Intelligent Identification Method and System for Quality Defects of Hidden Projects in Power Infrastructure

Abstract

The intelligent defect identification method and system for hidden projects in power infrastructure construction of the present application relate to the technical field of intelligent defect identification. By constructing a three-dimensional point cloud model, time-series data and visible light images for structural health monitoring are obtained; time-series features are extracted based on the time-series data, the visible light images are input into a defect segmentation model to obtain a defect probability map, geometric features of the defect region are generated based on the defect probability map, and the time-series features and geometric features are fused to obtain spatio-temporal features; an acceleration vector sequence for the expansion of the defect region is obtained based on the spatio-temporal features and defect pattern recognition is performed; the expansion direction vector of the defect region is determined according to the geometric features, the principal stress direction vector under the action of an external load is obtained, and the risk degree of the expansion of the defect region is judged; the defect state of the defect region is output according to the spatio-temporal features of the pixels in the defect region, and abnormal decision-making is carried out in combination with the defect pattern and the risk degree of the expansion of the defect region, realizing intelligent identification of defect detection.

Description

Technical Field

[0001] This application relates to the technical field of intelligent defect recognition, and particularly to an intelligent recognition method and system for quality defects in hidden projects of power infrastructure construction. Background Art

[0002] Transmission towers are key components in the power transmission system, and their structural health status directly affects the safe and stable operation of the power grid. However, due to the long-term exposure of transmission towers to the external environment, various defect problems such as cracks, rust, and deformation occur frequently, threatening transmission reliability. Traditional structural health monitoring of transmission towers mainly relies on manual regular inspections, which have problems such as missed inspections and misjudgments, and it is difficult to detect early signs of defects in a timely manner, resulting in a lag in maintenance decisions and passive risk prevention and control.

[0003] In recent years, with the rapid development of sensor technology and artificial intelligence, device health management driven by multi-source heterogeneous data has become possible. In the field of transmission tower structure monitoring, on the one hand, technologies such as lidar scanning and high-definition photography can obtain the three-dimensional topography information of the tower base; on the other hand, technologies such as fiber optic sensing and acoustic emission can collect dynamic signals such as strain and vibration of the structure in real time. At the same time, meteorological stations, line load monitoring devices, etc. can also provide external environment parameters such as wind speed, wind direction, and conductor tension. These multi-source data from different physical fields contain rich defect feature information.

[0004] Chinese patent application with publication number CN118133087A discloses a building engineering quality monitoring method and device based on artificial intelligence and big data. The method includes: collecting data to be monitored during the construction or use of the target building structure, and fusing the data to generate a vector to be recognized; inputting the vector to be recognized into a pre-constructed feature extraction model, and extracting the target features of the vector to be recognized through the feature extraction model; inputting the target features into a pre-constructed classifier model, and outputting a classification recognition result through the classifier model; the classification recognition result includes the building engineering quality grade corresponding to the target features; determining the quality status of the target building structure according to the classification recognition result.

[0005] The monitoring strategies of existing methods are fixed and cannot dynamically optimize the decision logic according to the defect evolution stage, resulting in resource waste or missed detection of key information. Summary of the Invention

[0006] This application aims to solve at least one of the technical problems in the related technologies. To this end, an object of this application is to propose an intelligent recognition method and system for quality defects in hidden projects of power infrastructure construction, realizing intelligent recognition of defect detection.

[0007] One aspect of this application provides an intelligent recognition method for quality defects in hidden projects of power infrastructure construction, including:

[0008] Step S100: Construct a three-dimensional point cloud model, obtain time-series data for structural health monitoring, and visible light images of the surface texture of the transmission tower base;

[0009] Step S200: Extract time-series features based on the time-series data, input the visible light image into a defect segmentation model to obtain a defect probability map, generate geometric features of the defect region based on the defect probability map, and fuse the time-series features and geometric features to obtain spatio-temporal features;

[0010] Step S300: Obtain an acceleration vector sequence for the expansion of the defect region based on the spatio-temporal features, and use the acceleration vector sequence for defect pattern recognition;

[0011] Step S400: Determine the expansion direction vector of the defect region according to the geometric features, obtain the principal stress direction vector under the action of the external load, and judge the risk degree of the expansion of the defect region;

[0012] Step S500: Output the defect state of the defect region according to the spatio-temporal features of the pixels in the defect region, and make an abnormal decision in combination with the defect pattern and the risk degree of the expansion of the defect region;

[0013] The specific method for constructing the three-dimensional point cloud model, obtaining the time-series data for structural health monitoring, and visible light images of the surface texture of the transmission tower base is as follows:

[0014] Step S110: Obtain three-dimensional point cloud data by laser radar scanning the transmission tower base and its surrounding environment. The attributes of each point include the three-dimensional coordinates and reflection intensity of the point, and generate a three-dimensional point cloud model of the transmission tower base based on the three-dimensional point cloud data;

[0015] Step S120: Obtain visible light images of the surface texture of the transmission tower base;

[0016] Step S130: Collect strain data, vibration data, and acoustic emission data through sensors in the transmission tower base, and integrate the time series of the strain data, vibration data, and acoustic emission data into time-series data for structural health monitoring;

[0017] Step S140: Take the center of the bottom surface of the transmission tower base as the origin, construct a three-dimensional space coordinate system, and synchronize the time of the time-series data for structural health monitoring and unify it into the three-dimensional space coordinate system;

[0018] The specific method for extracting time-series features based on the time-series data, inputting the visible light image into a defect segmentation model to obtain a defect probability map, generating geometric features of the defect region based on the defect probability map, and fusing the time-series features and geometric features to obtain spatio-temporal features is as follows:

[0019] Step S210: Perform time-frequency domain analysis on the time-series data of structural health monitoring, extract time-domain features and frequency-domain features, and perform feature fusion on the time-domain features and frequency-domain features to obtain time-series features;

[0020] Step S220: Align the visible light image and the 3D point cloud model, and establish a 3D point cloud model with texture information between the pixels of the visible light image and the 3D point cloud model;

[0021] Step S230: Use a convolutional neural network introducing prior physical knowledge for defect segmentation. The input data is the visible light image aligned with the 3D point cloud model, and the output data is the defect probability map. Calculate the geometric features of the defect area according to the defect probability map; the geometric features include the main direction of the defect area;

[0022] Step S240: Align the geometric features and the time-series features according to the time stamp, and associate them with the 3D spatial coordinates of the corresponding 3D point cloud model to obtain spatio-temporal features; the spatio-temporal features include the 3D spatial coordinates of the center of the defect area, the time stamp, and the geometric features and time-series features corresponding to the time stamp;

[0023] The specific method of using a convolutional neural network introducing prior physical knowledge for defect segmentation, where the input data is the visible light image aligned with the 3D point cloud model, the output data is the defect probability map, and calculating the geometric features of the defect area according to the defect probability map is as follows:

[0024] Design an end-to-end convolutional neural network model for defect segmentation. Input the visible light image aligned with the 3D point cloud model, label the defect labels for the visible light image to obtain a defect mask, construct a training sample based on the visible light image and the corresponding defect mask, and define minimizing the segmentation loss function as the training objective of the convolutional neural network model; Incorporate prior physical knowledge into the defect segmentation model, define a regularization term based on partial differential equations, and add the regularization term to the segmentation loss function to obtain the total segmentation loss function of the convolutional neural network introducing prior physical knowledge; Use the training sample to train the convolutional neural network introducing prior physical knowledge by minimizing the total segmentation loss function to obtain a defect segmentation model; Take the current visible light image aligned with the 3D point cloud model as the input data, output the defect probability map through the defect segmentation model, segment the defect area based on thresholding operation, and calculate the geometric features of the defect area;

[0025] The calculation method of the main direction is as follows:

[0026] Step S231: Use the 3D spatial coordinates corresponding to all pixels within a defect area as a 2D point cloud to form a coordinate matrix of B×2; B is the number of pixels in the defect area;

[0027] Step S232: Decentralize the coordinate matrix to obtain the coordinate matrix after mean removal;

[0028] Step S233: Calculate the covariance matrix of the coordinate matrix after mean removal;

[0029] Step S234: Perform eigenvalue decomposition on the covariance matrix, and take the eigenvector corresponding to the largest eigenvalue as the main direction of the defect area;

[0030] The specific method for obtaining the acceleration vector sequence of defect area expansion based on spatio-temporal features and using the acceleration vector sequence for defect pattern recognition is as follows:

[0031] Step S310: For the spatio-temporal features of the same defect area collected at consecutive tn data sampling times, form a spatio-temporal feature sequence, and calculate the displacement vector and velocity vector of the three-dimensional spatial coordinates of the defect area over time based on the spatio-temporal features of adjacent data sampling times;

[0032] Step S320: Based on the spatio-temporal feature sequence, obtain the displacement vector sequence and velocity vector sequence of the defect area, and calculate the acceleration vector sequence of defect area expansion according to the velocity vectors between adjacent data sampling times;

[0033] Step S330: Use a long short-term memory network to perform defect pattern recognition on the acceleration vector sequence of defect area expansion; the defect patterns include normal patterns and abnormal patterns;

[0034] The specific method for using a long short-term memory network to perform defect pattern recognition on the acceleration vector sequence of defect area expansion is as follows:

[0035] Use a long short-term memory network for defect pattern recognition, with the input being the acceleration vector sequence of defect area expansion and the output being the defect pattern corresponding to the data sampling time. Define a binary cross-entropy loss function, and use the historical acceleration vector sequence with defect pattern labels to perform supervised learning training on the long short-term memory network; when the output for consecutive m data sampling times is an abnormal pattern, it is determined that there is a risk of structural instability.

[0036] The specific method for determining the expansion direction vector of the defect area according to geometric features, obtaining the principal stress direction vector under external load, and judging the danger degree of defect area expansion is as follows:

[0037] Step S410: For the sequence of three-dimensional spatial coordinates of the defect area at consecutive tn data sampling times, obtain the main direction of this defect area as the expansion direction vector;

[0038] Step S420: Obtain the principal stress direction vector under the current external load;

[0039] Step S430: Calculate the spatial angle between the expansion direction vector of the defect area and the principal stress direction vector;

[0040] Step S440: Judge the consistency between the expansion direction vector of the defect area and the principal stress direction vector according to the spatial angle threshold, and determine the danger degree of the expansion of the defect area;

[0041] The specific method for judging the consistency between the expansion direction vector of the defect area and the principal stress direction vector according to the spatial angle threshold and determining the danger degree of the expansion of the defect area is as follows: Set the spatial angle threshold. When the spatial angle is less than the spatial angle threshold, it is determined that the defect area is expanding in the high-risk direction. Otherwise, the defect area is considered safe. The danger degree of the expansion of the defect area includes the high-risk direction and safety;

[0042] The specific method for outputting the defect state according to the spatio-temporal characteristics of the pixels in the defect area and making an abnormal decision by combining the defect mode and the danger degree of the expansion of the defect area is as follows:

[0043] Step S510: Conduct a clustering analysis on the spatio-temporal characteristics of the defect area in the historical time, divide the defect state into three stages: the germination stage, the stable stage and the diffusion stage, obtain the clustering model, input the spatio-temporal characteristics of the current defect area pixels, and output the corresponding defect state;

[0044] Step S520: Adaptively adjust the data adoption frequency of the defect area according to the output of the clustering model;

[0045] Step S530: Make an abnormal decision according to the defect state, defect mode and danger degree of the expansion of the defect area.

[0046] One aspect of the present application provides an intelligent identification system for quality defects in hidden projects of power infrastructure construction, including:

[0047] A multi-modal data collection module, which is used to construct a three-dimensional point cloud model, obtain the time-series data of structural health monitoring and the visible light image of the surface texture of the transmission tower foundation;

[0048] A spatio-temporal feature fusion module, which extracts time-series features based on the time-series data, inputs the visible light image into the defect segmentation model to obtain a defect probability map, generates geometric features of the defect area based on the defect probability map, and fuses the time-series features and the geometric features to obtain spatio-temporal features;

[0049] A defect mode recognition module, which obtains the acceleration vector sequence of the expansion of the defect area based on the spatio-temporal features, and uses the acceleration vector sequence for defect mode recognition;

[0050] A danger degree judgment module, which determines the expansion direction vector of the defect area according to the geometric features, obtains the principal stress direction vector under the action of external loads, and judges the danger degree of the expansion of the defect area;

[0051] The decision-making comprehensive judgment module outputs the defect status according to the spatiotemporal characteristics of the pixels in the defect area, and makes abnormal decisions based on the defect mode and the degree of danger of the defect area expansion.

[0052] One aspect of the present application provides a readable storage medium storing a computer program suitable for loading by a processor to execute the steps of a method for intelligently identifying quality defects of hidden power infrastructure projects.

[0053] The intelligent identification method and system for quality defects of concealed power infrastructure projects proposed in this application have the following advantages over existing technologies:

[0054] This application extracts discriminant features related to defects in multimodal sensor data and fuses them with three-dimensional spatial information to obtain a unified spatiotemporal feature representation. It uses physical prior knowledge to guide the convolutional neural network model for feature learning, realizing an organic combination of data-driven and knowledge-driven approaches and improving the accuracy and robustness of defect segmentation.

[0055] This application characterizes the spatial morphology of the defect by extracting the geometric features of the defect area, takes the main direction of the defect area as its expansion direction vector, reveals the local expansion trend of the defect, introduces the main stress direction vector under the action of external load, combines defect analysis with structural force analysis, explains the evolution behavior of the defect from the perspective of physical mechanism, and provides quantitative indicators for assessing the degree of danger of the defect.

[0056] This application calculates the displacement vector and velocity vector of the defect area based on the spatiotemporal feature sequence, characterizes the dynamic evolution process of the defect, further calculates the acceleration vector sequence, introduces high-order physical quantities to describe the dynamic behavior of the defect, and uses the LSTM network to perform defect pattern recognition on the acceleration vector sequence, realizing abnormal detection and early warning of defect evolution. Through modeling and pattern recognition of the dynamic evolution of defects, early warning of structural instability risks is achieved.

[0057] This application clusters the spatiotemporal characteristics of historical defects, adaptively divides defect states, reveals the inherent stages of defect evolution, dynamically adjusts data sampling frequency according to defect states, realizes self-optimization of monitoring strategies, comprehensively considers defect states, defect modes and danger levels, and constructs a multi-factor fusion abnormal decision-making mechanism. Through hierarchical classification of state modeling and abnormal decision-making, a global assessment from local to overall, from state to trend is achieved, which solves the problem of dynamically optimizing the decision-making logic of defect repair according to the defect evolution stage, and improves the sensitivity and reliability of defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1Flowchart of the intelligent identification method for quality defects of hidden projects in power infrastructure construction provided by this application;

[0059] Figure 2 Flowchart of the fusion method for spatio-temporal features provided by this application;

[0060] Figure 3 Flowchart of the defect pattern recognition method provided by this application;

[0061] Figure 4 Schematic diagram of the extended direction vector and the principal stress direction vector provided by this application;

[0062] Figure 5 Functional module diagram of the intelligent identification system for quality defects of hidden projects in power infrastructure construction provided by this application. Detailed implementation manners

[0063] To better understand this application, more detailed descriptions of various aspects of this application will be made with reference to the accompanying drawings. It should be understood that these detailed descriptions are only descriptions of the exemplary embodiments of this application and do not limit the scope of this application in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0064] In the accompanying drawings, for the sake of convenience of illustration, the sizes, dimensions and shapes of the elements have been slightly adjusted. The accompanying drawings are only examples and are not drawn strictly to scale. As used herein, terms such as "substantially", "about" and similar terms are used as terms indicating approximation and not as terms indicating degree, and are intended to illustrate the inherent deviations in measured values or calculated values that would be recognized by those of ordinary skill in the art. Additionally, in this application, the order of description of the processing steps does not necessarily represent the order in which these processes occur in actual operation, unless there are clear other limitations or can be deduced from the context.

[0065] It should also be understood that expressions such as "comprising", "including", "having", "containing" and / or "including having" are open-ended rather than closed-ended expressions in this specification, which indicate the existence of the stated features, elements and / or components, but do not exclude the existence of one or more other features, elements, components and / or their combinations. In addition, when an expression such as "at least one of..." appears after a list of listed features, it modifies the entire list of features rather than just an individual element in the list. In addition, when describing the embodiments of this application, the use of "may" means "one or more embodiments of this application". And, the term "exemplary" is intended to refer to an example or illustration.

[0066] Unless otherwise defined, all terms used herein (including engineering and scientific terms) shall have the same meaning as commonly understood by those of ordinary skill in the art to which this application pertains. It should also be understood that, unless explicitly stated in this application, words defined in common dictionaries shall be interpreted as having a meaning consistent with their meaning in the context of the relevant art, and shall not be interpreted in an idealized or overly formal sense.

[0067] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments may be combined with each other. The following will detail this application by referring to the drawings and in combination with the embodiments.

[0068] Embodiment 1

[0069] As Figure 1 shown, the intelligent identification method for quality defects of hidden projects in power infrastructure provided by this application includes:

[0070] Step S100: Construct a three-dimensional point cloud model, and obtain the time-series data of structural health monitoring and the visible light image of the surface texture of the transmission tower foundation;

[0071] The specific method for constructing the three-dimensional point cloud model and obtaining the time-series data of structural health monitoring and the visible light image of the surface texture of the transmission tower foundation is as follows:

[0072] Step S110: Obtain three-dimensional point cloud data by laser radar scanning the transmission tower foundation and its surrounding environment. The attributes of each point include the three-dimensional coordinates and reflection intensity of the point, and generate a three-dimensional point cloud model of the transmission tower foundation based on the three-dimensional point cloud data;

[0073] The three-dimensional point cloud data is used to reconstruct the three-dimensional geometric model of the transmission tower foundation, providing a data basis for subsequent defect location.

[0074] Step S120: Obtain the visible light image of the surface texture of the transmission tower foundation;

[0075] The visible light image of the surface texture of the transmission tower foundation is obtained by using a high-resolution camera or a drone to photograph the surface texture of the transmission tower foundation;

[0076] The three-dimensional point cloud data and the visible light image are synchronously collected and matched, so that in subsequent defect analysis, the defect information in the image space can be accurately mapped to the three-dimensional physical space, so that the position, size and shape of the defect can be intuitively displayed on the three-dimensional point cloud model, and correlated analysis can be performed with the time-series data of subsequent structural health monitoring.

[0077] Step S130: Collect strain data, vibration data, and acoustic emission data through sensors in the transmission tower foundation, and integrate the time series of strain data, vibration data, and acoustic emission data into the time series data for structural health monitoring;

[0078] The strain data can be expressed as: ; is the strain data collected at the th data sampling time;

[0079] The vibration data can be expressed as:

[0080] ; is the vibration data collected at the th data sampling time;

[0081] The acoustic emission data can be expressed as:

[0082] ; is the acoustic emission data collected at the th data sampling time;

[0083] Among them, represents the sensor number, is the data sampling time;

[0084] The time series data for structural health monitoring are the time series signal data such as strain data, vibration data, and acoustic emission data obtained through sensors such as strain gauges, accelerometers, and ultrasonic flaw detectors, which are used to reflect the dynamic response of the structure under external loads to analyze the damage and deterioration state of the structure.

[0085] Furthermore, the acquisition frequencies and sampling times of multi-modal time series data need to be synchronized. Time synchronization is achieved by embedding a high-precision clock module in the sensor nodes to ensure that the data collected by different sensors are aligned in the time dimension for multi-modal information fusion and time series analysis.

[0086] Step S140: Taking the center of the bottom surface of the transmission tower foundation as the origin, construct a three-dimensional space coordinate system, and synchronize the time series data for structural health monitoring and unify them into the three-dimensional space coordinate system;

[0087] The three-dimensional space coordinates in the three-dimensional space coordinate system can be expressed as (x, y, z).

[0088] Synchronizing the time series data for structural health monitoring and unifying them into the three-dimensional space coordinate system is to achieve the spatio-temporal consistency of the data, which is convenient for subsequent multi-modal data fusion and defect location.

[0089] The data acquisition process of the above steps designs the synchronous acquisition and spatio-temporal alignment of multi-source heterogeneous data, which is the basis for subsequent defect analysis and prediction. Among them, the three-dimensional point cloud data provides an accurate geometric model of the transmission tower base, the visible light image provides the visual features of surface defects, and the time-series data of structural health monitoring reflects the physical quantity indicators of the structural health state. These data complement each other and jointly constitute comprehensive health monitoring information of the transmission tower base.

[0090] Step S200: Extract time-series features based on the time-series data, input the visible light image into the defect segmentation model to obtain the defect probability map, generate the geometric features of the defect region based on the defect probability map, and fuse the time-series features and the geometric features to obtain spatio-temporal features;

[0091] As Figure 2 shown, the specific method for extracting time-series features based on the time-series data, inputting the visible light image into the defect segmentation model to obtain the defect probability map, generating the geometric features of the defect region based on the defect probability map, and fusing the time-series features and the geometric features to obtain spatio-temporal features is as follows:

[0092] Step S210: Perform time-frequency domain analysis on the time-series data of structural health monitoring, extract time-domain features and frequency-domain features, and fuse the time-domain features and the frequency-domain features to obtain time-series features;

[0093] The time-domain features include the mean and variance of the time-series data, which respectively reflect the average amplitude and fluctuation degree of the time-series data. The frequency-domain features include the Fourier transform coefficients of the time-series data and the power spectral density of the time-series data, which are respectively used to transform the time-domain signal to the frequency domain and reflect the energy distribution of the time-series data on different frequency components;

[0094] The calculation formula for the mean of the time-series data is:

[0095] ; where represents the time-series data at the data sampling time ti, tn is the last data sampling time, represents the mean of the time-series data sx;

[0096] The calculation formula for the variance of the time-series data is: ; where represents the variance of the time-series data sx;

[0097] The calculation formula for the Fourier transform coefficients of the time-series data is: ; where represents the Fourier transform coefficients of the time-series data sx, t represents the data sampling time variable of the time-domain signal, f represents the frequency variable of the frequency-domain signal, and j is the imaginary unit, satisfying , represents the time series data of the data sampling time t;

[0098] The calculation formula for the power spectral density of the time series data is: ; where represents the power spectral density of the time series data sx, and TS is the total duration of the time domain signal;

[0099] The time series feature is represented as , and is obtained by combining multiple time-frequency domain features;

[0100] Step S220: Align the visible light image and the three-dimensional point cloud model to establish a three-dimensional point cloud model with texture information between the pixels of the visible light image and the three-dimensional point cloud model;

[0101] The specific method for aligning the visible light image and the three-dimensional point cloud model to establish a three-dimensional point cloud model with texture information between the pixels of the visible light image and the three-dimensional point cloud model is:

[0102] Step S221: Input the visible light image and the three-dimensional point cloud model, and extract image features and point cloud features from the visible light image and the three-dimensional point cloud model;

[0103] The feature descriptor used for the extraction of local feature points is SIFT, and the process of extracting the image features of the visible light image I is represented as , and the process of extracting the point cloud features of the three-dimensional point cloud model P is represented as ; where , respectively represent the SIFT feature vectors of the visible light image and the three-dimensional point cloud model, is the image feature, is the point cloud feature;

[0104] Step S222: Perform feature matching on the image features and the point cloud features to obtain an initial feature matching set;

[0105] Step S223: Remove outliers from the initial feature matching set to obtain an inlier matching set, and according to the inlier matching set, solve the optimal transformation matrix from the image features to the point cloud features through the least squares optimization algorithm ;

[0106] The RANSAC method is used for outlier removal, and the transformation matrix is ; where T is the transformation model from the image features to the point cloud features, and respectively represent the coordinates of the image feature points and the three-dimensional coordinates of the point cloud feature points, is the indicator function, is the threshold for interior point determination; the argmax function represents searching for the transformation matrix that makes the value maximum, represents the Euclidean distance, represents the Euclidean distance between the image feature points and the point cloud feature points obtained by projecting through the transformation model T, represents that the mi-th image feature and the mk-th point cloud feature are a pair of matching pairs, is the feature matching set between the image features and the point cloud features;

[0107] Step S224: Map each pixel point in the visible light image to the coordinates of the corresponding point in the three-dimensional point cloud model according to the transformation matrix, and obtain a three-dimensional point cloud model with texture information ;

[0108] Step S230: Perform defect segmentation using a convolutional neural network introducing prior physical knowledge. The input data is the visible light image aligned with the three-dimensional point cloud model, and the output data is the defect probability map. Calculate the geometric features of the defect area according to the defect probability map; the geometric features include the main direction of the defect area;

[0109] Preferably, the convolutional neural network model selects the U-Net model;

[0110] The specific method of performing defect segmentation using a convolutional neural network introducing prior physical knowledge, where the input data is the visible light image aligned with the three-dimensional point cloud model, the output data is the defect probability map, and calculating the geometric features of the defect area according to the defect probability map is as follows:

[0111] Design an end-to-end convolutional neural network model for defect segmentation. Input the visible light image I aligned with the three-dimensional point cloud model, and label the defect labels for the visible light image to obtain the defect mask , construct training samples based on the visible light image and the corresponding defect mask, and define minimizing the segmentation loss function as the training objective of the convolutional neural network model; incorporate prior physical knowledge into the defect segmentation model, define a regularization term based on partial differential equations, add the regularization term to the segmentation loss function, and obtain the total segmentation loss function of the convolutional neural network introducing prior physical knowledge; use the training samples to train the convolutional neural network introducing prior physical knowledge by minimizing the total segmentation loss function to obtain the defect segmentation model; use the current visible light image aligned with the three-dimensional point cloud model as the input data, output the defect probability map through the defect segmentation model, segment the defect area based on thresholding operation, and calculate the geometric features of the defect area;

[0112] The segmentation loss function has the following calculation formula: ; Where N is the number of training samples, and are the height and width of the visible light image, n≤N, i≤H, k≤W, and Represents the pixel row and column index in the visible light image, Represents pixels in a visible light image The probability of being predicted as a defect, is the parameter of the convolutional neural network model; the segmentation loss function is the binary cross entropy loss at the pixel level. is the pixel of the visible light image in the nth training sample Defect mask of

[0113] The partial differential equation is a defect expansion partial differential equation derived based on the Paris law, and the equation expression is: ,in, represents the characteristic function of the defect area, represents the gradient of the characteristic function u of the defect area, and is a constant related to material properties;

[0114] The total segmentation loss function The calculation formula is: ; in, is the regularization coefficient, is the visible light image domain;

[0115] The defect area obtained by segmentation based on thresholding operation refers to setting a probability threshold, and for the output defect probability map , where pixels with a probability greater than the probability threshold are determined as defective areas, the defect mask of the defective area is 1, and the defect mask of the non-defective area is 0, and a binary defect mask is obtained. ;

[0116] The main direction is calculated as follows:

[0117] Step S231: The three-dimensional spatial coordinates corresponding to all pixels in a defect area are used as a two-dimensional point cloud to form a B×2 coordinate matrix;

[0118] The coordinate matrix Expressed as:

[0119] , where B is the number of pixels in the defect area; is the row index of the pixel in the defect area, is the column index of the pixel in the defect area;

[0120] Step S232: Decentralize the coordinate matrix to obtain the coordinate matrix after mean removal;

[0121] The coordinate matrix after mean removal is:

[0122] ; where and are the means of each column of the coordinate matrix respectively;

[0123] Step S233: Calculate the covariance matrix of the coordinate matrix after mean removal;

[0124] The covariance matrix of the coordinate matrix after mean removal has the following calculation formula:

[0125] ; where and represent the variances of the row index ' and column index of the coordinate matrix after mean removal respectively, represents the covariance of the row index ' and column index of the coordinate matrix after mean removal; is 's transpose;

[0126] Step S234: Perform eigenvalue decomposition on the covariance matrix, and take the eigenvector corresponding to the largest eigenvalue as the main direction of the defect region;

[0127] The calculation formula for performing eigenvalue decomposition on the covariance matrix is: , where is a diagonal matrix, and the elements on its diagonal are the eigenvalues and of the covariance matrix, is an orthogonal matrix, and its column vectors and are the eigenvectors corresponding to and respectively, and they satisfy the relationship ;

[0128] The step of taking the eigenvector corresponding to the largest eigenvalue as the main direction of the defect region includes: taking the eigenvector and corresponding to the largest eigenvalue among the eigenvalues on the diagonal of the diagonal matrix as the main direction of the defect region.

[0129] The main direction of the defective area can be expressed as:

[0130] , where represents the angle between the main direction and the horizontal direction;

[0131] The main direction captures the direction of the maximum variance in the pixel distribution of the defective area, reflecting the overall extension trend of the defect. The other eigenvector orthogonal to it represents the secondary extension direction of the defect.

[0132] By solving the main direction through the principal component analysis method, a quantitative direction feature can be extracted from the segmented defective area, providing a data basis for subsequent defect classification, recognition, and evolution analysis. For example, for crack-like defects, its main direction can indicate the trend of crack propagation, and for corrosion-like defects, the main direction may be related to the direction of stress concentration. Combining the direction feature with geometric features such as area, perimeter, and major axis length results in a comprehensive description of the defect morphology, providing more abundant information for automated defect analysis.

[0133] The geometric features also include the area, perimeter, and major axis perimeter of the defective area;

[0134] The area of the defective area The calculation formula is: ;

[0135] The calculation formula for the perimeter of the defective area is: ;

[0136] where

[0137] represents the difference between the defect mask of the current pixel and its left pixel, represents the difference between the defect mask of the current pixel and its upper pixel; represents the pixel predicted defect mask; <

[0138] If the defect mask values of two pixels are different, it means there is a vertical boundary, and the corresponding side length is 1, otherwise the side length is 0; similarly, the difference between the defect mask of the current pixel and its upper pixel is used to determine whether there is a horizontal boundary between two pixels;

[0139] The major axis length The calculation formula is: ;

[0140] [[ID=,55]]Step S240: Align the geometric features and the temporal features according to the timestamp and associate them with the three-dimensional spatial coordinates of the corresponding three-dimensional point cloud model to obtain spatio-temporal features;

[0141] The spatio-temporal features include the three-dimensional spatial coordinates of the center of the defect area, the time stamp, and the geometric features and temporal features corresponding to the time stamp.

[0142] Exemplarily, assuming that a crack image is captured on the concrete surface of a certain transmission tower foundation, the traditional CNN model may only be able to segment the general outline of the crack, while the CNN model with the physical regularization term can estimate the crack propagation direction and morphology according to the Paris law, making the segmentation result more refined and accurate. At the same time, the physical regularization term can also compensate for the problem of insufficient training data, enabling the CNN model to learn reasonable defect features even in small-sample scenarios.

[0143] Step S300: Obtain the acceleration vector sequence of the defect area expansion based on the spatio-temporal features, and use the acceleration vector sequence for defect pattern recognition;

[0144] As Figure 3 shown, the specific method for obtaining the acceleration vector sequence of the defect area expansion based on the spatio-temporal features and using the acceleration vector sequence for defect pattern recognition is as follows:

[0145] Step S310: For the spatio-temporal features of the same defect area collected at consecutive tn data sampling times, form a spatio-temporal feature sequence, and calculate the displacement vector and velocity vector of the three-dimensional spatial coordinates of the defect area over time according to the spatio-temporal features at adjacent data sampling times;

[0146] The spatio-temporal features extracted at the ti-th data sampling time can be expressed as , where is the three-dimensional spatial coordinates of the center of the defect area, and are the geometric features and temporal features at the ti-th data sampling time respectively;

[0147] The calculation methods for the displacement vector and velocity vector are as follows:

[0148] For between the ti-th data sampling time and the ti+1-th data sampling time, calculate the displacement vector and velocity vector of the defect area. The displacement vector is calculated by the formula:

[0149] ; The velocity vector is calculated by the formula: ; where is the three-dimensional spatial coordinates of the center of the defect area at the ti+1-th data sampling time, is the time difference between adjacent data sampling times;

[0150] Step S320: Obtain the displacement vector sequence and velocity vector sequence of the defect area based on the spatio-temporal feature sequence, and calculate the acceleration vector sequence of the defect area expansion according to the velocity vectors between adjacent data sampling times;

[0151] The acceleration vector is calculated by the formula: ; where is the velocity vector at the (ti + 1)-th data sampling time;

[0152] Step S330: Use a long short-term memory network to perform defect pattern recognition on the acceleration vector sequence of the defect area expansion; the defect patterns include normal patterns and abnormal patterns;

[0153] The specific method of using a long short-term memory network to perform defect pattern recognition on the acceleration vector sequence of the defect area expansion is as follows:

[0154] Use a long short-term memory network for defect pattern recognition. The input is the acceleration vector sequence of the defect area expansion, and the output is the defect pattern corresponding to the data sampling time. Define a binary cross-entropy loss function, and use the historical acceleration vector sequence with defect pattern labels to perform supervised learning training on the long short-term memory network. When the output for m consecutive data sampling times is an abnormal pattern, it is determined that there is a risk of structural instability.

[0155] The acceleration vector sequence can be expressed as representing the acceleration vector corresponding to the data sampling time ;

[0156] By introducing the concept of defect acceleration, the above steps construct physical quantities describing the dynamic evolution of defects, and use an LSTM network to learn their normal and abnormal patterns, realizing early warning of structural instability.

[0157] Exemplarily, a crack defect was found in the foundation of a transmission tower during continuous monitoring. Through the analysis in step S300, it was found that the crack had a significant accelerating expansion trend in the past week, and its acceleration vector continued to increase in a certain direction. The LSTM network recognized its acceleration sequence, so the instability risk of the crack was timely warned. Compared with the traditional static evaluation method based on defect size, this application identified the instability signs of the crack earlier by analyzing the dynamic evolution trend of the defect, and gained a valuable time window for subsequent maintenance decisions. This reflects the innovative role of step S300 in early warning of defects.

[0158] Step S400: Determine the expansion direction vector of the defect area according to geometric features, obtain the principal stress direction vector under external load, and judge the danger degree of the defect area expansion;

[0159] The specific method for determining the expansion direction vector of the defect area according to geometric features, obtaining the principal stress direction vector under the action of external loads, and judging the risk degree of the expansion of the defect area is as follows:

[0160] Step S410: For the sequence of three-dimensional space coordinates of the defect area at consecutive tn data sampling times, obtain the principal direction of the defect area as the expansion direction vector ;

[0161] Step S420: Obtain the principal stress direction vector under the action of the current external load;

[0162] The method for obtaining the principal stress direction vector under the action of the current external load is as follows: Use finite element analysis software to calculate the stress of the transmission tower foundation structure, obtain the stress distribution nephogram under the action of the actual external load, and take the stress direction of the element with the maximum stress as the principal stress direction vector of the structure .

[0163] Step S430: Calculate the expansion direction vector of the defect area and the principal stress direction vector to obtain the spatial angle between them;

[0164] The expansion direction vector of the defect area and the principal stress direction vector The calculation formula for the spatial angle between them is: ;

[0165] where, represents the vector modulus length, represents the inverse cosine function, and the spatial angle .

[0166] Step S440: Judge the consistency between the expansion direction vector of the defect area and the principal stress direction vector according to the spatial angle threshold, and determine the risk degree of the expansion of the defect area;

[0167] The specific method for judging the consistency between the expansion direction vector of the defect area and the principal stress direction vector according to the spatial angle threshold and determining the risk degree of the expansion of the defect area is as follows: Set the spatial angle threshold , when the spatial angle is less than the spatial angle threshold, it is determined that the defect area is expanding towards the high-risk direction, otherwise it is considered that the defect area is safe. The risk degree of the expansion of the defect area includes the high-risk direction and safety;

[0168] Furthermore, for the defect area determined to be expanding towards the high-risk direction, the maintenance priority of this defect area needs to be increased.

[0169] Preferably, the spatial angle threshold is 30°;

[0170] The above steps explore the spatial correlation between the defect expansion direction and the structural force direction, combine defect monitoring with load analysis, explain the evolution behavior of defects from a physical mechanism perspective, and guide the formulation of maintenance decisions.

[0171] like Figure 4 As shown, this is a schematic diagram of the expansion direction vector and the principal stress direction vector provided by this application. The gray cube in the figure represents the transmission tower base, and the center of the bottom surface of the transmission tower base is the origin. A three-dimensional spatial coordinate system is constructed. There is a black crack on the transmission tower base, which is a simple schematic diagram of the defect area. There are two arrows with an angle of 15° on the right side of the figure, where the red arrow represents the expansion direction vector of the defect area, and the blue arrow represents the principal stress direction vector under the action of external load. The spatial angle between the expansion direction vector of the defect area and the principal stress direction vector under the action of external load is 15°. When the spatial angle threshold is 30°, the size of the spatial angle is compared with the spatial angle threshold. At this time, it is determined that the defect area is expanding in a high-risk direction, and the maintenance priority of the defect area needs to be increased.

[0172] Step S500: Outputting the defect status of the defect area based on the spatiotemporal characteristics of the pixels, and making an abnormality decision based on the defect mode and the degree of danger of the defect area expansion;

[0173] The specific method of outputting the defect status according to the spatiotemporal characteristics of the pixels in the defect area and making an abnormality decision based on the defect mode and the degree of danger of the defect area expansion is as follows:

[0174] Step S510: Time-space characteristics of defect areas in historical time Perform cluster analysis to divide the defect state into three stages: the initiation stage, the stable stage, and the diffusion stage. A clustering model is obtained. The temporal and spatial characteristics of the pixels in the current defect area are input, and the corresponding defect state is output. 、 Represent geometric features and temporal features respectively;

[0175] The cluster analysis uses the K-means clustering algorithm to determine the number of clusters to be 3, corresponding to three defect states. The spatiotemporal features of all defect areas in the historical time period are used as the training set. Each data point corresponds to a spatiotemporal feature. K data points are randomly selected as the initial cluster centers. For each data point, its distance to the K cluster centers is calculated and assigned to the category to which the closest cluster center belongs. For each category, the cluster center is updated to the geometric center of all data points in that category. The above steps are repeated until the cluster center no longer changes, thereby obtaining the final clustering model.

[0176] Exemplarily, according to the trained clustering model, when the spatio-temporal features of the current defective area pixels are input, the clustering model calculates the distances between the spatio-temporal features and each clustering center, and classifies the spatio-temporal features into the category of the clustering center with the closest distance, and this category is the defective state corresponding to the spatio-temporal features of the current defective area pixels.

[0177] Step S520: Adaptively adjust the data adoption frequency of the defective area according to the output of the clustering model;

[0178] Specifically, the method of adaptively adjusting the data adoption frequency of the defective area according to the output of the clustering model is as follows: when the output of the clustering model determines that the defective area is in the budding stage or the stable stage, maintain the data adoption frequency of the defective area at the default slow detection frequency; when the output of the clustering model determines that the defective area enters the diffusion stage, increase the data adoption frequency of the defective area to the high-speed detection rate.

[0179] Preferably, the slow detection frequency is once per hour, and the high-speed detection rate is once per minute.

[0180] Step S530: Make an anomaly decision according to the defective state, defective mode and the risk degree of the expansion of the defective area;

[0181] Exemplarily, the method of making an anomaly decision according to the defective state, defective mode and the risk degree of the expansion of the defective area is as follows: design a rule-based three-color decision system: green indicates safety, the defective state of each defective area pixel is in the budding stage or the stable stage, and the defective mode and the risk degree of the expansion of the defective area are normal and safe; yellow indicates early warning, no more than 30% of the pixels enter the diffusion stage, or the defective mode and the risk degree of the expansion of the defective area are in the abnormal mode and the high-risk direction, but have not reached the red standard; red indicates alarm and maintenance, more than 30% of the pixels enter the diffusion stage, and the defective mode and the risk degree of the expansion of the defective area are in the abnormal mode and the high-risk direction;

[0182] When the red alarm is output for 3 consecutive decision cycles, start the emergency maintenance plan; when the yellow early warning is output for 3 consecutive decision cycles, start the daily maintenance plan and pay close attention; when the green safety state is continuously output, only perform regular periodic inspections. Through the cooperation of the hierarchical state machine, the overall defect detection sensitivity can be improved while reducing the false alarm rate.

[0183] The above steps design a micro-macro hierarchical state machine framework, taking the defect state of pixels in the defect area as the micro basis for abnormal decision-making, and taking the defect mode in the defect area and the risk level of defect area expansion as the macro basis for abnormal decision-making. It fuses multi-source heterogeneous monitoring data at different scales, and proposes an adaptive multi-sensor collaborative optimization method, balancing the real-time performance and reliability of detection.

[0184] Through the analysis of the evolution trend of spatio-temporal features, load correlation analysis and hierarchical decision fusion, this application forms a complete closed-loop solution from local to global, from single point to sequence, from detection to early warning, jointly constituting a "physics + data"-driven, hierarchical distributed defect analysis and early warning system, which can play an important role in the online structural health monitoring of complex working conditions.

[0185] Embodiment 2

[0186] As Figure 5 shown, the intelligent identification system for quality defects of hidden projects in power infrastructure provided by this application includes:

[0187] A multi-modal data collection module, used to construct a three-dimensional point cloud model, obtain the time-series data of structural health monitoring and the visible light image of the surface texture of the transmission tower foundation;

[0188] A spatio-temporal feature fusion module, which extracts time-series features based on the time-series data, inputs the visible light image into a defect segmentation model to obtain a defect probability map, generates geometric features of the defect area based on the defect probability map, and fuses the time-series features and geometric features to obtain spatio-temporal features;

[0189] A defect mode recognition module, which obtains the acceleration vector sequence of the defect area expansion based on the spatio-temporal features, and uses the acceleration vector sequence for defect mode recognition;

[0190] A risk level judgment module, which determines the expansion direction vector of the defect area according to the geometric features, obtains the principal stress direction vector under the action of external loads, and judges the risk level of the defect area expansion;

[0191] A decision comprehensive judgment module, which outputs the defect state of the defect area pixels according to the spatio-temporal features, and makes abnormal decisions in combination with the defect mode and the risk level of the defect area expansion.

[0192] Embodiment 3

[0193] This application also provides a readable storage medium. Computer-readable instructions are stored on the readable storage medium. When the computer-readable instructions are run by a processor, the intelligent identification method for quality defects of hidden projects in power infrastructure according to the embodiments of this application described above can be executed.

[0194] In addition, in the above technical solutions provided in the embodiments of the present application, the parts that are the same as the corresponding technical solutions in the prior art in terms of implementation principles are not described in detail to avoid excessive repetition.

[0195] As described above in the specific embodiments, the purpose, technical solutions and beneficial effects of the present invention have been further described in detail. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent identification method for quality defects of hidden power infrastructure projects, characterized by: include: Construct a 3D point cloud model to obtain time-series data for structural health monitoring and visible light images of the surface texture of the transmission tower base; Extracting time series features based on time series data, inputting the visible light image into the defect segmentation model to obtain a defect probability map, generating geometric features of the defect area based on the defect probability map, and fusing the time series features with the geometric features to obtain spatiotemporal features; Based on the spatiotemporal characteristics, the acceleration vector sequence of the defect area expansion is obtained, and the defect pattern recognition is performed using the acceleration vector sequence; Determine the expansion direction vector of the defect area based on geometric characteristics, obtain the principal stress direction vector under external load, and judge the degree of danger of the defect area expansion; Output the defect status based on the spatiotemporal characteristics of the pixels in the defect area, and make abnormal decisions based on the defect pattern and the degree of danger of the defect area expansion; The specific method of constructing a three-dimensional point cloud model and obtaining time series data for structural health monitoring and a visible light image of the surface texture of the transmission tower base includes: collecting strain data, vibration data, and acoustic emission data through sensors in the transmission tower base, and integrating the time series of the strain data, vibration data, and acoustic emission data into time series data for structural health monitoring; The specific method of extracting time series features based on time series data, inputting visible light images into the defect segmentation model to obtain a defect probability map, generating geometric features of the defect area based on the defect probability map, and fusing time series features with geometric features to obtain spatiotemporal features is as follows: Perform time-frequency domain analysis on the time series data of structural health monitoring, extract time domain features and frequency domain features, and perform feature fusion on the time domain features and frequency domain features to obtain time series features; Align the visible light image and the 3D point cloud model to establish a 3D point cloud model with texture information between the visible light image pixels and the 3D point cloud model; A convolutional neural network that incorporates prior physical knowledge is used for defect segmentation. The input data is a visible light image aligned with a 3D point cloud model, and the output data is a defect probability map. The geometric features of the defect area are calculated based on the defect probability map. The geometric features include the main direction of the defect area. The geometric features and temporal features are aligned according to the timestamps and associated with the three-dimensional spatial coordinates of the corresponding three-dimensional point cloud model to obtain spatiotemporal features; the spatiotemporal features include the three-dimensional spatial coordinates of the center of the defect area, the timestamp, and the geometric features and temporal features corresponding to the timestamp.

2. The intelligent identification method for quality defects of hidden power infrastructure projects according to claim 1, characterized in that: The specific method of constructing a three-dimensional point cloud model and obtaining time series data of structural health monitoring and visible light images of the surface texture of the transmission tower base also includes: The 3D point cloud data is obtained by scanning the transmission tower base and its surrounding environment through LiDAR. The attributes of each point include the 3D coordinates and reflection intensity of the point. Based on the 3D point cloud data, a 3D point cloud model of the transmission tower base is generated. Obtain visible light images of the surface texture of the transmission tower base; A three-dimensional spatial coordinate system is constructed with the center of the transmission tower base surface as the origin, and the time series data of structural health monitoring are synchronized and unified into the three-dimensional spatial coordinate system.

3. The intelligent identification method for quality defects of hidden power infrastructure projects according to claim 2, characterized in that: The convolutional neural network that introduces prior physical knowledge is used to perform defect segmentation. The input data is a visible light image aligned with the three-dimensional point cloud model, and the output data is a defect probability map. The specific method for calculating the geometric features of the defect area based on the defect probability map is as follows: An end-to-end convolutional neural network model is designed for defect segmentation. The visible light image aligned with the three-dimensional point cloud model is input, and the defect labels are annotated on the visible light image to obtain a defect mask. Training samples are constructed based on the visible light image and the corresponding defect mask, and the minimization of the segmentation loss function is defined as the training objective of the convolutional neural network model. Prior physical knowledge is integrated into the defect segmentation model, and a regularization term based on a partial differential equation is defined. The regularization term is added to the segmentation loss function to obtain the total segmentation loss function of the convolutional neural network that introduces prior physical knowledge. The convolutional neural network that introduces prior physical knowledge is trained using the training samples by minimizing the total segmentation loss function to obtain a defect segmentation model. The current visible light image aligned with the three-dimensional point cloud model is used as input data, and the defect probability map is output through the defect segmentation model. The defect area is obtained by segmentation based on the thresholding operation, and the geometric features of the defect area are calculated.

4. The intelligent identification method for quality defects of hidden power infrastructure projects according to claim 3 is characterized in that: The main direction is calculated as follows: The three-dimensional spatial coordinates corresponding to all pixels in a defect area are used as a two-dimensional point cloud to form a B×2 coordinate matrix; B is the number of pixels in the defect area; Decenter the coordinate matrix to obtain the coordinate matrix after removing the mean; Calculate the covariance matrix of the coordinate matrix after removing the mean; Perform eigenvalue decomposition on the covariance matrix and take the eigenvector corresponding to the maximum eigenvalue as the main direction of the defect area.

5. The intelligent identification method for quality defects of hidden power infrastructure projects according to claim 4, characterized in that: The specific method of obtaining the acceleration vector sequence of defect area expansion based on the spatiotemporal characteristics and using the acceleration vector sequence to perform defect pattern recognition is as follows: The spatiotemporal features of the same defect area collected at tn consecutive data sampling times constitute a spatiotemporal feature sequence, and the displacement vector and velocity vector of the three-dimensional spatial coordinates of the defect area over time are calculated based on the spatiotemporal features of adjacent data sampling times; Based on the spatiotemporal feature sequence, the displacement vector sequence and velocity vector sequence of the defect area are obtained. According to the velocity vector between adjacent data sampling times, the acceleration vector sequence of the defect area expansion is calculated. Defect pattern recognition is performed on the acceleration vector sequence of defect area expansion using long short-term memory network; The defect mode includes a normal mode and an abnormal mode.

6. The method for intelligently identifying quality defects of hidden power infrastructure projects according to claim 5, characterized in that: The specific method of using the long short-term memory network to perform defect pattern recognition on the acceleration vector sequence of the defect area expansion is: Long short-term memory (LSTM) networks are used for defect pattern recognition. The input is the acceleration vector sequence of the defect area extension, and the output is the defect pattern corresponding to the data sampling time. A binary cross-entropy loss function is defined, and the LSTM network is trained for supervised learning using the historical acceleration vector sequence with defect pattern labels. When the output of m consecutive data sampling times is an abnormal pattern, it is determined that there is a risk of structural instability.

7. The method for intelligently identifying quality defects of hidden power infrastructure projects according to claim 6, characterized in that: The specific method of determining the expansion direction vector of the defect area based on the geometric characteristics, obtaining the principal stress direction vector under the external load, and judging the danger level of the defect area expansion is as follows: For a sequence of three-dimensional spatial coordinates of a defect area at consecutive tn data sampling times, obtaining the main direction of the defect area as an expansion direction vector; Get the principal stress direction vector under the current external load; Calculate the spatial angle between the expansion direction vector of the defect area and the principal stress direction vector; The consistency between the expansion direction vector of the defect area and the principal stress direction vector is judged based on the spatial angle threshold, and the degree of danger of the defect area expansion is determined; The specific method for judging the consistency between the expansion direction vector of the defect area and the principal stress direction vector based on the spatial angle threshold and determining the danger level of the defect area expansion is as follows: setting the spatial angle threshold; when the spatial angle is less than the spatial angle threshold, it is determined that the defect area is expanding in a high-risk direction; otherwise, the defect area is considered safe. The danger level of the defect area expansion includes high-risk directions and safety.

8. The intelligent identification method for quality defects of hidden power infrastructure projects according to claim 7, characterized in that: The specific method of outputting the defect status according to the spatiotemporal characteristics of the pixels in the defect area and making an abnormality decision based on the defect mode and the degree of danger of the defect area expansion is as follows: Cluster analysis is performed on the spatiotemporal characteristics of defect areas in historical time, and the defect state is divided into three stages: the germination stage, the stable stage, and the diffusion stage. A clustering model is obtained, and the spatiotemporal characteristics of the pixels in the current defect area are input to output the corresponding defect state. Adaptively adjust the data sampling frequency of the defect area according to the output of the clustering model; Make abnormal decisions based on the defect status, defect mode and dangerous degree of defect area expansion.

9. An intelligent identification system for quality defects of concealed power infrastructure projects, used to implement the intelligent identification method for quality defects of concealed power infrastructure projects according to any one of claims 1 to 8, characterized in that: include: A multimodal data collection module, used to construct a 3D point cloud model, acquire time-series data for structural health monitoring, and obtain visible light images of the surface texture of the transmission tower base; The spatiotemporal feature fusion module extracts temporal features based on time series data, inputs the visible light image into the defect segmentation model to obtain a defect probability map, generates geometric features of the defect area based on the defect probability map, and fuses the temporal features with the geometric features to obtain spatiotemporal features. The defect pattern recognition module obtains the acceleration vector sequence of the defect area expansion based on the spatiotemporal characteristics and uses the acceleration vector sequence to perform defect pattern recognition; The danger level judgment module determines the expansion direction vector of the defect area based on geometric characteristics, obtains the principal stress direction vector under the action of external load, and judges the danger level of the defect area expansion; The decision-making comprehensive judgment module outputs the defect status according to the spatiotemporal characteristics of the pixels in the defect area, and makes abnormal decisions based on the defect mode and the degree of danger of the defect area expansion.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which is suitable for being loaded by a processor to execute the steps of the method for intelligently identifying quality defects of hidden power infrastructure projects as described in any one of claims 1 to 8.

Images