A method and device for detecting defects in a tower structure

Abstract

The application provides a tower structure defect detection method and device, and belongs to the tower detection field. The method first acquires tower structure image data and environmental influence parameters at multiple predetermined angles, then pre-processes and fuses the two to obtain a tower structure-environment fusion feature vector, and further generates a defect detection result. During pre-processing and feature fusion, the image data and environmental parameters are first converted into multidimensional matrices, then multi-scale feature analysis is performed to obtain corresponding multi-scale feature vectors, and finally the tower structure-environment fusion feature vector is obtained. Through the combination of multi-angle images and environmental parameters, multi-dimensional conversion, multi-scale analysis and feature fusion, the method improves the comprehensiveness and accuracy of tower defect detection, can more accurately identify defects, and is suitable for efficient detection of tower structures.

Description

Technical Field

[0001] This application belongs to the field of tower inspection technology, and more specifically, relates to a method and device for detecting structural defects in towers. Background Technology

[0002] In infrastructure sectors such as power and communications, transmission towers serve as the core structure supporting critical equipment like cables and antennas, and their operational status directly impacts the stability and safety of the entire system. For a long time, transmission towers have been exposed to the natural environment, facing the continuous effects of wind, sun, rain, and temperature fluctuations. They are also susceptible to factors such as material aging, residual installation stress, and external impacts, making them highly prone to various structural defects, such as corrosion, cracks, deformation, and loose connections. If these defects are not detected and addressed promptly, they can worsen over time, ultimately leading to a decrease in the tower's structural strength and even causing serious safety accidents like collapse, resulting in significant economic losses and adverse social impacts.

[0003] Currently, the main methods for detecting defects in steel tower structures fall into two categories: manual inspection and traditional automated inspection. Manual inspection relies on professionals carrying inspection tools to the site to determine the presence of defects through visual observation, tapping, and instrument measurement. This method not only consumes significant manpower, resources, and time, but is also limited by the experience and responsibility of the inspectors, as well as the site conditions, making it prone to missed or false detections. For example, in complex environments such as mountainous areas or high altitudes, manual inspection is extremely difficult, inefficient, and carries high safety risks.

[0004] Traditional automated inspection methods mostly rely on single image acquisition devices to obtain images of the tower surface, followed by defect identification using image processing techniques. While these methods reduce human intervention to some extent, they still have significant limitations. Images from a single angle or a few angles cannot fully reflect the three-dimensional structure of the tower, and often fail to effectively capture defects hidden behind the structure or in complex connection areas. These methods typically focus only on the image features of the tower itself, ignoring the influence of environmental factors on defect generation and development. For example, high humidity accelerates the corrosion of metal components in the tower, and strong winds can cause fatigue damage to tower connectors. Traditional methods fail to integrate these environmental parameters with image information, significantly impacting the accuracy of defect detection results.

[0005] With technological advancements, some detection methods have begun to incorporate multi-source data fusion techniques, but numerous problems remain in practical applications. Some methods fail to perform reasonable dimensionality transformations on image data and environmental parameters during data preprocessing, resulting in chaotic data structures and an inability to fully extract valuable information. Other methods employ simple splicing or weighting methods during feature fusion, failing to consider the differences in scale and dimension between different types of data. This results in fused feature vectors that cannot accurately reflect the intrinsic relationship between the tower structure and its environment, thus affecting the accuracy of defect detection. Furthermore, existing technologies lack in-depth analysis of multi-scale features, making it difficult to simultaneously identify defects of different scales, such as minute cracks and large-scale deformations on the tower surface, further limiting the practicality of the detection methods. Summary of the Invention

[0006] The purpose of this application is to provide a method and apparatus for detecting defects in the structure of iron towers, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for detecting defects in iron tower structures, the method comprising:

[0008] Acquire tower structure image data from multiple predetermined angles within a predetermined area acquired by an image acquisition device, and environmental impact parameters from the multiple predetermined angles acquired by an environmental sensor;

[0009] The image data of the iron tower structure at the multiple predetermined angles and the environmental impact parameters at the multiple predetermined angles are preprocessed and feature fused to obtain the iron tower structure-environment fusion feature vector;

[0010] Based on the aforementioned tower structure-environment fusion feature vector, defect detection results are generated;

[0011] Specifically, the tower structure image data and environmental impact parameters at the multiple predetermined angles are preprocessed and feature-fused to obtain a tower structure-environment fusion feature vector, including:

[0012] The tower structure image data from the multiple predetermined angles are converted into a multi-dimensional matrix of tower structure images according to the spatial distribution and pixel intensity dimensions, and the environmental impact parameters from the multiple predetermined angles are converted into a multi-dimensional matrix of environmental impact parameters according to the acquisition location and parameter type dimensions.

[0013] Multi-scale feature analysis was performed on the multi-dimensional matrix of the tower structure image and the multi-dimensional matrix of the environmental impact parameters to obtain the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters.

[0014] The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental influence parameters are fused to obtain the tower structure-environment fused feature vector.

[0015] Preferably, the environmental impact parameters include light intensity, wind speed, precipitation, and air humidity.

[0016] Preferably, the tower structure image data includes image resolution, color channel values, and shooting distance parameters.

[0017] Preferably, multi-scale feature analysis is performed on the multi-dimensional matrix of the tower structure image and the multi-dimensional matrix of the environmental impact parameters to obtain the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters, respectively, including:

[0018] The multi-dimensional matrix of the tower structure image is processed by an image feature extractor based on Gaussian pyramid decomposition to obtain the multi-scale feature vector of the tower structure image.

[0019] The multi-dimensional matrix of environmental impact parameters is processed by an environmental parameter feature extractor based on a statistical regression model to obtain the multi-scale feature vector of the environmental impact parameters.

[0020] Preferably, feature fusion is performed on the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental influence parameters to obtain the tower structure-environment fused feature vector, including:

[0021] The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters are input into the principal component analysis network to obtain the preliminary fusion feature matrix of the tower structure-environment.

[0022] The preliminary fusion feature matrix of the tower structure-environment is input into a domain knowledge-guided weight allocation network to obtain a domain knowledge-optimized fusion feature matrix of the tower structure-environment.

[0023] The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters are respectively input into the adaptive feature enhancement module to obtain the optimized multi-scale feature vector of the tower structure image and the optimized multi-scale feature vector of the environmental impact parameters.

[0024] The optimized multi-scale feature vector of the tower structure image and the optimized multi-scale feature vector of the environmental impact parameters are subjected to feature fusion modulation to obtain the tower structure-environment fusion feature vector.

[0025] Preferably, the process of inputting the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters into a principal component analysis network to obtain a preliminary fusion feature matrix of the tower structure and environment includes: performing a weighted summation of the transposes of the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters to obtain a joint matrix;

[0026] The joint matrix is ​​input into the principal component analysis network to obtain the preliminary fusion feature matrix of the tower structure and environment.

[0027] Preferably, the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters are respectively input into the adaptive feature enhancement module to obtain the optimized multi-scale feature vector of the tower structure image and the optimized multi-scale feature vector of the environmental impact parameters, including:

[0028] The multi-scale feature vectors of the tower structure image are normalized to obtain the standard feature vector and weighted feature vector of the tower structure image.

[0029] Using the domain knowledge to optimize the tower structure-environment fusion feature matrix as a reference matrix, the standard feature vector of the tower structure image, the weighted feature vector of the tower structure image, and the reference matrix are input into the adaptive feature enhancement module to obtain the optimized multi-scale feature vector of the tower structure image.

[0030] The multi-scale feature vectors of the environmental impact parameters are normalized to obtain the standard feature vectors and weighted feature vectors of the environmental impact parameters.

[0031] Using the domain knowledge to optimize the tower structure-environment fusion feature matrix as a reference matrix, the standard feature vector of the environmental impact parameter, the weighted feature vector of the environmental impact parameter, and the reference matrix are input into the adaptive feature enhancement module to obtain the optimized multi-scale feature vector of the environmental impact parameter.

[0032] Preferably, the defect detection result is generated based on the tower structure-environment fusion feature vector, including:

[0033] The tower structure-environment fusion feature vector is used through a decision tree model to obtain a classification result, which is used to indicate whether structural defects exist.

[0034] Based on the classification results, the defect detection results are generated.

[0035] Preferably, the present invention further includes a tower structure defect detection device for implementing the tower structure defect detection method described above, the device comprising:

[0036] The image and parameter acquisition module is used to acquire tower structure image data from multiple predetermined angles within a predetermined area acquired by an image acquisition device, and environmental impact parameters from the multiple predetermined angles acquired by an environmental sensor.

[0037] The preprocessing and feature fusion module is used to preprocess and fuse the multiple predetermined angle tower structure image data and the multiple predetermined angle environmental impact parameters to obtain the tower structure-environment fusion feature vector.

[0038] The defect detection result generation module is used to generate defect detection results based on the feature vector fused from the tower structure environment.

[0039] Preferably, the tower structure image data from the multiple predetermined angles and the environmental impact parameters from the multiple predetermined angles are preprocessed and feature-fused to obtain a tower structure-environment fusion feature vector, including:

[0040] The tower structure image data from the multiple predetermined angles are converted into a multi-dimensional matrix of tower structure images according to the spatial distribution and pixel intensity dimensions, and the environmental impact parameters from the multiple predetermined angles are converted into a multi-dimensional matrix of environmental impact parameters according to the acquisition location and parameter type dimensions.

[0041] Multi-scale feature analysis is performed on the multi-dimensional matrix of the tower structure image and the multi-dimensional matrix of the environmental impact parameters to obtain the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters.

[0042] The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental influence parameters are fused to obtain the tower structure-environment fused feature vector.

[0043] The beneficial effects of the method and device for detecting structural defects in iron towers provided by this invention are as follows:

[0044] By employing innovative data processing and feature fusion methods, this approach effectively addresses numerous issues present in existing detection technologies. Instead of being limited to image information from a single angle, it acquires tower structure image data from multiple predetermined angles, comprehensively capturing the structural features of the tower from different spatial perspectives and avoiding missed defects due to limited viewing angles. Simultaneously, it incorporates environmental impact parameters collected from multiple predetermined angles by environmental sensors, organically combining the tower's structural information with its surrounding environmental factors. This ensures that defect detection is no longer isolated from the environment but fully considers the impact of the environment on the generation and development of tower defects, such as corrosion trends in high-humidity environments and structural deformation risks in strong wind environments, thus making the detection results more consistent with actual conditions.

[0045] In the data preprocessing stage, this method transforms tower structure image data from multiple predetermined angles into a multi-dimensional matrix of tower structure images according to spatial distribution and pixel intensity dimensions, and transforms environmental impact parameters into a multi-dimensional matrix of environmental impact parameters according to acquisition location and parameter type dimensions. This dimensional transformation is not simply data organization, but rather starts from the inherent attributes of the data, forming a structured and organized matrix form for image data and environmental parameter data, laying a solid data foundation for subsequent feature analysis. Compared to traditional single-dimensional data processing, this multi-dimensional matrix can retain more original information, including pixel differences at different locations in the image and numerical changes of environmental parameters at different acquisition points, thus fully preserving the effective features contained in the data.

[0046] Multi-scale feature analysis was performed on the multi-dimensional matrices of the tower structure images and the multi-dimensional matrices of environmental impact parameters, respectively, enabling the extraction of feature information from the data at different scale levels. For the tower structure images, multi-scale feature analysis can simultaneously capture small-scale defect features such as fine cracks and spots, as well as large-scale defect features such as overall deformation and large-scale corrosion. For environmental impact parameters, multi-scale analysis can reflect the impact of short-term environmental fluctuations and long-term environmental trends on the tower. This multi-scale feature analysis method breaks through the limitation of single-scale extraction in traditional feature extraction, resulting in richer and more comprehensive features that cover different levels of information, providing more sufficient feature basis for subsequent defect detection.

[0047] In the feature fusion stage, this method does not simply concatenate the multi-scale feature vectors of the tower structure image and the multi-scale feature vectors of environmental influence parameters. Instead, it organically combines these two different types of feature vectors through a scientific fusion approach, forming a fused feature vector that comprehensively reflects the relationship between the tower structure and its environment. This fusion not only preserves the advantages of both image features and environmental parameter features but also uncovers the intrinsic connections between them, such as the correspondence between specific environmental parameters and specific structural defects. This makes the generated fused feature vector more representative and discriminative, and consequently, the defect detection results generated based on this can more accurately identify various defects in the tower, including those hidden defects that are accelerated by environmental factors.

[0048] The entire method, from data acquisition to feature fusion and defect detection result generation, forms a complete and logically rigorous processing flow. The seamless integration of each step prevents information loss or distortion during data processing. In this way, the method can significantly improve the accuracy and reliability of tower structural defect detection, reduce the probability of missed and false detections, help relevant personnel to promptly identify potential safety hazards in towers, and take proactive maintenance measures. This extends the service life of towers, reduces the incidence of safety accidents caused by tower failures, and plays a truly effective role in ensuring the safe operation of infrastructure. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram illustrating the working principle of the tower structure defect detection method described in this invention.

[0051] Figure 2 The flowchart for multi-scale feature analysis;

[0052] Figure 3 Flowchart for feature fusion optimization;

[0053] Figure 4 A flowchart for adaptive feature enhancement. Detailed Implementation

[0054] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0056] Please see Figure 1 This invention provides a method for detecting defects in the structure of iron towers, and the specific implementation steps are as follows:

[0057] The system acquires image data of the tower structure from multiple predetermined angles within a predetermined area, collected by an image acquisition device. Environmental impact parameters from these predetermined angles are collected by environmental sensors. The image acquisition device can be a high-definition camera, an image acquisition module mounted on a drone, etc., capable of capturing images of the tower structure from multiple predetermined angles, such as horizontal, tilted, and overhead views, to obtain comprehensive image information. Environmental sensors can be deployed at multiple predetermined angles corresponding to the image acquisition location, simultaneously collecting environmental impact parameters at those angles to ensure the spatial angular correspondence between the image data and the environmental impact parameters.

[0058] The tower structure image data and environmental impact parameters from multiple predetermined angles are preprocessed and feature fused to obtain a tower structure-environment fused feature vector. The preprocessing mainly involves standardizing the acquired raw data, such as removing noise from the image data and unifying the dimensions of the environmental impact parameters, laying the foundation for subsequent feature extraction and fusion. Feature fusion organically combines the features of the tower structure image data and the environmental impact parameters to form a fused feature vector that comprehensively reflects the information from both.

[0059] Based on the fused feature vector of the tower structure-environment, defect detection results are generated. By analyzing and processing the fused feature vector, it is determined whether there are defects in the tower structure and the specific details of the defects, and finally the defect detection results are output.

[0060] Specifically, the tower structure image data and environmental impact parameters at the multiple predetermined angles are preprocessed and feature-fused to obtain a tower structure-environment fusion feature vector, including:

[0061] The tower structure image data from multiple predetermined angles are converted into a multi-dimensional tower structure image matrix according to spatial distribution and pixel intensity dimensions. Similarly, the environmental impact parameters from these predetermined angles are converted into a multi-dimensional environmental impact parameter matrix according to acquisition location and parameter type dimensions. The spatial distribution dimension reflects the spatial positional relationship of the tower structure at different angles, while the pixel intensity dimension reflects the brightness of each pixel in the image. The acquisition location dimension corresponds to the spatial information of the environmental impact parameters, and the parameter type dimension distinguishes different types of environmental parameters. This conversion transforms the original unstructured data into a structured matrix form, facilitating subsequent feature analysis.

[0062] Multi-scale feature analysis is performed on the multi-dimensional matrix of the tower structure image and the multi-dimensional matrix of the environmental impact parameters to obtain the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters. Multi-scale feature analysis can extract features of the data from different scale levels, such as from local details to overall outline, from short-term fluctuations to long-term trends, thereby capturing key information in the data more comprehensively.

[0063] The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental influence parameters are fused to obtain the tower structure-environment fused feature vector. Through a specific fusion algorithm, the effective information in the two feature vectors is integrated to form a more representative and discriminative fused feature vector, providing a more reliable basis for subsequent defect detection.

[0064] Example 1: See Figure 2 Environmental impact parameters include light intensity, wind speed, precipitation, and air humidity. Light intensity directly relates to the lighting conditions during image acquisition; different light intensities can lead to differences in the sharpness of object outlines and color saturation in images of the tower structure. Wind speed reflects the airflow intensity of the tower's environment; long-term wind action can cause vibration, displacement, or loosening of tower components. Precipitation reflects the total amount of rainfall over a certain period; rainwater may seep into the tower's joints or metal surfaces, causing corrosion and other problems. Air humidity indicates the water vapor content in the air; high humidity accelerates the oxidation process of metal materials. These parameters characterize the natural environmental conditions of the tower from different dimensions and are intrinsically related to the integrity of the tower structure.

[0065] The image data of the iron tower structure includes image resolution, color channel values, and shooting distance parameters. Image resolution refers to the number of pixels per unit length. The higher the resolution, the clearer the details of the iron tower in the image, such as the size of bolts and the texture of components, which can be displayed more finely. Color channel values ​​describe the pixel brightness of different channels such as red, green, and blue in the image. These values ​​can be used to distinguish the color differences in different parts of the iron tower, such as the color difference between normal metal surfaces and rusted areas. The shooting distance parameter records the distance between the image acquisition device and the part of the iron tower being photographed. Different distances affect the proportion and perspective of the iron tower structure in the image. Close-up shots can capture local details, while long-distance shots can present the overall structural layout.

[0066] When performing multi-scale feature analysis on the multi-dimensional matrix of the tower structure image and the multi-dimensional matrix of environmental influence parameters, a Gaussian pyramid decomposition-based image feature extractor is used to obtain multi-scale feature vectors for the tower structure image. The Gaussian pyramid decomposition process involves first applying Gaussian blur to the original image, then downsampling to retain specific pixels, generating a scaled-down image. This process is repeated multiple times to form a series of images at different scales. Extracting features from these images at different scales can cover the feature information of the tower structure from the overall framework to local components. For example, in larger-scale images, the overall structural morphology of the tower and the relative positions of its main components can be extracted, while in smaller-scale images, the shape of bolts and details of welds can be extracted. The multi-scale feature vectors of the tower structure image obtained in this way integrate image features from different observation scales and can comprehensively reflect the visual information of the tower structure.

[0067] For the multi-dimensional matrix of environmental impact parameters, a multi-scale feature vector of environmental impact parameters is obtained through an environmental parameter feature extractor based on a statistical regression model. The statistical regression model can analyze environmental impact parameters from multiple perspectives, including the numerical variation patterns of each parameter and the interaction relationships between different parameters. In multi-scale analysis, different time or spatial scales can be set. For example, at a shorter time scale, the synchronous changes in precipitation and air humidity during a specific precipitation event can be analyzed; at a longer time scale, the cumulative impact of seasonal wind force trends on the tower structure can be analyzed; at a smaller spatial scale, the changes in light intensity at a specific location on the tower can be analyzed; and at a larger spatial scale, the distribution characteristics of environmental parameters throughout a predetermined area can be analyzed. The statistical regression model extracts and integrates the parameter features at these different scales to form a multi-scale feature vector of environmental impact parameters. This vector reflects the potential impact of environmental factors on the tower structure at different temporal and spatial scales.

[0068] In practice, when the image acquisition device takes pictures at multiple predetermined angles within a designated area, it simultaneously records the image resolution, color channel values, and shooting distance parameters at each angle, forming image data of the tower structure. Simultaneously, environmental sensors collect real-time data on light intensity, wind speed, precipitation, and air humidity at the same multiple predetermined angles, forming environmental impact parameters. Subsequently, this tower structure image data is organized according to spatial distribution and pixel intensity dimensions, converting it into a multi-dimensional matrix of tower structure images. The spatial distribution dimension corresponds to the shooting angle, while the pixel intensity dimension integrates brightness-related information from image resolution and color channel values, as well as the pixel density distribution affected by the shooting distance parameter. The environmental impact parameters are arranged according to the acquisition location (i.e., multiple predetermined angles) and parameter type (i.e., light intensity, wind speed, precipitation, and air humidity), converting it into a multi-dimensional matrix of environmental impact parameters.

[0069] After matrix transformation, an image feature extractor based on Gaussian pyramid decomposition is activated to process the multi-dimensional matrix of the tower structure image. This extractor first performs multi-scale decomposition on the image data corresponding to the matrix, generating image representations at different levels. Then, it extracts vectors from each level that reflect features such as the tower structure's outline, component morphology, and surface texture, ultimately combining them to form a multi-scale feature vector of the tower structure image. Simultaneously, an environmental parameter feature extractor based on a statistical regression model analyzes the multi-dimensional matrix of environmental impact parameters. By setting different analysis windows and parameter combinations, it uncovers the variation characteristics and interrelationships of various environmental parameters across different time intervals and spatial ranges, quantifying these features into vector form to construct the multi-scale feature vector of environmental impact parameters.

[0070] Example 2: See Figure 3 The process of fusing multi-scale feature vectors from tower structure images and multi-scale feature vectors from environmental impact parameters to obtain a fused feature vector for the tower structure and environment involves several sequential steps. First, the two feature vectors are input into a principal component analysis (PCA) network to generate a preliminary fused feature matrix for the tower structure and environment. The PCA network processes the high-dimensional input feature vectors, identifying the main directions of change in the data and retaining those feature components that best reflect the original information while reducing redundant information. In this process, the multi-scale feature vectors from the tower structure images carry visual information such as the tower's appearance and structural details extracted from different angles and scales, while the multi-scale feature vectors from the environmental impact parameters contain the variation characteristics of environmental factors such as illumination, wind, precipitation, and humidity at different spatiotemporal scales. The PCA network then performs joint analysis on these two vectors to identify common dimensions that comprehensively reflect the core information of both, combining features from these dimensions to form a preliminary fused feature matrix.

[0071] The resulting preliminary fusion feature matrix of the tower structure and environment is input into a domain knowledge-guided weight allocation network. The domain knowledge encompasses tower structure design specifications, the correlation between common defect types and environmental factors, and the correspondence between image features and structural integrity. The weight allocation network evaluates each element in the preliminary fusion feature matrix based on this knowledge. For example, matrix elements corresponding to image features related to key load-bearing components of the tower, and matrix elements corresponding to parameter features related to long-term high humidity environments, are assigned relatively high values; while matrix elements corresponding to image features belonging to background interference or transient environmental fluctuation parameters are assigned relatively low values. After this adjustment, the preliminary fusion feature matrix is ​​optimized into a domain knowledge-optimized tower structure-environment fusion feature matrix, making the importance of each element in the matrix more closely aligned with the actual needs of tower structure defect detection.

[0072] The multi-scale feature vectors of the tower structure image and the environmental impact parameters need to be optimized separately. This process is achieved through an adaptive feature enhancement module. For the multi-scale feature vector of the tower structure image, the adaptive feature enhancement module first analyzes the distribution of its internal feature components, identifying those components that clearly reflect key information such as the tower structure's edges, connection nodes, and surface damage, while weakening interference components caused by shooting angle deviations, light reflections, etc. For the multi-scale feature vector of environmental impact parameters, the module focuses on parameter change features related to tower material aging and structural stability, such as feature components corresponding to continuous high humidity and parameter fluctuation features under strong winds, strengthening these components and suppressing those parameter fluctuation components that occur occasionally and have a minor impact. After this processing, the optimized multi-scale feature vectors of the tower structure image and the environmental impact parameters are obtained.

[0073] The two optimized feature vectors are fused and modulated to generate the final tower structure-environment fused feature vector. The feature fusion and modulation process establishes a correlation mechanism between the two vectors. For example, it matches image features of a certain part of the tower with environmental parameter features of that location, analyzing the correlation between surface rust features shown in the image and the corresponding long-term high humidity and high precipitation features. This correlation is then incorporated into the fused vector as feature components. Simultaneously, the modulation process unifies the feature dimensions of the two vectors to ensure dimensional consistency in the fused vector, facilitating subsequent defect detection and analysis. During modulation, domain knowledge is referenced to optimize the weight information in the tower structure-environment fused feature matrix, ensuring that features with higher weights occupy more prominent positions in the fused vector.

[0074] In practice, the principal component analysis network processes the multi-scale feature vectors of the tower structure image and the multi-scale feature vectors of environmental impact parameters simultaneously. By extracting the common feature dimensions of both through parallel computation, processing time is shortened. The domain knowledge-guided weight allocation network requires pre-loading relevant domain knowledge datasets. These datasets contain correlation data between various defects recorded in historical detection cases and their corresponding image features and environmental parameters. The network learns from this data to form a criterion for judging feature importance. The adaptive feature enhancement module relies on real-time analysis capabilities, dynamically adjusting the enhancement strategy based on the specific situation of the input feature vectors without manual intervention. The feature fusion modulation stage establishes a feature mapping table, combining each component of the two optimized feature vectors according to preset rules. For example, for image feature components and environmental parameter feature components belonging to the same area of ​​the tower, a weighted superposition method is used to integrate them into a new component in the fusion vector, ensuring that the fused vector can simultaneously reflect the structural appearance information and environmental impact information of that area.

[0075] Example 3: The multi-scale feature vectors of the tower structure image and the multi-scale feature vectors of environmental influence parameters are input into a principal component analysis network to obtain a preliminary fusion feature matrix of the tower structure and environment. The specific operation revolves around the weighted summation of the two feature vectors. This process requires first clarifying the proportion of the two feature vectors in the fusion, i.e., their respective weight values. The setting of these weight values ​​needs to be combined with the actual scenario of tower structure defect detection, comprehensively considering the correlation between image features and environmental parameter features on defect judgment.

[0076] The multi-scale feature vector of the iron tower structure image consists of multiple dimensions, each corresponding to the image features of the tower at different angles and scales, such as the outline changes of the tower body and the texture details of its components. The multi-scale feature vector of environmental impact parameters also contains multiple dimensions, corresponding to the characteristics of light intensity, wind force, precipitation, and air humidity at different spatiotemporal scales, such as the long-term humidity trend at a certain angle and the parameter fluctuations of short-term strong winds. Since the original dimensions of the two vectors may differ, direct calculation will result in a dimension mismatch problem. Therefore, it is necessary to transpose the multi-scale feature vector of environmental impact parameters. After transposition, its dimensional structure will be consistent with the multi-scale feature vector of the iron tower structure image, allowing for subsequent mathematical operations.

[0077] After the transpose operation is completed, the weighted calculation stage begins. Each element in the multi-scale feature vector of the tower structure image is multiplied by its corresponding weight to obtain the weighted result of the vector. Simultaneously, each element in the transposed multi-scale feature vector of the environmental impact parameters is multiplied by its corresponding weight to obtain another set of weighted results. Then, these two sets of weighted results are summed element-wise to obtain the matrix, which is the preliminary fusion feature matrix of the tower structure and environment.

[0078] In practice, the principal component analysis network first preprocesses the multi-scale feature vectors of the tower structure image and the environmental influence parameters to remove outliers and noise. Outliers may originate from sudden interference during image acquisition, such as image data anomalies caused by lens reflection, or erroneous parameter values ​​generated by temporary malfunctions of environmental sensors. Removing these outliers improves the stability of the feature vectors.

[0079] The determination of weights requires multiple rounds of debugging. Initial weights can be set based on the frequency of occurrence of image features and environmental parameters in defect detection in historical data. For example, if image features have successfully identified defects more often in past detection cases, they can be temporarily assigned a higher initial weight. Subsequently, by comparing the fusion results with the actual defect situation, the weight values ​​are gradually adjusted until the weight allocation enables the preliminary fused feature matrix to more accurately reflect the comprehensive information of the tower structure and environment.

[0080] The transpose operation is implemented in computer programs using a matrix transpose function. The program automatically adjusts the rows and columns of the multi-scale feature vector of the environmental impact parameters to match its dimensions with the multi-scale feature vector of the tower structure image. For example, if the multi-scale feature vector of the tower structure image is an m x n matrix, the transposed multi-scale feature vector of the environmental impact parameters will also be adjusted to m x n, ensuring that each element corresponds one-to-one when the two are weighted and summed.

[0081] The specific calculation method for weighted summation can be expressed as follows:

[0082]

[0083] in, This represents the initial fusion feature matrix of the tower structure and its environment. The weights represent the multi-scale feature vectors of the tower structure image. This represents the multi-scale feature vector of the tower structure image. The weights of the multi-scale eigenvectors representing environmental impact parameters are given. This represents the multi-scale feature vector of the environmental impact parameters after transposition.

[0084] During the calculation, the operation of each element is performed independently. For example, the element in the i-th row and j-th column of the multi-scale feature vector of the tower structure image is multiplied by the weight 'a', and the element in the i-th row and j-th column of the transposed multi-scale feature vector of the environmental influence parameters is multiplied by the weight 'b'. The sum of the two products is the value of the element in the i-th row and j-th column of the preliminary fused feature matrix.

[0085] After the initial feature matrix is ​​generated, the principal component analysis network performs dimensionality reduction. The purpose of dimensionality reduction is to reduce redundant information in the matrix and retain those features that are more valuable for defect detection. For example, there may be multiple feature dimensions in the matrix that express similar information; after dimensionality reduction, these dimensions will be merged, making the matrix structure more concise.

[0086] Dimensionality reduction is achieved by extracting the principal components of the matrix. The principal components are the directions of maximum variance in the matrix, reflecting the main trends in data variation. After extracting the principal components, the dimensionality of the initial fused feature matrix will be reduced, but the core information contained within will not be lost.

[0087] Throughout the process, the principal component analysis network records various parameters in real time, including weight values, vector dimensions before and after transposition, and matrix dimensions after summation. These records will be used for subsequent tracking and adjustment. If the initial fused feature matrix performs poorly in subsequent defect detection, the problem can be found by analyzing these records, such as unreasonable weight settings or deviations in the transposition operation, and targeted corrections can be made.

[0088] Example 4: See Figure 4 The multi-scale feature vectors of the tower structure image and the multi-scale feature vectors of environmental influence parameters are respectively input into the adaptive feature enhancement module to obtain two optimized feature vectors. The operation process revolves around normalization processing and feature adjustment guided by the reference matrix.

[0089] For multi-scale feature vectors of tower structure images, normalization is performed first. This process transforms the values ​​of each element in the vector to the same range. For example, if a feature vector contains pixel intensity values ​​of a certain part of the tower, its value range may be between 0 and 255, while another feature may involve shooting distance parameters, with a value range between 10 and 50 meters. Normalization will map these values ​​from different ranges to a uniform range of 0 to 1. During the transformation process, two results are generated simultaneously: a standard feature vector of the tower structure image and a weighted feature vector of the tower structure image. The standard feature vector is a vector that directly reflects the core features of the tower structure image after normalization, such as the clarity of the tower outline after processing and the pixel changes at component connection points; the weighted feature vector reflects the relative importance of each feature after normalization. For example, image features that can clearly distinguish between normal and defective structures will have higher weight values.

[0090] The automatic identification of discriminative features uses a large amount of image feature data of iron towers labeled "normal" and "defective" to calculate the distribution difference of each feature in the two types of samples. Screening is performed using information gain (higher values ​​indicate stronger discriminative power) or analysis of variance (e.g., a p-value less than 0.05 indicates a significant difference), while also incorporating domain knowledge (e.g., edge features of cracks, color features of rust) to assist in confirmation.

[0091] The determination of the weight values ​​is based on the results of information gain or variance analysis of the dependent features. The coefficients output from the training of the logistic regression model are used as the initial weights (the larger the absolute value of the coefficient, the higher the weight). Then, referring to domain knowledge (such as knowing that crack features are more critical than texture features), the initial weights are adjusted proportionally (e.g., the weight of crack features is multiplied by 1.2) to finally obtain the specific values.

[0092] After normalization, the tower structure-environment fusion feature matrix optimized using domain knowledge is used as a reference matrix. The standard feature vector of the tower structure image, the weighted feature vector of the tower structure image, and the reference matrix are input into the adaptive feature enhancement module. The reference matrix contains the feature importance distribution adjusted based on professional knowledge of the tower structure, such as which image features are related to bolt loosening and which environmental parameter features are related to corrosion development. The adaptive feature enhancement module compares the standard feature vector with the corresponding features in the reference matrix. If the value of a feature in the standard feature vector deviates from the expected range of that feature in the reference matrix, the module will adjust it according to the corresponding weight value in the weighted feature vector. For example, if the reference matrix indicates that the feature of the tower beam connection point at a certain angle should have a high value, but the value of this feature in the standard feature vector is low, and it has a high weight in the weighted feature vector, the module will appropriately increase the value of this feature to enhance its performance in the vector.

[0093] The processing flow for multi-scale feature vectors of environmental impact parameters is similar to that for multi-scale feature vectors of tower structure images. First, normalization is performed to unify the feature values ​​of parameters such as light intensity, wind speed, precipitation, and air humidity into the same range. For example, light intensity might be measured in lux, ranging from 0 to 10000, and wind speed might be represented by levels 0 to 12; after normalization, these values ​​will be converted to the range of 0 to 1. Similarly, during the normalization process, standard feature vectors and weighted feature vectors for environmental impact parameters are generated. The standard feature vectors reflect the core changing characteristics of each environmental parameter, such as the humidity fluctuation trend at a certain angle; the weighted feature vectors reflect the relative importance of each environmental parameter feature, for example, the weight of long-term high humidity characteristics might be higher than that of short-term precipitation characteristics.

[0094] Subsequently, using the same domain knowledge-optimized tower structure-environment fusion feature matrix as a reference matrix, the standard feature vector of environmental impact parameters, the weighted feature vector of environmental impact parameters, and the reference matrix are input into the adaptive feature enhancement module. The module adjusts the standard feature vector based on the correlation between environmental parameter features and tower defects in the reference matrix. For example, if the reference matrix indicates a strong correlation between high humidity at a certain angle and corrosion at the base of the tower, and the standard feature vector shows a low value for this humidity feature but a high weight in the weighted feature vector, the module will increase the value of this feature to strengthen its influence in the vector.

[0095] In practice, the adaptive feature enhancement module first performs dimension matching on the input standard feature vector and the reference matrix to ensure that they can be compared element-wise. If there is a difference in dimensions between the vector and the matrix, the module will first expand the dimensions of the vector or compress the dimensions of the matrix to achieve alignment. For example, if the standard feature vector is a one-dimensional vector and the reference matrix is ​​a two-dimensional matrix, the module will convert the vector into a form with the same row or column dimensions as the matrix.

[0096] During the adjustment process, the module avoids over-correction to prevent introducing new biases. For example, if the value of a certain feature in the standard eigenvector deviates too much from the reference matrix, but that feature has a low weight in the weighted eigenvector, the module will only make minor adjustments to maintain the overall stability of the vector. After adjustment, the output optimized environmental impact parameter multi-scale eigenvector will have a more consistent numerical distribution of each feature with the professional knowledge orientation reflected in the reference matrix, and can more accurately reflect the potential impact of environmental factors on the tower structure.

[0097] Throughout the process, the adaptive feature enhancement module's adjustment strategy dynamically changes according to the different input data, without requiring manual intervention. For example, for different types of iron towers, due to differences in their structural design, the feature distribution in the reference matrix will be different. The module will automatically adapt to these differences, adjusting the magnitude and direction of feature enhancement so that the optimized feature vector can better serve the subsequent generation of defect detection results.

[0098] Example 5: The process of generating defect detection results based on the fused feature vector of the tower structure and environment begins by inputting the fused feature vector of the tower structure and environment into a decision tree model. The decision tree model consists of a series of nodes, each node representing a judgment condition for a certain feature in the fused feature vector. These features cover various types of information after the fusion of the tower structure image and environmental impact parameters, such as the comprehensive performance of the image features of a certain part of the tower and the long-term humidity parameters of the corresponding location, and the combination of structural texture and wind force level features at different angles.

[0099] The decision tree model is built upon a large amount of historical data, including the fused feature vectors of the tower structure and environment recorded in past inspections, along with the corresponding actual defects. The model learns from this data to determine the judgment criteria and branch direction for each node. For example, a node might use the fused value of the "tower leg image features and high humidity environmental parameters" from the fused feature vector as its judgment criterion. If this value is greater than a certain threshold, the branch points to "potential corrosion defects," otherwise it points to "no corrosion-related features found."

[0100] When the new tower structure-environment fusion feature vector is input into the decision tree model, the model starts from the root node and propagates downwards layer by layer according to preset judgment conditions. At each node, depending on whether the value of the corresponding feature in the vector is satisfied, it enters different branches until it reaches the leaf node. The output of the leaf node is the classification result, which clearly indicates whether the tower structure has defects.

[0101] When generating defect detection results based on classification results, if the classification result indicates the presence of structural defects, further analysis of the specific feature information in the fused feature vector is required. For example, if the classification result shows the presence of corrosion defects, the approximate location can be determined by using corrosion-related features in the fused feature vector. These features may be associated with image features of the tower component from a specific angle (such as surface color changes and texture roughness) and corresponding environmental parameter features (such as long-term high humidity and high precipitation). By combining the distribution of these features, it is possible to infer that the corrosion may have occurred in a specific part of the tower. Furthermore, the severity of the corrosion can be roughly judged based on the magnitude of the feature values; higher values ​​indicate more pronounced corrosion characteristics.

[0102] If the classification result indicates no structural defects, the defect detection result must clearly state the conclusion that no defects were found, and record the main feature distribution of the fused feature vector during the detection process for subsequent comparative analysis. For example, it can be stated that the features of the tower structure image from various angles do not show any abnormalities, and the corresponding environmental parameter features are also within the normal range, with no significant features that could cause defects.

[0103] In practice, decision tree models are updated regularly. As new detection data accumulates, the model relearns and adjusts the judgment conditions and branch directions of nodes to adapt to the changing characteristics of tower structural defects under different environments. For example, after a new type of defect is discovered in a certain area, the fused feature vector corresponding to that defect and the actual situation are incorporated into historical data. The model will then add or adjust relevant nodes accordingly to make the classification results more in line with actual detection needs.

[0104] The generated defect detection results must be presented in a standardized format, including the tower number detected, detection time, predetermined area range, information summary from multiple predetermined angles, classification results, and a detailed description of the defect (if any). For existing defects, the possible types, approximate locations, and characteristic features must be described in detail to provide clear guidance for subsequent maintenance work. Additionally, the results should include a list of numerical values ​​for key features in the fused feature vector for review and verification.

[0105] Throughout the entire process, from inputting the fused feature vectors into the decision tree model to generating the final defect detection results, everything is automated, minimizing human intervention. The program records processing logs for each step, including the input fused feature vectors, the classification path of the decision tree model, the output classification results, and the feature analysis process during defect detection result generation. These logs can be used to trace and optimize the detection process.

[0106] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting structural defects in iron towers, characterized in that, include: Acquire tower structure image data from multiple predetermined angles within a predetermined area acquired by an image acquisition device, and environmental impact parameters from the multiple predetermined angles acquired by an environmental sensor; The image data of the iron tower structure at the multiple predetermined angles and the environmental impact parameters at the multiple predetermined angles are preprocessed and feature fused to obtain the iron tower structure-environment fusion feature vector; Based on the aforementioned tower structure-environment fusion feature vector, defect detection results are generated; Specifically, the tower structure image data and environmental impact parameters at the multiple predetermined angles are preprocessed and feature-fused to obtain a tower structure-environment fusion feature vector, including: The tower structure image data from the multiple predetermined angles are converted into a multi-dimensional matrix of tower structure images according to the spatial distribution and pixel intensity dimensions, and the environmental impact parameters from the multiple predetermined angles are converted into a multi-dimensional matrix of environmental impact parameters according to the acquisition location and parameter type dimensions. Multi-scale feature analysis was performed on the multi-dimensional matrix of the tower structure image and the multi-dimensional matrix of the environmental impact parameters to obtain the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters. The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental influence parameters are fused to obtain the tower structure-environment fused feature vector. Multi-scale feature analysis is performed on the multi-dimensional matrix of the tower structure image and the multi-dimensional matrix of the environmental impact parameters to obtain the multi-scale feature vectors of the tower structure image and the environmental impact parameters, respectively, including: The multi-dimensional matrix of the tower structure image is processed by an image feature extractor based on Gaussian pyramid decomposition to obtain the multi-scale feature vector of the tower structure image. The multi-dimensional matrix of environmental impact parameters is processed by an environmental parameter feature extractor based on a statistical regression model to obtain the multi-scale feature vector of the environmental impact parameters. The tower structure image multi-scale feature vector and the environmental influence parameter multi-scale feature vector are fused to obtain the tower structure-environment fused feature vector, including: The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters are input into a principal component analysis network to obtain a preliminary fusion feature matrix of the tower structure and environment. The preliminary fusion feature matrix of the tower structure-environment is input into a domain knowledge-guided weight allocation network to obtain a domain knowledge-optimized fusion feature matrix of the tower structure-environment. The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters are respectively input into the adaptive feature enhancement module to obtain the optimized multi-scale feature vector of the tower structure image and the optimized multi-scale feature vector of the environmental impact parameters. The optimized multi-scale feature vector of the tower structure image and the optimized multi-scale feature vector of the environmental impact parameters are subjected to feature fusion modulation to obtain the tower structure-environment fusion feature vector. The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters are respectively input into the adaptive feature enhancement module to obtain the optimized multi-scale feature vector of the tower structure image and the optimized multi-scale feature vector of the environmental impact parameters, including: The multi-scale feature vectors of the tower structure image are normalized to obtain the standard feature vector and weighted feature vector of the tower structure image. Using the domain knowledge to optimize the tower structure-environment fusion feature matrix as a reference matrix, the standard feature vector of the tower structure image, the weighted feature vector of the tower structure image, and the reference matrix are input into the adaptive feature enhancement module to obtain the optimized multi-scale feature vector of the tower structure image. The multi-scale feature vectors of the environmental impact parameters are normalized to obtain the standard feature vectors and weighted feature vectors of the environmental impact parameters. Using the domain knowledge to optimize the tower structure-environment fusion feature matrix as a reference matrix, the standard feature vector of the environmental impact parameter, the weighted feature vector of the environmental impact parameter, and the reference matrix are input into the adaptive feature enhancement module to obtain the optimized multi-scale feature vector of the environmental impact parameter.

2. The method for detecting structural defects in iron towers according to claim 1, characterized in that, The environmental impact parameters include light intensity, wind speed, precipitation, and air humidity.

3. The method for detecting structural defects in iron towers according to claim 2, characterized in that, The tower structure image data includes image resolution, color channel values, and shooting distance parameters.

4. The method for detecting structural defects in iron towers according to claim 1, characterized in that, The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters are input into a principal component analysis network to obtain a preliminary fusion feature matrix of the tower structure and environment. This includes: weighted summation of the transposes of the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters to obtain a joint matrix. The joint matrix is ​​input into the principal component analysis network to obtain the preliminary fusion feature matrix of the tower structure and environment.

5. The method for detecting structural defects in iron towers according to claim 1, characterized in that, Defect detection results are generated based on the aforementioned tower structure-environment fusion feature vector, including: The tower structure-environment fusion feature vector is used through a decision tree model to obtain a classification result, which is used to indicate whether structural defects exist. Based on the classification results, the defect detection results are generated.

6. A tower structure defect detection device, used to implement the tower structure defect detection method as described in any one of claims 1 to 5, characterized in that, include: The image and parameter acquisition module is used to acquire tower structure image data from multiple predetermined angles within a predetermined area acquired by an image acquisition device, and environmental impact parameters from the multiple predetermined angles acquired by an environmental sensor. The preprocessing and feature fusion module is used to preprocess and fuse the multiple predetermined angle tower structure image data and the multiple predetermined angle environmental impact parameters to obtain the tower structure-environment fusion feature vector. The defect detection result generation module is used to generate defect detection results based on the feature vector fused from the tower structure environment.

7. The tower structure defect detection device according to claim 6, characterized in that, Preprocessing and feature fusion are performed on the tower structure image data from the multiple predetermined angles and the environmental impact parameters from the multiple predetermined angles to obtain the tower structure-environment fused feature vector, including: The tower structure image data from the multiple predetermined angles are converted into a multi-dimensional matrix of tower structure images according to the spatial distribution and pixel intensity dimensions, and the environmental impact parameters from the multiple predetermined angles are converted into a multi-dimensional matrix of environmental impact parameters according to the acquisition location and parameter type dimensions. Multi-scale feature analysis is performed on the multi-dimensional matrix of the tower structure image and the multi-dimensional matrix of the environmental impact parameters to obtain the multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental impact parameters. The multi-scale feature vector of the tower structure image and the multi-scale feature vector of the environmental influence parameters are fused to obtain the tower structure-environment fused feature vector.

Images