: capture crack features of different scales through multi-scale convolution operations and define the input image The characteristics are ,in Representing the scale level, the convolution operation is defined as:

[0025] ;

[0026] in, Indicates convolution kernels, represents the convolution operation, is the number of scale layers;

[0027] Through multi-scale convolution, the feature representation of cracks at different detail levels is obtained;

[0028] Adaptive weight allocation mechanism: Based on the characteristics of the cracks, different scale levels of capture are allocated. The mechanism is as follows:

[0029] ;

[0030] in, is the adaptive weight, representing the importance of different scales, It is the nonlinear activation function ReLU;

[0031] Adaptive weights dynamically adjust the weights of each scale level, making the model pay more attention to the significant features of the crack area;

[0032] Crack contour enhancement mechanism: by combining the Sobel operator Extract crack edge information and combine it with adaptive weight features:

[0033]

[0034] in, The edge image extracted by the Sobel operator, is a tuning parameter used to balance the weight of adaptive weight and edge information;

[0035] This mechanism ensures that the boundaries of the cracks are identified with greater precision;

[0036] Crack area prediction output: The final crack area segmentation result is obtained through the activation function Perform binarization to obtain the segmentation mask of the crack :

[0037] ;

[0038] Through the above processing, The output value is in the range of [0,1], indicating the probability value of each pixel belonging to a crack. By selecting an appropriate threshold, the probability map is converted into a binary mask map to identify the crack area on the building surface.

[0039] Further optimizing the technical solution, in step S3, by extracting the geometric features of the length, width, depth, direction, and shape complexity of the crack, and calculating the texture information of the crack, the Gabor filter extracts the texture features, and the chromaticity histogram extracts the color change features;

[0040] Time series analysis is introduced to use the data of building cracks changing over time to extract the characteristics of crack expansion speed and crack growth trend.

[0041] Further optimizing the technical solution, in step S4, in the self-supervised learning model, it is assumed that there is an unlabeled crack image set , where each For a building image containing cracks, deep crack features are learned through contrastive learning tasks and crack structure preservation tasks;

[0042] For each input image , and generate two different enhanced views, respectively denoted as and , represents the performance of the same crack under different conditions. The goal is to let the model learn that the crack features in these two views are the same;

[0043] The loss function of contrastive learning is defined as:

[0044]

[0045] in,

[0046] It is a feature extraction network that maps the image to the feature space;

[0047] is the cosine similarity function, which measures the similarity between two images in the feature space;

[0048] is the temperature parameter used to adjust the sensitivity of contrastive learning;

[0049] In the crack structure preservation task, the crack structure is modeled using a graph structure. It is assumed that a set of crack structure points is extracted from the image. , the structural points represent the key locations of the cracks, and the Euclidean distance between each pair of points is defined as:

[0050] ;

[0051] Enhanced view of the same crack and The point set on and For comparison, the model keeps the distance relationship between these points unchanged, and the loss function for crack structure preservation is defined as:

[0052] ;

[0053] By keeping the relative distance between the key points of the cracks unchanged, the model can learn the characteristics of the cracks while retaining the geometric structure of the cracks.

[0054] To further optimize the technical solution, the self-supervised learning model combines two tasks, contrastive learning and structure preservation, and the total loss function of the model is defined as:

[0055]

[0056] in, is the balance coefficient, which is used to control the weights of contrastive learning and structure preservation tasks; is the loss function for contrastive learning; Loss function for fracture structure preservation;

[0057] Based on this total loss function, the model not only learns the similarities of cracks under different perspectives and lighting conditions, but also maintains the consistency of the geometric structure of the cracks, thereby improving the robustness of feature learning.

[0058] To further optimize the technical solution, in step S6, in the dynamic evolution model of the crack, the stress field of a certain area of ​​the building is assumed to be , represents the stress value of the region over time, and the expansion of the crack under the action of material stress is modeled by the energy release rate and the crack expansion rate;

[0059] The dynamic equation of the stress field is as follows:

[0060]

[0061] in,

[0062] Represents the effect of external loads on the stress field, which changes with time and building structure;

[0063] is the temperature field, is the temperature influence factor, which indicates the effect of temperature change on stress;

[0064] is the stress attenuation coefficient, which indicates the rate at which stress inside the material is released;

[0065] The driving force of crack propagation is the energy release rate caused by the stress field The energy release rate reflects the energy released per unit area of ​​the crack under stress;

[0066] Crack growth rate Determined by the relationship between energy release rate and material crack resistance, when When the crack expands at a preset speed, is the critical energy release rate, which represents the minimum energy required for crack propagation.

[0067] To further optimize the technical solution, the dynamic evolution model of the crack, under the action of the stress field and the energy release rate, the expansion path of the crack is represented by the following dynamic equation:

[0068]

[0069] in,

[0070] represents the position of the crack front as a function of time;

[0071] is the crack growth rate;

[0072] is the normal vector of the crack front, indicating the direction of crack extension;

[0073] The crack path evolution equation combines the crack propagation speed and direction information to simulate the dynamic propagation process of the crack under the action of material stress.

[0074] To further optimize the technical solution, in step S7, first, the building surface is regarded as a two-dimensional or three-dimensional space grid, and the cracks are regarded as edges or nodes in the grid;

[0075] Then, the graph convolutional network (GCN) is used to analyze the correlation between cracks and find potential crack extension paths or key crack points.

[0076] By analyzing the topological structure, we can identify which cracks will become the starting point or key point of structural damage, thus providing a scientific basis for crack repair.

[0077] A machine learning-based building crack monitoring system is constructed based on the above-mentioned machine learning-based building crack monitoring method, including a crack area adaptive segmentation module, a crack feature extraction and analysis module, a crack feature learning module, a crack morphology dynamics modeling module, and a model training and updating module;

[0078] The crack area adaptive segmentation module adopts deep learning technology to automatically segment the input crack image into crack areas, and accurately separates the cracks from the background in complex scenes through an adaptive segmentation algorithm;

[0079] The crack feature extraction and analysis module analyzes the segmented crack area and extracts the geometric features of the crack, including length, width, shape, and edge features;

[0080] The crack feature learning module learns the deep features of cracks through a self-supervised learning model. Through contrast learning and structure preservation tasks, the system can better understand the manifestations of cracks in different environments.

[0081] The crack morphology dynamics modeling module, combined with the machine learning model, simulates the dynamic expansion process of cracks over time and predicts future changes of cracks through stress fields, energy release rates and crack expansion rates;

[0082] The model training and updating module continuously trains the self-supervised learning model and the dynamic evolution model through regularly updated data to improve the accuracy of crack detection and prediction.

[0083] Compared with the prior art, the present invention provides a building crack monitoring method and system based on machine learning, which has the following beneficial effects:

[0084] This machine learning-based building crack monitoring method and system effectively improves the accuracy and robustness of crack monitoring through adaptive segmentation, self-supervised learning, and crack morphology dynamics modeling, and solves the problems of false detection and missed detection in crack detection by traditional methods. The system can adaptively learn the manifestation of cracks in different environments and materials, accurately extract crack features and predict their expansion paths. Through physical modeling of stress fields and crack expansion rates, the system can also dynamically track the evolution of cracks, significantly improving the prediction accuracy of crack expansion and providing more reliable technical guarantees for building structure safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 A flow chart of a building crack monitoring method based on machine learning proposed by the present invention;

[0086] Figure 2 A schematic diagram of the architecture of an adaptive segmentation model in a building crack monitoring method based on machine learning proposed by the present invention;

[0087] Figure 3 This is a structural schematic diagram of a building crack monitoring system based on machine learning proposed by the present invention. DETAILED DESCRIPTION

[0088] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0089] Embodiment 1:

[0090] See also Figure 1 , a building crack monitoring method based on machine learning, comprising the following steps:

[0091] S1. Multi-band image data enhancement and preprocessing

[0092] By collecting images of buildings at different angles, multi-band data enhancement is performed, and images of different bands of ultraviolet, infrared and visible light are used to enrich crack feature information.

[0093] In this embodiment, multi-band imaging technology can capture subtle changes in buildings that are difficult to see with the naked eye. By fusing these images, multi-layer image data containing rich information is formed. In the preprocessing stage, the image needs to be denoised, contrast enhanced, and normalized so that the subsequent steps can better identify crack details. This step can significantly improve the diversity and robustness of the data, making the model more generalizable.

[0094] S2. Adaptive segmentation of crack regions

[0095] An adaptive segmentation model based on deep learning, such as a fully convolutional network (FCN) or U-Net, is used to automatically learn the features of the crack area using a convolutional neural network (CNN). During model training, based on a dataset of building cracks, a multi-scale convolutional feature hierarchy is used to capture cracks of different scales.

[0096] In this embodiment, in step S2, the adaptive segmentation model of the crack region is used for accurate crack region segmentation by combining feature extraction and image processing technology of deep learning;

[0097] In the adaptive segmentation model, it is assumed that the input image is , represents a two-dimensional image, and are the coordinates of the image, and the goal of the model is to find a set of functions , used to automatically segment crack regions from images;

[0098] like Figure 2 As shown, the adaptive segmentation model includes a multi-scale feature extraction layer, an adaptive weight allocation mechanism, a crack contour enhancement mechanism, and a crack area prediction output.

[0099] Furthermore, in the adaptive segmentation model:

[0100] Multi-scale feature extraction layer: capture crack features of different scales through multi-scale convolution operations and define the input image The characteristics are ,in Representing the scale level, the convolution operation is defined as:

[0101] ;

[0102] in, Indicates convolution kernels, represents the convolution operation, is the number of scale layers;

[0103] Through multi-scale convolution, the feature representation of cracks at different detail levels is obtained.

[0104] In the crack detection process, cracks vary in shape and size, and traditional single-scale convolution may not be able to capture all the details. The multi-scale convolution extraction layer can obtain detailed information about cracks from multiple levels, ensuring that cracks of different sizes and locations can be effectively captured. For example, wider cracks may be identified in the large-scale feature layer, while smaller cracks can be captured by small-scale convolution kernels. In this way, the model can adaptively handle the diversity of cracks.

[0105] Adaptive weight allocation mechanism: Based on the characteristics of the cracks, different scale levels of capture are allocated. The mechanism is as follows:

[0106] ;

[0107] in, is the adaptive weight, representing the importance of different scales, It is a nonlinear activation function such as ReLU;

[0108] Adaptive weights dynamically adjust the weights of each scale level, allowing the model to pay more attention to the salient features of the crack area.

[0109] The cracks in each image are different. In some scenes, coarse-scale information may be more important, while in other scenes, fine-scale features are crucial. The adaptive weight mechanism automatically assigns weights to convolution results of different scales by learning the crack features in the image, ensuring that the model can flexibly adjust the focus of feature extraction according to the actual crack situation. This module improves the robustness of the model, enabling it to cope with complex and changing crack morphologies.

[0110] Crack contour enhancement mechanism: by combining the Sobel operator Extract crack edge information and combine it with adaptive weight features:

[0111] ;

[0112] in, The edge image extracted by the Sobel operator, is a tuning parameter used to balance the weight of adaptive weight and edge information;

[0113] This mechanism ensures that the boundaries of the cracks are determined with greater precision.

[0114] Edge detection is particularly important in crack detection because the boundary area of ​​the crack is often the key part of segmentation. By introducing the Sobel operator to extract the edge and combining it with the adaptive feature weight, it is ensured that the model does not lose the boundary information during the segmentation process. The role of this mechanism is to enhance the boundary accuracy of the crack, especially for small and fuzzy crack areas, which can further improve the clarity of the segmentation results.

[0115] Crack area prediction output: The final crack area segmentation result is obtained through the activation function Perform binarization to obtain the segmentation mask of the crack :

[0116] ;

[0117] Through the above processing, The output value is in the range of [0,1], indicating the probability value of each pixel belonging to a crack. By selecting an appropriate threshold, the probability map is converted into a binary mask map to identify the crack area on the building surface.

[0118] The final prediction output uses the sigmoid function to map features to probability space. The value of each pixel represents the probability that the pixel belongs to a crack. By setting a threshold, these probability values ​​can be converted into binary segmentation results. This step ensures the interpretability of the segmentation results, and the threshold can also be adjusted according to the specific application scenario to control the sensitivity of crack detection.

[0119] The adaptive weight distribution mechanism and crack contour enhancement mechanism ensure accurate segmentation of the crack area. Compared with the traditional fixed threshold or simple convolution model, this model can dynamically adjust the level of crack feature extraction to adapt to different crack forms and scenarios. At the same time, the edge-based enhancement mechanism can improve the segmentation accuracy of the crack boundary, which plays an important role in improving the accuracy of the detection results.

[0120] S3. Vectorized representation of crack features

[0121] After detecting the crack area, the morphological information of the crack is converted into a feature vector for the machine learning model.

[0122] In this embodiment, in step S3, the geometric features of the length, width, depth, direction, and shape complexity of the crack are extracted, and the texture information of the crack, such as the texture features extracted by the Gabor filter, and the color change features, such as the chromaticity histogram, are calculated;

[0123] Time series analysis is introduced to use the data of building cracks changing over time to extract the characteristics of crack expansion speed and crack growth trend.

[0124] S4. Crack feature learning based on self-supervised learning

[0125] Based on the self-supervised learning model, the model learns the structure and manifestation of cracks through pre-training tasks, uses unlabeled data to improve the ability to understand crack characteristics, and enhances the generalization performance in different scenarios.

[0126] In this embodiment, in step S4, in the self-supervised learning model, it is assumed that there is an unlabeled crack image set , where each For a building image containing cracks, deep crack features are learned through contrastive learning tasks and crack structure preservation tasks;

[0127] For each input image , and generate two different enhanced views, respectively denoted as and , represents the performance of the same crack under different conditions. The goal is to let the model learn that the crack features in these two views are the same;

[0128] The loss function of contrastive learning is defined as:

[0129]

[0130] in,

[0131] It is a feature extraction network that maps the image to the feature space;

[0132] is the cosine similarity function, which measures the similarity between two images in the feature space;

[0133] is the temperature parameter used to adjust the sensitivity of contrastive learning.

[0134] The similarity between different enhanced views of the same crack in the feature space is maximized, while the similarity between views of other cracks and it is minimized.

[0135] In crack detection, the appearance of cracks in an image will change with different viewing angles, lighting, and even materials. By designing contrastive learning tasks, the model learns the invariance of cracks in the feature space, and can identify the features of the same crack even under different conditions. Contrastive learning enhances the contrast of views, allowing the model to effectively capture the core features of cracks in an unsupervised manner, maximize the similarity between different views of the same crack, and minimize the confusion of other cracks.

[0136] In the crack structure preservation task, the crack structure is modeled using a graph structure. It is assumed that a set of crack structure points is extracted from the image. , the structural points represent the key locations of the cracks, and the Euclidean distance between each pair of points is defined as:

[0137] ;

[0138] Enhanced view of the same crack and The point set on and For comparison, the model keeps the distance relationship between these points unchanged, and the loss function for crack structure preservation is defined as:

[0139] ;

[0140] By keeping the relative distance between the key points of the cracks unchanged, the model can learn the characteristics of the cracks while retaining the geometric structure of the cracks.

[0141] The geometric morphology and structural information of cracks are crucial for crack monitoring. By designing the crack structure preservation task, the model can learn the spatial geometric relationship of cracks. For example, the distance between the key points of the same crack should remain consistent under different views or enhancement conditions. This task ensures that the model captures the overall characteristics of the crack while not ignoring the microstructural changes of the crack. Ultimately, the model is able to show stronger generalization capabilities in more complex scenarios, especially in architectural environments with diverse crack morphologies.

[0142] The self-supervised learning model combines two tasks, contrastive learning and structure preservation, and the total loss function of the model is defined as:

[0143] ;

[0144] in, is the balance coefficient, which is used to control the weights of contrastive learning and structure preservation tasks; is the loss function for contrastive learning; Loss function for fracture structure preservation;

[0145] Based on this total loss function, the model not only learns the similarities of cracks under different perspectives and lighting conditions, but also maintains the consistency of the geometric structure of the cracks, thereby improving the robustness of feature learning.

[0146] The combination of contrastive learning and crack structure preservation tasks is introduced in the existing crack detection scenario, which is an innovation of the traditional supervised learning method. In the absence of a large amount of labeled data, through self-supervised learning, the model can automatically learn the invariance and geometric structure of cracks, greatly improving its understanding and generalization ability of crack characteristics. Especially in complex and changeable scenarios such as building crack monitoring, this method has strong adaptability and practical application value.

[0147] S5. Crack Generation Based on Image Generative Adversarial Network (GAN)

[0148] A generative adversarial network (GAN) is used to synthesize crack images. GAN generates realistic crack samples. By generating crack images and comparing them with real crack images, the model learns more abstract crack features.

[0149] In this embodiment, in order to improve the robustness of the model to crack features, a generative adversarial network (GAN) can be used to synthesize crack images. GAN can generate realistic crack samples, making the types of cracks in the training set richer, especially supplementing rare complex crack samples. The generated cracks can not only be used to expand the training data set, but also to enhance the model's ability to recognize different crack forms. By generating crack images and comparing them with real crack images, the model can learn more abstract crack features, so that it has higher detection accuracy in practical applications.

[0150] S6. Crack morphology dynamics modeling

[0151] Through the formula models of physics and mechanics, the expansion trend of cracks in buildings under different external stress conditions is simulated. The physical model is combined with the machine learning model to construct a dynamic evolution model of cracks.

[0152] In this embodiment, in step S6, in the dynamic evolution model of cracks, the stress field of a certain area of ​​the building is assumed to be , represents the stress value of the region over time, and the expansion of the crack under the action of material stress is modeled by the energy release rate and the crack expansion rate;

[0153] The dynamic equation of the stress field is as follows:

[0154] ;

[0155] in,

[0156] Represents the effect of external loads on the stress field, which changes with time and building structure;

[0157] is the temperature field, is the temperature influence factor, which indicates the effect of temperature change on stress;

[0158] is the stress attenuation coefficient, which indicates the rate at which stress inside the material is released;

[0159] The driving force of crack propagation is the energy release rate caused by the stress field The energy release rate reflects the energy released per unit area of ​​the crack under stress as follows:

[0160] ;

[0161] in, is the energy release constant of the material, indicating the energy threshold required for crack expansion inside the material. The energy release rate is the key factor in determining the crack expansion speed. Exceeding a certain threshold The crack will expand.

[0162] Crack growth rate It is determined by the relationship between the energy release rate and the material's crack resistance, as follows:

[0163]

[0164] when When , the crack expands at a certain speed, where is the critical energy release rate, which indicates the minimum energy required for crack propagation; It is the speed factor of crack extension, which depends on the material's crack resistance and crack morphology.

[0165] The formula shows that the crack extension speed increases with the increase of energy release rate, but only when Once the critical value is exceeded, the crack will begin to expand.

[0166] Furthermore, in the dynamic evolution model of the crack, under the action of the stress field and the energy release rate, the crack expansion path is represented by the following dynamic equation:

[0167] ;

[0168] in,

[0169] represents the position of the crack front as a function of time;

[0170] is the crack growth rate;

[0171] is the normal vector of the crack front, indicating the direction of crack extension;

[0172] The crack path evolution equation combines the crack propagation speed and direction information to simulate the dynamic propagation process of the crack under the action of material stress.

[0173] Through the crack path evolution equation, the model can simulate the dynamic expansion process of cracks and predict the possible shape and direction of cracks in the future. For example, a crack in a building may expand in the direction of maximum stress and gradually form a complex branching structure.

[0174] Compared with traditional static crack monitoring methods, this model can dynamically track the evolution of cracks in the time dimension and provide more accurate early warning and prediction functions.

[0175] S7. Crack topology analysis based on graph structure

[0176] By constructing the topological structure of the cracks, the correlation between the cracks is analyzed.

[0177] In this embodiment, in step S7, first, the building surface is regarded as a two-dimensional or three-dimensional space grid, and the cracks are regarded as edges or nodes in the grid;

[0178] Then, a graph convolutional network (GCN) is used to analyze the correlation between cracks and find potential crack extension paths or key crack points.

[0179] By analyzing the topological structure, we can identify which cracks will become the starting point or key point of structural damage, thus providing a scientific basis for crack repair.

[0180] Embodiment 2:

[0181] See also Figure 3 A building crack monitoring system based on machine learning is constructed based on the building crack monitoring method based on machine learning described in Example 1, including a crack area adaptive segmentation module, a crack feature extraction and analysis module, a crack feature learning module, a crack morphology dynamics modeling module, and a model training and updating module.

[0182] The crack area adaptive segmentation module adopts deep learning technology to automatically segment the input crack image into crack areas, and accurately separates the cracks from the background in complex scenes through an adaptive segmentation algorithm.

[0183] Using a convolutional neural network trained with deep learning (such as U-Net), the crack areas in the building are identified and segmented through multi-scale feature extraction. The details of cracks of different sizes are captured to ensure that the system can effectively detect cracks of different sizes.

[0184] The crack feature extraction and analysis module analyzes the segmented crack area and extracts the geometric features of the crack, such as length, width, shape, and edge features.

[0185] Based on image processing technology, the geometric parameters of the cracks, such as the length, width, and depth of the cracks, are calculated. Edge detection algorithms such as Canny are used to capture the outline of the cracks and further analyze the complex morphology of the cracks. Based on the morphological characteristics of the cracks, the model can classify whether the cracks are linear, curved, or bifurcated.

[0186] The crack feature learning module learns the deep features of cracks through a self-supervised learning model. Through comparative learning and structure preservation tasks, the system can better understand the manifestations of cracks in different environments.

[0187] Contrastive learning task: The model learns the deep features of the crack by comparing enhanced views of the same crack under different viewpoints or lighting conditions.

[0188] Fracture structure preservation task: By maintaining the consistency of the fracture geometry, the model can maintain its understanding of the fracture structure in different environments and improve generalization ability.

[0189] The crack morphology dynamics modeling module, combined with a machine learning model, simulates the dynamic expansion process of cracks over time and predicts future changes in cracks through stress fields, energy release rates, and crack expansion rates.

[0190] The stress distribution in different parts of the building is simulated by the stress field dynamics equation.

[0191] Based on the changes in the stress field, the energy release rate of crack extension is calculated, and the crack extension rate is predicted in combination with the material properties.

[0192] The possible future expansion direction and morphological changes of the cracks are predicted by the expansion path equation.

[0193] The model training and updating module continuously trains the self-supervised learning model and the dynamic evolution model through regularly updated data to improve the accuracy of crack detection and prediction.

[0194] When new crack data is input into the system, the model automatically updates its parameters through incremental training to maintain the latest performance. Combined with the newly collected crack data, the loss function of contrastive learning and structure preservation tasks is optimized to further improve the model capabilities.

[0195] The beneficial effects of the present invention are:

[0196] This machine learning-based building crack monitoring method and system effectively improves the accuracy and robustness of crack monitoring through adaptive segmentation, self-supervised learning, and crack morphology dynamics modeling, and solves the problems of false detection and missed detection in crack detection by traditional methods. The system can adaptively learn the manifestation of cracks in different environments and materials, accurately extract crack features and predict their expansion paths. Through physical modeling of stress fields and crack expansion rates, the system can also dynamically track the evolution of cracks, significantly improving the prediction accuracy of crack expansion and providing more reliable technical guarantees for building structure safety management.

[0197] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0198] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A building crack monitoring method based on machine learning, characterized in that: The following steps are involved: S1, multi-band image data enhancement and preprocessing; By collecting images of buildings from different angles, multi-band data enhancement is performed, and images of different bands of ultraviolet, infrared and visible light are used to enrich crack feature information. By fusing images, multi-layer image data containing rich information is formed. S2, adaptive segmentation of crack area; Adopting an adaptive segmentation model based on deep learning to adaptively segment multi-layer image data, the adaptive segmentation model is a fully convolutional network FCN or U-Net, which uses a convolutional neural network CNN to automatically learn the features of the crack area. During model training, based on the building crack data set, the multi-scale convolution feature hierarchy is used to capture cracks of different scales. S3, quantized representation of crack features; After detecting cracks of different scales, the morphological information of the cracks is converted into feature vectors for machine learning models; S4, crack feature learning based on self-supervised learning; Based on the self-supervised learning model, the transformed feature vectors are pre-trained to allow the model to learn the structure and manifestation of cracks, use unlabeled data to improve the understanding of crack characteristics, and enhance the generalization performance in different scenarios; S5. Crack generation based on image generation adversarial network GAN; Generative adversarial networks (GANs) are used to process multi-layer image data to synthesize crack images. GANs generate realistic crack samples. By generating crack images and comparing them with real crack images, the model learns more abstract crack features. S6, crack morphology dynamics modeling; Through the formula model of physics and mechanics, the expansion trend of cracks in buildings under different external stress conditions is simulated, and the physical model is combined with the machine learning model to construct a dynamic evolution model of cracks; S7, Crack topology analysis based on graph structure; By constructing the topological structure of the cracks, the correlation between the cracks is analyzed.

2. A building crack monitoring method based on machine learning according to claim 1, characterized in that: In step S2, the adaptive segmentation model of the crack area is used for accurate crack area segmentation by combining feature extraction and image processing technology of deep learning; In the adaptive segmentation model, it is assumed that the input image is , represents a two-dimensional image, and are the coordinates of the image, and the goal of the model is to find a set of functions , used to automatically segment crack regions from images; The adaptive segmentation model includes a multi-scale feature extraction layer, an adaptive weight assignment mechanism, a crack contour enhancement mechanism, and a crack area prediction output.

3. A building crack monitoring method based on machine learning according to claim 2, characterized in that: In the adaptive segmentation model: Multi-scale feature extraction layer: capture crack features of different scales through multi-scale convolution operations and define the input image The characteristics are ,in Representing the scale level, the convolution operation is defined as: ； in, Indicates convolution kernels, represents the convolution operation, is the number of scale layers; Through multi-scale convolution, the feature representation of cracks at different detail levels is obtained; Adaptive weight allocation mechanism: Based on the characteristics of the cracks, different scale levels of capture are allocated. The mechanism is as follows: ； in, is the adaptive weight, representing the importance of different scales, It is the nonlinear activation function ReLU; Adaptive weights dynamically adjust the weights of each scale level, making the model pay more attention to the significant features of the crack area; Crack contour enhancement mechanism: by combining the Sobel operator Extract crack edge information and combine it with adaptive weight features: ； in, The edge image extracted by the Sobel operator, is a tuning parameter used to balance the weight of adaptive weight and edge information; This mechanism ensures that the boundaries of the cracks are identified with greater precision; Crack area prediction output: The final crack area segmentation result is obtained through the activation function Perform binarization to obtain the segmentation mask of the crack : ； Through the above processing, The output value is in the range of [0,1], indicating the probability value of each pixel belonging to a crack. By selecting an appropriate threshold, the probability map is converted into a binary mask map to identify the crack area on the building surface.

4. A building crack monitoring method based on machine learning according to claim 1, characterized in that: In step S3, the geometric features of the length, width, depth, direction, and shape complexity of the crack are extracted, and the texture information of the crack is calculated. The texture features are extracted by using a Gabor filter, and the color change features are extracted by using a chromaticity histogram. Time series analysis is introduced to use the data of building cracks changing over time to extract the characteristics of crack expansion speed and crack growth trend.

5. The building crack monitoring method based on machine learning according to claim 1 is characterized in that: In step S4, in the self-supervised learning model, it is assumed that there is an unlabeled crack image set , where each For a building image containing cracks, deep crack features are learned through contrastive learning tasks and crack structure preservation tasks; For each input image , and generate two different enhanced views, respectively denoted as and , represents the performance of the same crack under different conditions. The goal is to let the model learn that the crack features in these two views are the same; The loss function of contrastive learning is defined as: ； in, It is a feature extraction network that maps the image to the feature space; is the cosine similarity function, which measures the similarity between two images in the feature space; is the temperature parameter used to adjust the sensitivity of contrastive learning; In the crack structure preservation task, the crack structure is modeled using a graph structure. It is assumed that a set of crack structure points is extracted from the image. , the structural points represent the key locations of the cracks, and the Euclidean distance between each pair of points is defined as: ； Enhanced view of the same crack and The point set on and For comparison, the model keeps the distance relationship between these points unchanged, and the loss function for crack structure preservation is defined as: ； By keeping the relative distance between the key points of the cracks unchanged, the model can learn the characteristics of the cracks while retaining the geometric structure of the cracks.

6. A building crack monitoring method based on machine learning according to claim 5, characterized in that: The self-supervised learning model combines two tasks, contrastive learning and structure preservation, and the total loss function of the model is defined as: ; in, is the balance coefficient, which is used to control the weights of contrastive learning and structure preservation tasks; is the loss function for contrastive learning; Loss function for fracture structure preservation; Based on this total loss function, the model not only learns the similarities of cracks under different perspectives and lighting conditions, but also maintains the consistency of the geometric structure of the cracks, thereby improving the robustness of feature learning.

7. A building crack monitoring method based on machine learning according to claim 1, characterized in that: In step S6, in the dynamic evolution model of cracks, the stress field of a certain area of ​​the building is assumed to be , represents the stress value of the region over time, and the expansion of the crack under the action of material stress is modeled by the energy release rate and the crack expansion rate; The dynamic equation of the stress field is as follows: ； in, Represents the effect of external loads on the stress field, which changes with time and building structure; is the temperature field, is the temperature influence factor, which indicates the effect of temperature change on stress; is the stress attenuation coefficient, which indicates the rate at which stress inside the material is released; The driving force of crack propagation is the energy release rate caused by the stress field The energy release rate reflects the energy released per unit area of ​​the crack under stress; Crack growth rate Determined by the relationship between energy release rate and material crack resistance, when When the crack expands at a preset speed, is the critical energy release rate, which represents the minimum energy required for crack propagation.

8. A building crack monitoring method based on machine learning according to claim 7, characterized in that: In the dynamic evolution model of the crack, under the action of stress field and energy release rate, the crack expansion path is expressed by the following dynamic equation: ； in, represents the position of the crack front as a function of time; is the crack growth rate; is the normal vector of the crack front, indicating the direction of crack extension; The crack path evolution equation combines the crack propagation speed and direction information to simulate the dynamic propagation process of the crack under the action of material stress.

9. The building crack monitoring method based on machine learning according to claim 1, characterized in that: In step S7, first, the building surface is regarded as a two-dimensional or three-dimensional space grid, and the cracks are regarded as edges or nodes in the grid; Then, the graph convolutional network (GCN) is used to analyze the correlation between cracks and find potential crack extension paths or key crack points. By analyzing the topological structure, we can identify which cracks will become the starting point or key point of structural damage, thus providing a scientific basis for crack repair.

10. A machine learning-based building crack monitoring system, constructed based on the machine learning-based building crack monitoring method according to any one of claims 1 to 9, characterized in that: It includes the module of adaptive segmentation of crack area, the module of extraction and analysis of crack features, the module of learning crack features, the module of modeling of crack morphology and dynamics, and the module of model training and updating; The crack area adaptive segmentation module adopts deep learning technology to automatically segment the input crack image into crack areas, and accurately separates the cracks from the background in complex scenes through an adaptive segmentation algorithm; The crack feature extraction and analysis module analyzes the segmented crack area and extracts the geometric features of the crack, including length, width, shape, and edge features; The crack feature learning module learns the deep features of cracks through a self-supervised learning model. Through contrast learning and structure preservation tasks, the system can better understand the manifestations of cracks in different environments. The crack morphology dynamics modeling module, combined with the machine learning model, simulates the dynamic expansion process of cracks over time and predicts future changes of cracks through stress fields, energy release rates and crack expansion rates; The model training and updating module continuously trains the self-supervised learning model and the dynamic evolution model through regularly updated data to improve the accuracy of crack detection and prediction.

Images