Bolt loosening detection method and server

Abstract

This application discloses a bolt loosening detection method and server, relating to the field of electrical data processing technology. The disclosed bolt loosening detection method includes: acquiring bolt images at the support connection points of photovoltaic strings in a photovoltaic power station; determining the reconstructed image corresponding to the bolt image using a target generative adversarial network (PGN) model, and determining the reconstruction error between the bolt image and the reconstructed image. The PGN model is trained using samples of fastened bolt images, and the reconstruction error characterizes the degree of looseness of the bolt to be detected in the bolt image, with the reconstruction error being positively correlated with the degree of looseness; based on the reconstruction error, determining the first loosening detection result for the bolt to be detected. This technical solution improves the detection rate of loose bolts in photovoltaic power stations.

Description

Technical Field

[0001] This application relates to the field of electrical data processing technology, and in particular to a bolt loosening detection method and server. Background Technology

[0002] In the structure of a photovoltaic (PV) power station, photovoltaic panels are fixed to the support structure with bolts. This connection method is crucial for ensuring the stability of PV panel installation. However, PV power stations are typically deployed outdoors, making them highly susceptible to environmental corrosion and external forces, which can lead to malfunctions such as loose bolts. As a key connecting component between the PV panel and the support structure, if loose bolts are not detected and addressed in a timely manner, it can result in decreased installation accuracy and displacement of the PV panels. In severe cases, it can even cause PV panels to detach or the support structure to collapse, resulting in safety accidents. This not only damages the PV modules and reduces power generation but also poses a threat to the personal safety of on-site maintenance personnel, leading to significant economic losses and safety hazards.

[0003] Currently, images of bolts at the connection points of photovoltaic string supports can be collected. A supervised learning-based bolt loosening detection model can be used to analyze and process the images to determine whether the bolts are loose, thus enabling rapid detection of bolt faults.

[0004] However, the performance of supervised learning-based bolt loosening detection models relies on a large number of bolt image samples clearly labeled as loose. The training effect of the model is directly related to the quantity and diversity of the samples. In actual photovoltaic power plant operation and maintenance scenarios, image samples of loose bolts are extremely scarce, resulting in poor generalization ability of the trained model and a tendency to miss loose bolts. Summary of the Invention

[0005] The main purpose of this application is to provide a bolt loosening detection method and server, which aims to reduce the need for loose bolt samples, improve the generalization ability of the target generative adversarial network model, and thus improve the detection rate of loose bolts in photovoltaic power plants.

[0006] To achieve the above objectives, this application proposes a method for detecting bolt loosening, comprising: Obtain images of bolts at the support connections of photovoltaic strings in a photovoltaic power plant; The target generative adversarial network model is used to determine the reconstructed image corresponding to the bolt image and to determine the reconstruction error between the bolt image and the reconstructed image. The target generative adversarial network model is trained by fastening bolt image samples. The reconstruction error is used to characterize the looseness of the bolt to be detected in the bolt image, and the reconstruction error is positively correlated with the looseness. Based on the reconstruction error, the first loosening detection result of the bolt to be tested is determined.

[0007] In addition, to achieve the above objectives, this application also proposes a server, including: a memory, a processor, and a bolt loosening detection program stored in the memory and executable on the processor, the bolt loosening detection program being configured to implement the steps of the bolt loosening detection method as described above.

[0008] The proposed technical solutions in this application have at least the following technical effects: This application eliminates the dependence of existing supervised learning-based detection methods on scarce labeled samples of loose bolts. The target generative adversarial network model can be trained using readily available images of fastened bolts, fully grasping the image features of normal fastened bolts. This effectively solves the problems of insufficient model training, poor generalization ability, and missed detections caused by scarce samples, reducing the need for loose bolt samples and improving the generalization ability of the target generative adversarial network model. At the same time, the reconstruction error characterizes the degree of bolt looseness, and the two are positively correlated. When the bolt to be detected is loose, its image features differ from the normal fastened bolt features learned by the model. The model cannot accurately reconstruct a matching image, resulting in significant reconstruction errors. This allows for the accurate capture of subtle features of slight bolt loosening. Furthermore, the model learns the essential features of fastened bolts, making it more adaptable to the complex outdoor environment of photovoltaic power stations. This reduces missed detections caused by environmental factors and labeling biases, thereby improving the detection rate of loose bolts. Attached Figure Description

[0009] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating an embodiment of the bolt loosening detection method of this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the bolt loosening detection method of this application; Figure 3 This is a flowchart illustrating Embodiment 3 of the bolt loosening detection method of this application; Figure 4 This is a flowchart illustrating Embodiment 4 of the bolt loosening detection method of this application; Figure 5 This is a flowchart illustrating Embodiment 5 of the bolt loosening detection method of this application; Figure 6 This is a schematic diagram of the overall process of the bolt loosening detection method of this application; Figure 7 This is a schematic diagram of the server structure in this application.

[0012] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0013] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0014] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0015] In the structure of a photovoltaic (PV) power station, photovoltaic panels are fixed to the support structure with bolts. This connection method is crucial for ensuring the stability of PV panel installation. However, PV power stations are typically deployed outdoors, making them highly susceptible to environmental corrosion and external forces, which can lead to malfunctions such as loose bolts. As a key connecting component between the PV panel and the support structure, if loose bolts are not detected and addressed in a timely manner, it can result in decreased installation accuracy and displacement of the PV panels. In severe cases, it can even cause PV panels to detach or the support structure to collapse, resulting in safety accidents. This not only damages the PV modules and reduces power generation but also poses a threat to the personal safety of on-site maintenance personnel, leading to significant economic losses and safety hazards.

[0016] Currently, images of bolts at the connection points of photovoltaic string supports can be collected. A supervised learning-based bolt loosening detection model can be used to analyze and process the images to determine whether the bolts are loose, thus enabling rapid detection of bolt faults.

[0017] However, the performance of supervised learning-based bolt loosening detection models relies on a large number of bolt image samples clearly labeled as loose. The training effect of the model is directly related to the quantity and diversity of the samples. In actual photovoltaic power plant operation and maintenance scenarios, image samples of loose bolts are extremely scarce, resulting in poor generalization ability of the trained model and a tendency to miss loose bolts.

[0018] To address the aforementioned shortcomings, this application proposes a bolt loosening detection method. The main technical solution includes: acquiring bolt images at the support connection points of photovoltaic strings in a photovoltaic power station; determining the reconstructed image corresponding to the bolt image using a target generative adversarial network (PGN) model, and determining the reconstruction error between the bolt image and the reconstructed image. The PGN model is trained using samples of fastened bolt images, and the reconstruction error characterizes the degree of looseness of the bolt to be detected in the bolt image, with the reconstruction error being positively correlated with the degree of looseness; based on the reconstruction error, determining the first loosening detection result of the bolt to be detected.

[0019] The technical solution of this application breaks away from the dependence of existing supervised learning-based detection methods on scarce labeled samples of loose bolts. The target generative adversarial network model can be trained using readily available images of fastened bolts, fully mastering the image features of normal fastened bolts. This effectively solves the problems of insufficient model training, poor generalization ability, and missed detections caused by scarce samples, reducing the need for loose bolt samples and improving the generalization ability of the target generative adversarial network model. At the same time, the reconstruction error represents the degree of bolt looseness, and the two are positively correlated. When the bolt to be detected is loose, its image features differ from the normal fastened bolt features learned by the model. The model cannot accurately reconstruct a matching image, resulting in significant reconstruction errors. This allows for the accurate capture of subtle features of slight bolt loosening. Furthermore, the model learns the essential features of fastened bolts, making it more adaptable to the complex outdoor environment of photovoltaic power stations. This reduces missed detections caused by environmental factors and labeling biases, thereby improving the detection rate of loose bolts.

[0020] Furthermore, many related technologies employ the method of using drones to fly under photovoltaic panels and photograph the bottom of the bolts. From this perspective, the length characteristics of the bolts are compressed, making it difficult to effectively capture key lateral features of bolt loosening, such as bolt shank tilt and gaps between the nut and the connector. Simultaneously, the cluttered background and uneven lighting under the panels increase the difficulty of image analysis and recognition.

[0021] To address this deficiency, the main technical means employed in this application include: determining the flight path and flight control parameters of the UAV, wherein the flight control parameters include at least the target top-down angle; controlling the UAV to fly along the side of the photovoltaic string according to the flight path, and during the flight, controlling the UAV to capture images of the bolts at the support connection point below the photovoltaic string from the target top-down angle. This application uses the target top-down angle for image capture, and the bolt images acquired from this angle can reveal the loosening characteristics of the bolts on the side, providing high-quality input for subsequent analysis.

[0022] Furthermore, related technologies typically only output two results: loose or not loose, lacking analysis of the bolt's condition evolution trend. Maintenance personnel cannot predict which bolts are slowly loosening or at what rate, and can only respond passively, making it difficult to implement predictive maintenance.

[0023] To address this deficiency, the main technical approach employed in this application includes: constructing a time series of the reconstruction error of the target bolt based on the reconstruction error determined in each historical loosening detection, wherein the target bolt is any bolt at the support connection of any photovoltaic string; based on the time series of reconstruction errors, predicting the loosening trend of the target bolt, assessing the loosening rate and degradation risk level of the target bolt, and generating maintenance recommendations for the target bolt. This approach provides valuable predictive maintenance insights for power plant operation and maintenance, helping to eliminate potential safety hazards in advance and optimize operation and maintenance.

[0024] Furthermore, in related technologies, bolt loosening is usually assessed from a single dimension, which leads to a decrease in the accuracy of bolt loosening assessment results.

[0025] To address this deficiency, the main technical approach employed in this application includes: extracting bolt features from bolt images to be detected, wherein the bolt features include a first feature and / or a second feature. The first feature includes the length of the bolt extending beyond the nut, and the second feature includes at least one of the following: photovoltaic bracket gap, photovoltaic module alignment, and bolt posture. Based on the first feature and / or at least one second feature, a second loosening detection result for the bolt to be detected is determined. This approach enables comprehensive analysis of bolt loosening from multiple dimensions. Furthermore, in addition to analysis based on bolt features, the loosening detection result of the bolt to be detected can also be determined using a target generative adversarial network model, enriching the dimensions of bolt loosening detection. Combining unsupervised and multi-source geometric features forms a dual verification, significantly improving the accuracy of bolt loosening detection results and reducing the false alarm rate.

[0026] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or server capable of performing the above functions. The following description uses a server as an example to illustrate this embodiment and the subsequent embodiments.

[0027] It should be noted that the server in this application is connected to the drone for acquiring bolt images at the bracket connection points of the photovoltaic strings in the photovoltaic power station and for subsequent bolt loosening detection.

[0028] Based on this, the embodiments of this application provide a method for detecting bolt loosening, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the bolt loosening detection method of this application.

[0029] In this embodiment, the bolt loosening detection method includes steps S10~S30: Step S10: Obtain images of bolts at the bracket connections of the photovoltaic strings in the photovoltaic power station; The bracket connection refers to the connection point where the support brackets of the photovoltaic string are connected to each other, or the part where the bracket is fixedly connected to the photovoltaic panel. This bracket connection point is where the bolts are installed, and loose bolts at this point will directly affect the structural stability of the photovoltaic string.

[0030] Bolt images refer to visual images captured through specific imaging methods that include the bolts at the connection points of photovoltaic string supports. These images serve as the raw data for subsequent inspection. Specifically, the bolt images can be captured by drones, robots, or other equipment with image acquisition capabilities. The bolt images can be acquired in real-time or from historically acquired and cached images. This allows for analysis of both real-time and historically acquired and cached images, meeting bolt inspection needs in various scenarios. A bolt image may contain only one bolt or multiple bolts; that is, at least one bolt must be present in the image.

[0031] In one alternative approach, an industrial-grade high-definition visual acquisition device is deployed at the photovoltaic power station site. This device is fixedly installed at pre-set inspection points on the photovoltaic support structure, with the device facing the bolts connecting the photovoltaic string support structure to the columns and beams. The device triggers the shooting action through a pre-set timed acquisition command. During the acquisition process, the device's built-in supplementary lighting device is activated to eliminate interference such as uneven lighting and shadows, directly acquiring clear original images of the bolt areas. After acquisition, the image data is uploaded to the server in real time via wired transmission, enabling the server to obtain bolt images at the connection points of the photovoltaic string support structure in the photovoltaic power station and perform subsequent bolt loosening detection.

[0032] In another alternative approach, an automated inspection robot equipped with a vision acquisition module is used. This robot moves along a pre-set inspection path within the photovoltaic power station area. When the robot reaches the connection point of the photovoltaic string support, its built-in positioning module precisely aligns with the bolt inspection area. Using a gimbal, the robot adjusts the shooting angle and focus of the vision acquisition module to acquire multi-angle images of bolts of different specifications and installation positions. After preliminary noise reduction processing by the robot's built-in system, the acquired images are transmitted to a server for storage via a wireless communication module. When the server needs to detect bolt loosening, it can retrieve and analyze historically acquired and stored bolt images. This automated inspection robot can be a drone, or other robots or devices with image acquisition and mobility capabilities.

[0033] Step S20: Determine the reconstructed image corresponding to the bolt image through the target generative adversarial network model, and determine the reconstruction error between the bolt image and the reconstructed image. The target generative adversarial network model is trained through fastened bolt image samples. The reconstruction error is used to characterize the looseness of the bolt to be detected in the bolt image, and the reconstruction error is positively correlated with the looseness. The target generative adversarial network model refers to a generative adversarial network model that has been specially trained on images of fastened bolts. It consists of a generator network and a discriminator network that are trained against each other. Its function is to extract and reconstruct features from the input bolt images. It can accurately restore the image features of normally fastened bolts, but it cannot accurately reconstruct images of abnormally loose bolts.

[0034] The reconstructed image refers to the image output by the target generative adversarial network model after feature encoding and decoding of the input bolt image. It serves as a reference image for comparison with the original input bolt image. It should be noted that the target generative adversarial network model in this embodiment is trained with a large number of normally tightened bolt images. The training objective is to enable the encoder and decoder to accurately extract the structural features of normal bolts and reconstruct a reconstructed image highly consistent with the normal bolt image from the low-dimensional feature space. This process makes the reconstructed image a standardized template for normal bolts. That is, for the input normally tightened bolt image, the model can reconstruct an image almost indistinguishable from the original image, with extremely low reconstruction error. However, for abnormal bolt images that are loose, tilted, or have gaps, their features deviate significantly from the normal bolt features learned by the model. The encoder cannot extract matching normal features, and the decoder cannot reconstruct an image that closely matches the original abnormal image, ultimately resulting in a high reconstruction error. Therefore, by comparing the reconstructed image with the original input bolt image, the pain point of scarce loose bolt samples in real-world photovoltaic power plant scenarios is solved. Abnormal bolts can be identified using only normal samples, without relying on labeled loose samples.

[0035] Reconstruction error refers to the quantified value of the feature difference between the original bolt image and the corresponding reconstructed image. Its function is to characterize the degree of bolt loosening. It can be calculated based on differences in pixel and structural features of the images and is a quantitative indicator for judging whether a bolt is loose. A larger reconstruction error indicates a greater degree of bolt loosening, while a smaller reconstruction error indicates a lesser degree of loosening, thus enabling the assessment of bolt loosening based on reconstruction error.

[0036] The bolt image samples refer to images of bolts in a normally tightened and secure state in a photovoltaic power station. These images are the foundational data for training the target generative adversarial network (GAN) model. The number of these bolt image samples needs to be sufficient and cover bolt image samples under different environments within the photovoltaic power station to ensure the generalization ability of the target GAN model. To better identify bolt loosening, the images in these bolt image samples can be side views of the bolts.

[0037] The bolts to be inspected refer to the bolts at the connection points of the photovoltaic string supports in a photovoltaic power station, which need to be checked for looseness. There can be one or more bolts to be inspected.

[0038] The degree of looseness refers to the extent to which the bolt under test deviates from its normal tightening state. Specifically, the higher the degree of looseness, the higher the risk of the bolt becoming loose; conversely, the lower the degree of looseness, the tighter the bolt is.

[0039] In one alternative approach, the acquired bolt image can be input into a pre-trained target generative adversarial network (GAN) model. The generator of this GAN model first encodes and extracts features such as bolt contours, textures, and connection gaps from the input bolt image, mapping the bolt image to low-dimensional features. Then, it decodes these low-dimensional features to generate a reconstructed image with the same dimensions as the input bolt image. Subsequently, the model's discriminator calculates the mean square error between the bolt image and the reconstructed image, which is taken as the reconstruction error. Furthermore, besides calculating the mean square error between the bolt image and the reconstructed image to determine the reconstruction error, the peak signal-to-noise ratio (PSNR) can also be used to determine the degree of distortion in the reconstructed image and thus the reconstruction error. For example, the more accurate the reconstruction, the higher the PNR; the looser the bolt, the greater the image distortion and the lower the PNR.

[0040] In another alternative approach, the input bolt image is first normalized to unify the pixel specifications and brightness standards before being input into the target generative adversarial network model. The model uses a multi-layer convolutional neural network to perform deep extraction and feature mapping of the bolt's structural features, generating a reconstructed image that fits the normal fastening state. Then, a structural similarity algorithm is used to comprehensively calculate the differences between the bolt image and the reconstructed image in three dimensions: texture, structure, and brightness. The differences in the three dimensions are then weighted and summed to obtain a reconstruction error that comprehensively reflects the differences between the images.

[0041] Understandably, by using a trained target generative adversarial network model to reconstruct the image and calculate the error, the physical state of bolt loosening is transformed into a quantifiable numerical indicator. This establishes a correlation between image differences and the degree of bolt loosening, providing a direct quantitative basis for subsequent determination of whether the bolt is loose.

[0042] Step S30: Determine the first loosening detection result of the bolt to be tested based on the reconstruction error.

[0043] The first loosening test result includes both loosening and tightening. In other application scenarios, it can be further divided into multiple loosening levels or multiple tightening levels. For example, a higher loosening level indicates a looser bolt, and vice versa. Similarly, a higher tightening level indicates a tighter bolt, and vice versa.

[0044] In one alternative approach, multiple error thresholds can be pre-set based on a large number of bolt samples with different degrees of looseness. The reconstructed error obtained above is compared with the preset thresholds one by one. When the reconstructed error is lower than the lowest threshold, the bolt is determined to be in a normal tightness state. When the error is between two adjacent thresholds, the bolt is determined to be in a slightly loose state. When the error is higher than the highest threshold, the bolt is determined to be in a severely loose state. This forms a corresponding first looseness detection result. For example, the first looseness detection result includes one of the following: normal tightness state, slightly looseness state, and severely looseness state.

[0045] In another optional approach, the calculated reconstruction error is input into a preset error grading judgment model. This model learns from historical detection data to form a mapping relationship between error and loosening level. The model automatically matches the corresponding loosening level according to the input reconstruction error, thereby forming the corresponding first loosening detection result. For example, the first loosening detection result includes the specific loosening level.

[0046] In this embodiment, the reliance on scarce labeled samples of loose bolts in existing supervised learning-based detection methods is eliminated. The target generative adversarial network model can be trained using readily available images of fastened bolts, fully grasping the image features of normally fastened bolts. This effectively solves the problems of insufficient model training, poor generalization, and missed detections caused by scarce samples, reducing the need for loose bolt samples and improving the generalization ability of the target generative adversarial network model. At the same time, the reconstruction error represents the degree of bolt looseness, and the two are positively correlated. When the bolt to be detected is loose, its image features differ from the features of normally fastened bolts learned by the model. The model cannot accurately reconstruct a matching image, resulting in significant reconstruction errors. This allows for the accurate capture of subtle features of slightly loose bolts. Furthermore, the model learns the essential features of fastened bolts, making it more adaptable to the complex outdoor environment of photovoltaic power stations. This reduces missed detections caused by environmental factors and labeling biases, thereby improving the detection rate of loose bolts.

[0047] In one feasible implementation, step S20 may include steps S21 to S23: Step S21: Map the bolt image to a low-dimensional bolt feature vector through the encoding network of the target generative adversarial network model; The encoding network, as a component of the target generative adversarial network model, is responsible for feature extraction and dimensionality reduction of the input bolt image. It maps high-dimensional image pixel data into low-dimensional feature vectors, mining the feature information of the bolt image and serving as a prerequisite for image reconstruction. The encoding network consists of sequentially connected convolutional layers, batch normalization layers, nonlinear activation layers, and downsampling layers. After extracting the structural features of a normal fastening bolt through three levels of convolution and downsampling, global pooling is used to obtain a fixed-dimensional low-dimensional bolt feature vector.

[0048] Low-dimensional bolt feature vectors refer to low-dimensional numerical vectors obtained by the encoding network after feature extraction and dimensionality reduction of bolt images. They contain structural and appearance feature information of bolt images. Compared with the original high-dimensional image data, they are easier for the model to perform subsequent decoding and reconstruction calculations. They are feature bridges connecting the original image and the reconstructed image.

[0049] In one alternative approach, a bolt image can be mapped to a low-dimensional bolt feature vector through three levels of convolution, max pooling downsampling, and global average pooling. Specifically, the first convolutional layer takes the bolt image as input and performs convolution operations to extract shallow texture features of the bolt; after convolution, it is processed by a batch normalization layer, and then the feature representation is enhanced by a ReLU nonlinear activation unit to obtain a first-level feature map. The second convolutional layer takes the first-level feature map as input and performs convolution operations to extract mid-level contour features of the bolt; simultaneously, it performs downsampling through a max pooling layer to compress the feature map size, and after batch normalization and ReLU activation, it obtains a second-level feature map. The third convolutional layer takes the second-level feature map as input and performs convolution operations to extract deep structural features such as the bolt's geometric contour; it again performs downsampling through a max pooling layer to further compress the feature map size, and after batch normalization and ReLU activation, it obtains a final deep feature map. The global pooling layer performs global average pooling on the final deep feature map, transforming the two-dimensional feature map into a one-dimensional vector, which is then mapped to a fixed-dimensional low-dimensional bolt feature vector, completing the encoding.

[0050] In another alternative approach, a three-stage convolutional multi-scale kernel, average pooling downsampling, and global max pooling are used to map the bolt image to a low-dimensional bolt feature vector. Specifically, the first convolutional layer takes the bolt image as input, fuses multi-scale convolutional kernels for convolution operations, extracts multi-scale shallow features of the bolt, and obtains a first-level feature map after batch normalization and LeakyReLU nonlinear activation. The second convolutional layer takes the first-level feature map as input to extract the mid-level contour features of the bolt; it compresses the feature size through downsampling by an average pooling layer, and obtains a second-level feature map after batch normalization and LeakyReLU activation. The third convolutional layer takes the second-level feature map as input to extract the deep structural features of the bolt; it further compresses the size through downsampling by an average pooling layer, and obtains a final deep feature map after batch normalization and LeakyReLU activation. Finally, global pooling is performed on the final deep feature map to extract key feature values ​​from the feature map, transforming them into a fixed-dimensional low-dimensional bolt feature vector, thus completing the encoding.

[0051] Understandably, for bolt images, hierarchical feature extraction through three-level convolution and downsampling accurately captures the structural features of normally fastened bolts, discarding redundant background and noise information in the image; global pooling is used to obtain a fixed-dimensional low-dimensional feature vector, realizing the mapping of high-dimensional bolt images to low-dimensional feature space, providing feature basis for image reconstruction of the decoding network.

[0052] Step S22: The low-dimensional bolt feature vector is reconstructed in reverse through the decoding network of the target generative adversarial network model to obtain the reconstructed image; The decoding network, as a component of the target generative adversarial network model, functions to perform inverse feature upscaling and reconstruction on the low-dimensional bolt feature vector output by the encoding network, transforming the low-dimensional feature vector back into high-dimensional image data and generating a reconstructed image corresponding to the original bolt image. The decoding network consists of deconvolutional layers, batch normalization layers, nonlinear activation layers, and upsampling layers. After three levels of deconvolution and upsampling to progressively restore the bolt outline and refine image details, the reconstructed image is obtained through the output reconstruction layer.

[0053] In one alternative approach, the low-dimensional bolt feature vector is reconstructed using a three-stage deconvolution, bilinear interpolation upsampling, and sigmoid normalized activation process to obtain the reconstructed image. Specifically, the first deconvolution layer maps the low-dimensional bolt feature vector output from the encoding network to a low-dimensional feature map through a fully connected layer, and then performs deconvolution. Simultaneously, the feature map size is magnified by a factor of 2 through bilinear interpolation upsampling, followed by batch normalization and ReLU activation to obtain the first-level reconstructed feature map, which initially restores the bolt's feature contour. The second deconvolution layer uses the first-level reconstructed feature map as input for deconvolution transformation; it again magnifies the feature map size by a factor of 2 through bilinear interpolation upsampling, followed by batch normalization and ReLU activation to obtain the second-level reconstructed feature map, gradually refining the bolt's geometric contour and structural proportions. The third deconvolutional layer takes the second-level reconstructed feature map as input and performs deconvolution operations. It then fine-tunes the feature map size through bilinear interpolation upsampling, followed by batch normalization and ReLU activation to obtain the third-level reconstructed feature map. This further restores the texture and structural information of the bolt region, making the feature map size close to that of the original bolt image. The output reconstruction layer performs convolution adjustments on the third-level reconstructed feature map and normalizes the pixel values ​​using a Sigmoid normalization activation unit to match the pixel value range of the original bolt image, ultimately outputting the reconstructed bolt image.

[0054] In another alternative approach, the low-dimensional bolt feature vector is reconstructed through a three-stage deconvolutional transpose convolution, nearest neighbor interpolation upsampling, and Tanh normalization activation process to obtain the reconstructed image. The first deconvolutional layer maps the low-dimensional bolt feature vector to a low-dimensional feature map, then performs transpose convolution. Nearest neighbor interpolation upsampling enlarges the feature size, followed by batch normalization and LeakyReLU activation to obtain the first-level reconstructed feature map, initially restoring the bolt's outline. The second deconvolutional layer performs transpose convolution on the first-level feature map; nearest neighbor interpolation upsampling further enlarges the size, followed by batch normalization and LeakyReLU activation to obtain the second-level reconstructed feature map, restoring the bolt's structural details and proportions. The third deconvolutional layer performs transpose convolution on the second-level feature map; nearest neighbor interpolation upsampling calibrates the feature map size, followed by batch normalization and LeakyReLU activation to obtain the third-level reconstructed feature map, refining the bolt's texture and edge features. The output reconstruction layer performs convolution optimization on the third-level reconstructed feature map, normalizes the pixel values ​​through the Tanh normalization activation unit, matches the pixel distribution of the original bolt image, and finally outputs the reconstructed bolt image.

[0055] Understandably, based on the low-dimensional bolt feature vector output by the encoding network, the feature map size is gradually enlarged and the details of the bolt image are refined through hierarchical feature restoration using three-level deconvolution and upsampling, accurately restoring the contour, structure, and texture features of a normal fastening bolt. Through the normalization processing of the output reconstruction layer, it is ensured that the size and pixel value range of the reconstructed image are completely consistent with the original bolt image, providing a comparable standard image for subsequent pixel-by-pixel calculation of reconstruction error.

[0056] Step S23: Determine the mean square error between the bolt image and the reconstructed image pixel by pixel, as the reconstruction error between the bolt image and the reconstructed image.

[0057] Pixel-by-pixel refers to using a single pixel in the bolt image as the basic unit of calculation, and performing feature comparison and calculation on the corresponding pixels in the original bolt image and the reconstructed image.

[0058] Mean squared error (MSE) is a quantitative indicator obtained by calculating the sum of squares of the differences in pixel values ​​between corresponding pixels in the original image and the reconstructed image, and then taking the average value. It is a classic indicator that characterizes the differences in pixel features between two images and can accurately and objectively reflect the overall pixel difference between the original bolt image and the reconstructed image.

[0059] In one alternative approach, both the original bolt image and the reconstructed image are first converted into single-channel grayscale images. The grayscale values ​​of corresponding pixels in the two grayscale images are obtained pixel by pixel. The square of the difference between the grayscale values ​​of each pixel is calculated. The sum of the squared differences of all pixels is then divided by the total number of pixels in the image to obtain the mean square error as the reconstruction error.

[0060] In another alternative approach, for the colored bolt image and the reconstructed image, pixel data of the red, green, and blue color channels are extracted respectively. The mean square error of the corresponding pixel in each channel is calculated pixel by pixel. Then, according to the influence of the three channels on the image features, different weight coefficients are set, and the mean square errors of the three channels are weighted and summed to obtain the overall mean square error as the reconstruction error.

[0061] Understandably, by using the pixel mean square error as a quantification standard, the feature differences between the original bolt image and the reconstructed image can be accurately calculated, thereby obtaining the reconstruction error that can objectively characterize the degree of bolt loosening, and providing a precise and unified quantitative indicator for subsequent loosening judgment.

[0062] In this embodiment, an encoding network is used to extract features, and a decoding network is used to reconstruct them. This ensures that the reconstructed image can accurately reproduce the features of normally fastened bolts. For loose bolts, due to abnormal features, accurate reconstruction is not possible. This strengthens the feature difference between normal bolts, which have small reconstruction errors, and loose bolts, which have large reconstruction errors, thus improving the accuracy of the reconstruction error in representing the loose state of the bolts. Furthermore, the pixel-wise mean square error is explicitly used as the calculation method for the reconstruction error. Compared with other image difference calculation methods, the mean square error can more accurately and objectively reflect the overall pixel difference between two images, avoiding errors caused by subjective judgment or single feature calculation, and improving the accuracy of the reconstruction error calculation.

[0063] Based on the first embodiment of this application, in the second embodiment of this application, the same or similar content as the above embodiment can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 Methods for detecting loose bolts include: Step S10: Obtain images of bolts at the bracket connections of the photovoltaic strings in the photovoltaic power station; Step S40: Identify the bolt region in the bolt image; The bolt region refers to the local image area containing the bolt to be detected in the bolt image. It is the effective area of ​​the bolt image. This area excludes interference from background and bracket-related parts of the image, retaining only the main features of the bolt, and is the area for subsequent model analysis.

[0064] In one alternative approach, object detection models such as YOLO and Faster R-CNN are trained using images containing bolts from photovoltaic power plants as training samples. The trained object detection model can accurately identify the main bolt region in the bolt image and output the bounding box coordinates of each bolt region, including the x and y coordinates of the top-left corner and the bottom-right corner, thus achieving precise localization of the bolt region. Specifically, taking the YOLO target detection model as an example, when processing images of bolts in photovoltaic power plants, the YOLO target detection model first uses the original bolt image as input. The preprocessing layer scales the image to a fixed size, performs pixel normalization and boundary filling, and outputs a standardized image with uniform size and numerical specifications. This standardized image is then input into the backbone network layer, where features are extracted through multiple convolutions, pooling, and residual structures, outputting a multi-scale feature map containing shallow edge contours and deep semantic information of the bolts. This multi-scale feature map is then input into the neck network layer, where upsampling, downsampling, and cross-layer feature concatenation are performed using feature pyramids and path aggregation structures, outputting an enhanced feature map that integrates fine-grained positional information and high-level semantic information. The enhanced feature map enters the detection head network layer, where regression calculations are performed based on grid cells and adapted bolt anchor frames, outputting prediction results including target confidence, bolt category probability, and bounding box offset and scaling factor. Finally, the prediction results are input into the post-processing layer, where non-maximum suppression removes redundant and duplicate boxes, and the precise upper-left and lower-right corner coordinates of each bolt region are output based on the original image coordinates, completing end-to-end localization and recognition of the bolt target.

[0065] In another alternative approach, the bolt image is first converted to grayscale and filtered for noise reduction to enhance the contrast between the bolt and the background. Then, the image is binarized and segmented by setting a grayscale threshold to separate the bolt region from the background region. Finally, the bolt contour is extracted from the segmented image, and the boundary range of the bolt region is determined based on the contour to achieve bolt region recognition.

[0066] Understandably, accurately locating and identifying the region containing only the bolts from the original bolt image, eliminating interference from irrelevant areas such as the background, brackets, and photovoltaic panels, and determining the scope of subsequent clipping can improve the efficiency and accuracy of model analysis and avoid interference from irrelevant background information on model reconstruction and error calculation.

[0067] Step S50: Crop the bolt image into multiple bolt region images, and input each bolt region image into the target generative adversarial network model respectively; Bolt region images refer to local images obtained by cropping the original bolt image according to the identified bolt regions. Each image contains only the complete features of the bolt to be detected and serves as the actual analysis data input to the target generative adversarial network model. Cropped images can reduce background interference and improve the model's analysis efficiency and accuracy.

[0068] In one alternative approach, the original bolt image is automatically cropped in batches using an image cropping algorithm based on the bounding box coordinates of each bolt region, resulting in multiple bolt region images, each containing a single bolt. Each bolt region image is then individually input into the target generative adversarial network model for analysis, following the order in which the images were generated.

[0069] Specifically, the image cropping algorithm described above can be as follows: First, read the pixel matrix of the original bolt image, determine the horizontal and vertical coordinates of the upper left and lower right corners of the bounding box, take the upper left corner coordinates as the starting point and the lower right corner coordinates as the ending point, delineate a rectangular cropping range in the pixel matrix, completely extract all pixels within this range and recombine them into a new image matrix, remove the background area that exceeds the bounding box in the original bolt image, and finally generate a cropped image containing only a single bolt target.

[0070] Alternatively, the image cropping algorithm described above can also be as follows: Obtain the original coordinates of the bolt bounding box; set a fixed expansion pixel value based on the bolt size and image resolution; expand the bounding box outwards by the corresponding number of pixels in each of the four directions (up, down, left, and right); during the expansion process, determine whether it exceeds the original bolt image boundary; if it does, use the image boundary as the cropping limit; then perform region cropping according to the expanded new coordinate range to obtain a region image that retains the complete bolt structure and a small amount of surrounding support background, ensuring that the subsequent target generative adversarial network model has sufficient contextual feature information for analysis. This method can improve the integrity of the cropped image, reduce the problem of missing bolt features due to overly tight cutting, and improve the accuracy of reconstruction error calculation.

[0071] Understandably, converting the original full bolt image into multiple individual bolt region images allows the model to perform independent reconstruction and error calculation for each bolt, avoiding interference between multiple bolts in the same image. This also significantly reduces the amount of input data for the model and improves the model's inference computation efficiency.

[0072] Step A21: Determine the reconstructed image corresponding to the bolt region image using the target generative adversarial network model, and determine the reconstruction error between the bolt region image and the reconstructed image corresponding to the bolt region image.

[0073] In one feasible approach, for a single bolt region image, the bolt region image is mapped to a low-dimensional bolt feature vector through the encoding network of the target generative adversarial network model; the low-dimensional bolt feature vector is then reconstructed using the decoding network of the target generative adversarial network model to obtain a reconstructed image of the bolt region image; the mean square error between the bolt region image and its reconstructed image is determined pixel-by-pixel as the reconstruction error between the bolt region image and its reconstructed image. It should be noted that step A21 is a refinement of step S20, and the specific processing procedure can be referred to steps S21 to S23 above.

[0074] Step S30: Determine the first loosening detection result of the bolt to be tested based on the reconstruction error.

[0075] In this embodiment, by identifying and cropping the bolt region, interference from irrelevant information such as background and brackets in the original bolt image is effectively eliminated. This allows the target generative adversarial network model to focus only on the main features of the bolt for reconstruction and error calculation, avoiding model reconstruction deviations caused by irrelevant information. This significantly improves the accuracy of reconstruction error calculation and thus enhances the accuracy of bolt loosening detection.

[0076] In one feasible implementation, step S30 in the first and second embodiments described above includes: Step S31: If the reconstruction error is greater than or equal to the preset error value, determine the first loosening detection result as the bolt to be tested being loose; The preset error value can be determined in advance based on experiments. For example, it can be a pre-set critical value for reconstruction error based on the bolt type, installation process, operating environment, operation and maintenance experience of the photovoltaic power station, as well as a large amount of bolt testing data. This serves as a quantitative standard for judging whether the bolts are loose. This preset error value can be fixed or dynamically adjusted according to the bolt's usage scenario and loosening risk level.

[0077] The bolt to be tested is loose when the reconstruction error of the bolt to be tested reaches or exceeds the preset error value. This indicates that it deviates from the normal tightening state and has loosening characteristics such as elongation, tilting, and gap between the nut and the connecting parts, and it needs to be tightened for maintenance.

[0078] Step S32: If the reconstruction error is less than the preset error value, determine the first loosening detection result as the bolt to be tested is tightened.

[0079] The bolt under test is considered tight when the reconstruction error of the bolt under test is lower than the preset error value, indicating that it is in a normal tight state with no obvious loosening characteristics and no need for immediate maintenance.

[0080] In this embodiment, a threshold-based determination method is used. If the reconstruction error is greater than or equal to a preset error value, the first loosening detection result is determined to be that the bolt to be tested is loose; if the reconstruction error is less than the preset error value, the first loosening detection result is determined to be that the bolt to be tested is tight. This enables rapid determination of the bolt loosening state, improving the real-time performance and efficiency of bolt detection.

[0081] Based on the above embodiments of this application, in the third embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 Methods for detecting loose bolts include: Step S10: Obtain images of bolts at the bracket connections of the photovoltaic strings in the photovoltaic power station; Step S60: Extract the bolt features of the bolt to be detected in the bolt image, wherein the bolt features include a first feature and / or a second feature, the first feature includes the length of the bolt extending out of the nut, and the second feature includes at least one of the following: photovoltaic bracket gap, photovoltaic module alignment, and bolt posture. Bolt features refer to the set of visual features extracted from the bolt images at the connection of the photovoltaic panel under bracket, which can characterize the state of the bolt and the surrounding connection parts. They are an important basis for judging whether the bolt is loose. They can reflect the tightness of the bolt from dimensions such as geometric size, surrounding structural fit, and bolt posture. Their function is to provide multi-dimensional judgment basis for bolt loosening detection and make up for the limitations of single reconstruction error judgment.

[0082] The first feature refers to the characteristic extracted from the bolt image that reflects the bolt's own geometric dimensions, specifically the length of the bolt extending beyond the nut. This length, representing the visual pixel distance or actual physical distance from the tip of the nut to the root of the bolt in the bolt image, is a key feature indicating whether the bolt is loose. It can be converted into actual length through image pixel calculation or dimensional calibration. This first feature is directly related to the bolt's looseness; a loose bolt will have a visual length greater than a normally tightened bolt due to the protruding bolt, making it a geometric feature for determining bolt looseness.

[0083] The second feature refers to the set of features extracted from the bolt image that reflects the state of the surrounding connection structure and the bolt's own spatial posture. It is an indirect characteristic of bolt loosening. Bolt loosening causes changes in the state of surrounding supports or photovoltaic modules, and also alters the bolt's own posture. This feature can indirectly verify the bolt's loosening state. Specifically, the photovoltaic support gap refers to the visual gap between the photovoltaic panel fixing support and the connector caused by bolt loosening. Normally tightened bolts ensure a tight fit between the support and connector with no obvious gap. After bolt loosening, a identifiable gap appears between the support and connector, and the size of the gap is positively correlated with the degree of bolt loosening. Photovoltaic module alignment refers to the degree of positional matching between adjacent photovoltaic modules and between upper and lower rows of photovoltaic modules in a photovoltaic string. It is determined by the bolt tightening state. Normally tightened bolts ensure accurate alignment of photovoltaic modules. Loose bolts cause displacement of photovoltaic modules, resulting in offset or misalignment, thus reducing alignment. Bolt posture refers to the spatial arrangement of bolts at the connection of photovoltaic brackets. It is a spatial characteristic that reflects whether the bolts are loose. Normally tightened bolts maintain a standard posture that is perpendicular to the connection surface of the bracket. After the bolts are loose, they will show posture changes such as tilting or swaying, which can be quantified by angle and direction.

[0084] In one alternative approach, key regions such as the bolt to be detected, the connection point of the photovoltaic support, and the edge of the photovoltaic module can be identified and located in the bolt image first, and the bounding box coordinates of each region can be output. Then, the bolt body, the gap region of the support, and the overlapping region of the photovoltaic module can be segmented by a semantic segmentation algorithm. The pixel distance of the bolt extending beyond the nut can be calculated based on the pixel coordinates. The size of the gap of the photovoltaic support can be calculated by the pixel area of ​​the gap region. The alignment of the photovoltaic module can be calculated by the pixel offset of the edge of the photovoltaic module. The axial direction of the bolt can be detected by Hough transform, and the bolt tilt angle and direction can be calculated to obtain the bolt posture.

[0085] Specifically, when extracting multi-dimensional features from bolt images, the bolt images are first input into the YOLO object detection model trained above. The model extracts multi-scale features from the image through the backbone network, and after feature fusion by the neck network, the head network accurately identifies and locates key regions in the image, such as the bolts to be detected, the connection parts of the photovoltaic bracket, and the edges of the photovoltaic modules. It outputs the rectangular bounding box coordinates of each region in pixels, achieving pixel-level localization of key regions. Based on the above bounding box coordinates, the bolt image is input into the semantic segmentation model to perform instance segmentation of each key region. The model generates pixel-level masks for the bolt body, the bracket gap area, and the overlapping area of ​​the photovoltaic module through the mask branch, accurately segmenting the pixel set of each feature region and removing background, irrelevant bracket, and other interfering pixels. For the bolt extension length feature, the extreme points of the profile are extracted from the bolt body mask to determine the pixel coordinates of the nut top and bolt root. The pixel distance between the two points is calculated using the Euclidean distance formula to obtain the pixel quantization value of the bolt extension length. For the photovoltaic bracket gap feature, based on the bracket gap area mask, the total number of pixels in the gap area is counted to obtain the pixel area. At the same time, the maximum pixel width of the gap area profile is detected and used as the quantization value of the photovoltaic bracket gap. For the photovoltaic module alignment feature, the pixel coordinates of the photovoltaic module edge mask are extracted. Combined with the module standard alignment edge pixel baseline calibrated by the photovoltaic power station installation specification, the horizontal and vertical pixel offsets of the actual edge relative to the standard baseline are calculated as 1 - (offset / module standard pixel width). The horizontal and vertical alignment are calculated separately, and the average of the two is taken as the final quantized value of the photovoltaic module alignment. For the bolt posture features, the bolt body mask is first binarized, and then the bolt axis line is detected by Hough linear transform. The pixel coordinates of the line are extracted and its slope is calculated. Then, the angle between the axis and the perpendicular line of the photovoltaic bracket connection surface is derived as the bolt tilt angle. At the same time, the direction vector of the axis line is calculated. The specific direction of the bolt is determined by the angle between the direction vector and the vertical positive direction. Finally, the bolt posture is quantified by combining the tilt angle and the direction.

[0086] Understandably, by extracting multi-dimensional visual features that characterize the loosening state of bolts from bolt images, the visual state of the bolt and its surroundings can be transformed into quantifiable feature indicators, providing a multi-dimensional basis for the judgment of the second loosening detection result, and realizing the transformation from single image data to feature indicators.

[0087] Step S70: Determine the second loosening detection result of the bolt to be tested based on the first feature and / or at least one second feature; The second loosening detection result is a judgment of the loosening state of the bolt under test based on the extracted bolt features. It is divided into two types: loose and tight. It is a preliminary judgment of bolt loosening obtained from the dimensions of bolt geometric features and surrounding structural features. This second loosening detection result can also be similar to the first loosening detection result.

[0088] In one alternative approach, if only the first feature is extracted, the calculated length of the bolt extending beyond the nut is compared with the average length of normally tightened bolts in the same row. If it exceeds the normal range, it is determined to be loose. If only a single second feature is extracted, the photovoltaic bracket gap is compared with the preset gap, the photovoltaic module alignment with the preset alignment, and the bolt posture with the preset posture. If any feature exceeds the normal range, it is determined to be loose. If multiple features are extracted, and any feature is determined to be loose, the overall second looseness detection result is loose.

[0089] In another optional approach, if the first feature and multiple second features are extracted simultaneously, a weight coefficient is set for each feature. The weight coefficient can be set such that the length of the bolt extending out of the nut has the highest weight, followed by the gap of the photovoltaic bracket, and the alignment of the photovoltaic module and the bolt posture decrease in that order. The deviation rate between the actual value and the standard value of each feature is multiplied by the corresponding weight to obtain the weighted deviation value of each feature. All weighted deviation values ​​are summed to obtain the total deviation value. If the total deviation value is greater than the preset total deviation threshold, the second loosening detection result is determined to be loose; otherwise, it is considered tight.

[0090] Step S80: Determine the final loosening test result of the bolt to be tested based on the first loosening test result and / or the second loosening test result.

[0091] The final loosening detection result refers to the comprehensive judgment of the loosening state of the bolt under test, obtained by combining the first loosening detection result and / or the second loosening detection result. It is a final judgment result that integrates the reconstruction error quantification index and the multi-dimensional visual characteristics of the bolt, and has higher accuracy and reliability compared to a single detection result. This final loosening detection result can also be divided into two types: loose and tight.

[0092] In one alternative approach, if only the first or second loosening detection result is obtained, that result is directly used as the final loosening detection result. If both results are obtained simultaneously, the first and second loosening detection results need to be processed according to the actual situation to obtain the final loosening detection result for the bolt to be inspected. For example, based on the importance of the first and second features in the current scenario, either the first or second loosening detection result can be selected as the final loosening detection result. The importance of each feature in each scenario can be set according to the actual situation.

[0093] In another optional approach, if two detection results are obtained simultaneously, a judgment confidence level is set for the first loosening detection result and the second loosening detection result. If the confidence level of one of the detection results is greater than the preset confidence threshold, then the detection result is taken as the final loosening detection result. If the confidence levels of both detection results are greater than the threshold, the judgment result that is consistent with the two is taken. If they are inconsistent, the features are re-extracted and the reconstruction error is recalculated for verification.

[0094] In another alternative approach, if the first loosening detection result is the same as the second loosening detection result, either the first loosening detection result or the second loosening detection result is taken as the final loosening detection result of the bolt to be tested; if the first loosening detection result is different from the second loosening detection result, the final loosening detection result of the bolt to be tested is determined comprehensively based on the reconstruction error and the quantification results of the bolt characteristics.

[0095] The quantitative result refers to the specific numerical result obtained by converting the extracted bolt features. It is a quantitative indicator that characterizes the bolt feature state, such as the specific value of the bolt extension length beyond the nut, the pixel width of the bracket gap, the percentage of photovoltaic module alignment, and the specific degree of bolt tilt angle. Its function is to provide accurate numerical basis for comprehensive judgment when the first loosening detection result and the second loosening detection result are inconsistent.

[0096] The final loosening detection result of the bolt to be tested is determined comprehensively based on the quantification results of reconstruction error and bolt features. This includes: assigning corresponding weights to the quantification results of reconstruction error and bolt features; performing a weighted summation of the quantification results of reconstruction error and bolt features according to the corresponding weights; determining the weighted summation result; and comparing the weighted summation result with a preset threshold. If the result is greater than the preset threshold, it indicates loosening; otherwise, it indicates that the bolt to be tested is tight. The preset threshold is a preset value used to judge the tightness and looseness of the bolt and can be set according to the actual situation. The first loosening detection result is mutually verified with the second loosening detection result to determine the final loosening detection result, thereby improving the accuracy of the final loosening detection result and avoiding misjudgment of bolt loosening due to judgment based on only a single dimension.

[0097] In this embodiment, multi-dimensional bolt features are extracted from bolt images for loosening detection, overcoming the limitations of relying solely on reconstruction error judgment. The method determines bolt loosening from multiple dimensions, including bolt geometry, surrounding connection structure status, and bolt spatial orientation, providing multiple verification bases for the detection results and significantly improving the accuracy of bolt loosening detection. Furthermore, the extracted bolt features cover both direct and indirect characteristics of bolt loosening. The length of the bolt extending beyond the nut directly reflects the loosening state, while the gap in the photovoltaic support, the alignment of the photovoltaic module, and the bolt orientation serve as auxiliary features for indirectly verifying the loosening state. The combination of direct and indirect features effectively identifies bolt loosening caused by different reasons, improving the adaptability of the detection method to various loosening scenarios.

[0098] In one feasible implementation, step S70, determining the second loosening detection result of the bolt to be tested based on the first feature, includes: Step A71: Obtain other bolts located in the same row as the bolt to be tested; The same row of bolts refers to all bolts in a photovoltaic power station that belong to the same photovoltaic string and are distributed sequentially along the string arrangement axis. These bolts use the same model, installation process and fastening standard. Under normal conditions, the length of the bolts extending out of the nut is basically the same, which can be used as a reference standard for mutual comparison. Its function is to provide a standard comparison sample for the length of the bolt to be tested.

[0099] In one alternative approach, the photovoltaic string row to which the bolt to be tested belongs is determined based on the 3D modeling data of the photovoltaic power station and the string clustering results; the identification information, location information and image information of all tested bolts in that row are retrieved, and all other bolts in the same row as the bolt to be tested are automatically matched to form a set of bolts in the same row.

[0100] In another alternative approach, in the original image of the bolt to be detected, all bolts distributed along the photovoltaic string arrangement axis are identified. Based on the pixel position of the bolt in the image, other bolts arranged at the same horizontal level as the bolt to be detected are determined, and the image regions and feature information of these bolts are extracted as samples of other bolts in the same row.

[0101] Understandably, this step provides a standard reference sample for comparing the lengths of the bolts to be tested, ensuring the rationality and accuracy of the length comparison and avoiding judgment deviations caused by comparing bolts from different rows or models.

[0102] Step A72: Determine the difference in the length of the bolt extending beyond the nut between the other bolts and the bolt to be tested; In one alternative approach, the average length of all other bolts in the same row is first calculated as the standard normal length for that row. Then, the length of the bolt to be tested is subtracted from this average length to obtain the length difference between the bolt to be tested and the bolts in the same row. If the difference is positive, it indicates that the length of the bolt to be tested extending beyond the nut is greater than the normal standard. Alternatively, the difference between the lengths of other bolts and the length of the bolt to be tested extending beyond the nut can be calculated directly.

[0103] Step A73: If the difference is greater than the preset difference, the second loosening detection result is determined to be that the bolt to be tested is loose.

[0104] The preset difference value is a critical value for the length difference of bolts extending beyond the nut, pre-set based on the bolt type, installation process, and a large amount of bolt testing data within the same row. It serves as a quantitative standard for judging whether the length of the bolt extending beyond the nut is abnormal. This difference value can effectively distinguish between normal, minor length differences and significant length increases caused by loosening. For example, if the bolt type and installation process are different in each row, then the preset difference value can also be different.

[0105] In one alternative approach, the calculated length difference is directly compared with a preset difference. If the difference is greater than the preset difference, the second loosening detection result is directly determined to be that the bolt to be tested is loose. Alternatively, areas with abnormal bolt extension lengths beyond the nut and the magnitude of the difference can be marked on the bolt image.

[0106] In this embodiment, bolts in the same row are used as length comparison references to ensure the homogeneity of the comparison samples and avoid length judgment deviations caused by differences in bolt type, installation process, and operating environment. This improves the accuracy and rationality of judging looseness based on the length of the bolt extending beyond the nut. The length of the bolt to be tested is compared with the length of the bolts in the same row to obtain the specific length difference. Then, the difference is compared with a preset difference to improve the accuracy of the detection result of whether the bolt is loose.

[0107] In one feasible implementation, determining the second loosening detection result of the bolt to be detected based on at least one second feature in step S70 includes: Step B71: If the bracket gap is greater than the preset gap, the second loosening test result is determined to be that the bolt to be tested is loose; The preset gap refers to the critical gap value between the photovoltaic bracket and the connector that is set in advance based on the model of the photovoltaic bracket, the installation process, the bolt tightening standard, etc. It is a quantitative standard for judging whether the bracket gap is abnormal. Under normal tightening conditions, the bracket gap is less than the preset gap, and the bracket gap will be greater than the preset gap after the bolts are loosened.

[0108] Step B72: If the alignment of the photovoltaic module is less than the preset alignment, the second loosening detection result is determined to be that the bolt to be tested is loose. Preset alignment refers to the critical value of photovoltaic module alignment set in advance according to the installation specifications or string arrangement requirements of photovoltaic modules. It is a quantitative standard for judging whether photovoltaic modules are misaligned. Under normal tightness, the alignment of photovoltaic modules should be greater than this alignment value. After the bolts are loosened, the photovoltaic modules will be displaced and the alignment value will be less than this alignment value.

[0109] Step B73, the bolt posture includes the bolt tilt angle and the bolt direction. If the bolt posture is the preset bolt tilt angle and the bolt direction is the preset bolt direction, the second loosening detection result is determined to be that the bolt to be detected is loose.

[0110] The preset bolt tilt angle value is a critical value for bolt tilt angle set in advance according to the bolt installation specifications or the connection requirements of the photovoltaic bracket. It is a quantitative standard for judging whether the bolt is tilted. Under normal tightening conditions, the bolt tilt angle should be less than the preset bolt tilt angle. After the bolt is loosened, the bolt will tilt obviously and the tilt angle will be greater than the preset bolt tilt angle.

[0111] The preset bolt direction refers to the normal installation direction of the bolts set in advance according to the bolt installation specifications or the connection structure of the photovoltaic bracket. It is the standard for judging whether the bolts are deviating. Under normal tightening conditions, the bolt direction should be consistent with this direction. After the bolts are loose, the bolt direction will deviate from this standard direction.

[0112] In this embodiment, bolt loosening determination based on the second feature is realized from three dimensions: photovoltaic bracket gap, photovoltaic module alignment, and bolt posture. This covers the changes in the surrounding structure state and the bolt's own posture caused by bolt loosening, forming a multi-dimensional indirect feature determination system. This system can verify the loosening state of the bolt from different perspectives, improving the comprehensiveness and reliability of the second loosening detection results.

[0113] Based on the above embodiments of this application, in the fourth embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 4 Methods for detecting loose bolts include: Step S01: Determine the flight path and flight control parameters of the UAV, wherein the flight control parameters include at least the target's top-down angle; The flight path refers to the flight trajectory of the drone when collecting bolt images in the photovoltaic power station. It is planned based on the 3D modeling data of the photovoltaic power station, the arrangement of the photovoltaic strings, and the location of the support connection. Its purpose is to ensure that the drone can collect bolt images of the support connection below the photovoltaic strings in a full and complete manner.

[0114] Flight control parameters refer to the set of various parameters that control the UAV's flight and image acquisition states. These parameters ensure the UAV flies according to plan and acquires high-quality bolt images. The flight control parameters include at least the target view angle, and may also include flight altitude, flight speed, shooting interval, gimbal rotation speed, etc. The target view angle refers to the preset downward shooting angle when the UAV's gimbal camera faces the support connection point below the photovoltaic string. It is a key parameter ensuring the camera can clearly capture the loosening characteristics of the bolt's side. This angle allows the camera's optical axis to form the optimal observation angle with the bolt's side, maximizing the presentation of the bolt's loosening characteristics, such as elongation, tilt, and gaps. This target view angle can be fixed or varied according to actual conditions; for example, it can be set to -60°.

[0115] In one alternative approach, a drone equipped with a lidar and a visible light camera acquires high-precision laser point cloud data and orthophotos of the photovoltaic power station. The orthophotos are used for visual segmentation, while the laser point cloud provides accurate 3D spatial information. The geographic coordinates of the boundaries and corners of each photovoltaic module are identified and extracted from the orthophotos. Subsequently, based on a preset 3D spatial distance threshold and the geographic coordinates of the boundaries and corners of each photovoltaic module, cluster analysis is performed on these discrete modules to group spatially adjacent modules into rows and determine the order of modules within each row. For each row of photovoltaic strings, the edge points of the first and last photovoltaic modules are taken, and combined with the laser point cloud data to generate safe drone hovering start and end points. The drone flies at a certain height above the plane where the photovoltaic panels are located, and flies along the axis of the photovoltaic strings, thereby realizing the planning of the drone's flight path. This ensures that the flight path extends along the side of the photovoltaic strings and covers all support connection points. At the same time, the flight control parameters are automatically set. These flight control parameters can be set according to actual conditions such as terrain. For example, the flight altitude is 2-5 meters above the photovoltaic panels, the flight speed is 1-2 meters / second, the target downward angle is -60°, and the shooting interval is 0.5 seconds.

[0116] In another optional approach, based on the actual design drawings and site layout of the photovoltaic power station, the side flight points of the photovoltaic strings and the shooting points at the bracket connections are manually marked in the power station, and the shooting points are connected in sequence to form the flight path of the drone. By taking test shots on site, the top-down angle of the gimbal camera is adjusted, and the angle that can most clearly show the bolt loosening characteristics is selected as the target top-down angle. At the same time, the flight control parameters such as flight altitude, flight speed, and shooting interval are optimized based on the results of the test shots on site to ensure that the bolt images collected are clear.

[0117] Step S02: Based on the flight path and flight control parameters, generate the first control command for the UAV and send the first control command to the UAV's flight control system. The flight control system responds to the first control command and controls the UAV to fly along the side of the photovoltaic string according to the flight path. During the flight of the UAV, the system controls the image acquisition module of the UAV to view the support connection point below the photovoltaic string from the target top angle, acquire bolt images of the support connection point below the photovoltaic string, and transmit the bolt images to the server.

[0118] The sides of a photovoltaic (PV) string refer to the areas on both sides along the axis of the PV string. This area is the flight zone for drones to photograph the connection points of the support structures beneath the PV string. Flying the drone in this area avoids collisions with the PV panels and support structures while ensuring that the camera captures a clear image of the bolts from a top-down perspective. The top-down perspective can be perpendicular to the bolt axis or along the tilt direction of the PV module plane. At this angle, a frontal view of the bolts can be captured with minimal distortion, enabling more accurate image detection.

[0119] In one alternative approach, the planned flight path and flight control parameters are imported into the UAV's flight control system. The UAV autonomously flies along the side of the photovoltaic string according to the instructions of the flight control system. During the flight, the flight control system automatically controls the image acquisition module, such as the gimbal camera, to maintain a downward angle towards the support connection point below the photovoltaic string and automatically triggers the camera to take pictures at preset shooting intervals, thereby realizing the automated acquisition of bolt images. The acquired images are stored in real time in the UAV's onboard storage unit.

[0120] In another optional approach, an operator remotely controls the drone using a remote controller. Following a planned flight path, the operator manually controls the drone to fly along the side of the photovoltaic strings. During flight, the operator adjusts the image acquisition module, such as the gimbal camera, using the remote controller to maintain a top-down view and align it with the support connection point. Upon reaching the preset shooting point, the operator manually triggers the camera to take a picture. Simultaneously, the operator observes the image acquisition quality through the drone's real-time image transmission, adjusting the flight attitude and shooting angle as needed to ensure clear and effective image acquisition. In this way, the drone's flight control system can respond to the operator's control commands and acquire bolt images.

[0121] Understandably, controlling the drone to fly and collect bolt images according to the preset path and parameters ensures the drone's safe flight on the side of the photovoltaic string, while ensuring that the gimbal camera captures bolt images at the connection point of the support below the photovoltaic string from a top-down angle, maximizing the presentation of bolt loosening characteristics and providing high-quality raw image data for subsequent inspection processes.

[0122] Step S10: Obtain bolt images at the bracket connections of the photovoltaic strings in the photovoltaic power station; wherein, the server obtains bolt images sent by the flight control system and performs subsequent bolt loosening detection operations.

[0123] Step S20: Determine the reconstructed image corresponding to the bolt image through the target generative adversarial network model, and determine the reconstruction error between the bolt image and the reconstructed image. The target generative adversarial network model is trained through fastened bolt image samples. The reconstruction error is used to characterize the looseness of the bolt to be detected in the bolt image, and the reconstruction error is positively correlated with the looseness. Step S30: Determine the first loosening detection result of the bolt to be tested based on the reconstruction error.

[0124] This embodiment clarifies the data acquisition method of the UAV flying along the side of the photovoltaic string, avoiding the collision risk and problems of background clutter and uneven lighting associated with UAVs flying under photovoltaic panels, thus improving the safety of UAV flight. It also effectively reduces background interference in the images, improving the quality of the bolt images. Furthermore, the target top-down angle of the UAV is set so that the camera optical axis forms the optimal observation angle with the side of the bolt, clearly capturing the bolt's side loosening features such as bolt elongation, tilting, and gaps between the nut and the connector. This solves the problem of feature compression caused by traditional UAVs shooting directly at the bolt, providing high-quality image data for subsequent feature extraction and model analysis, and improving the accuracy of bolt loosening detection.

[0125] Based on the above embodiments of this application, in the fifth embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 5 Methods for detecting loose bolts include: Step S01: Determine the flight path and flight control parameters of the UAV, wherein the flight control parameters include at least the target's top-down angle; Step S02: Based on the flight path and flight control parameters, generate the first control command for the UAV and send the first control command to the UAV's flight control system. The flight control system responds to the first control command and controls the UAV to fly along the side of the photovoltaic string according to the flight path. During the flight of the UAV, the system controls the image acquisition module of the UAV to view the support connection point below the photovoltaic string from the target top angle, acquire bolt images of the support connection point below the photovoltaic string, and transmit the bolt images to the server.

[0126] Step S10: Obtain images of bolts at the bracket connections of the photovoltaic strings in the photovoltaic power station; Step S20: Determine the reconstructed image corresponding to the bolt image through the target generative adversarial network model, and determine the reconstruction error between the bolt image and the reconstructed image. The target generative adversarial network model is trained through fastened bolt image samples. The reconstruction error is used to characterize the looseness of the bolt to be detected in the bolt image, and the reconstruction error is positively correlated with the looseness. Step S30: Determine the first loosening detection result of the bolt to be tested based on the reconstruction error.

[0127] Step S110: Determine the bolt loosening rate within the target area where the loose bolt is located; After the loosening test result is determined by the above method, each bolt to be tested can be marked according to the loosening test result, specifically as a loose bolt or a fastening bolt.

[0128] The target area can be a row of photovoltaic strings or a power station zone, and can be set according to the actual situation.

[0129] Bolt loosening rate refers to the percentage of bolts identified as loose within a target area out of the total number of bolts inspected in that area. It is an indicator that characterizes the overall risk of bolt loosening in the target area and can directly reflect the operational health status of the bolts in that area.

[0130] In one optional approach, based on the photovoltaic string clustering results completed in the early stage of 3D modeling of the photovoltaic power station, each row of photovoltaic strings is designated as an independent target area. The loose bolts are assigned to their corresponding target areas according to the string information. The number of loose bolts in each target area and the total number of bolts inspected are then counted. The bolt loosening rate is calculated as (number of loose bolts in the target area / total number of bolts inspected in the target area). The formula calculates the bolt loosening rate for each target area and automatically generates statistical results of the area loosening rate.

[0131] In another optional approach, the entire photovoltaic power station is divided into multiple regular geographic grids according to a preset geographic size, serving as the basic target area. Each grid is assigned a unique regional code. The geographic coordinates of the loose bolts are used to match the target area of ​​the geographic grid to which they belong. At the same time, the number of loose bolts in each grid and the total number of bolts detected are counted. If the bolt detection coverage rate in a certain grid is lower than the preset standard, the data is corrected by combining the bolt loosening situation of the surrounding adjacent grids. Then, the accurate bolt loosening rate of each target area is calculated using the above formula.

[0132] Step S120: If the bolt loosening rate is greater than the preset bolt loosening rate, adjust the flight path and flight control parameters; The preset bolt loosening rate refers to the critical value of bolt loosening rate set in advance based on the bolt installation process, operating years, natural environment and operation and maintenance safety standards of the photovoltaic power station. If the actual loosening rate exceeds this value, it indicates that the bolt loosening risk in the target area is high.

[0133] In one optional approach, risk levels are pre-defined for different bolt loosening rates exceeding the limit, and standardized flight paths and flight control parameter adjustment schemes are developed for each risk level. When the bolt loosening rate in the target area exceeds the preset value, the corresponding risk level is determined based on the extent of the exceedance, and the corresponding adjustment scheme is directly invoked. This involves adding waypoints and optimizing local paths on the flight path for that area, and standardizing adjustments to flight control parameters such as flight speed, shooting interval, and camera angle. The adjustments to the flight path and flight control parameters include at least one of the following: adding hovering points above the loose bolts; adjusting the UAV's pitch angle to acquire bolt images from multiple angles; reducing the UAV's flight altitude; and increasing the image acquisition resolution of the UAV's image acquisition module.

[0134] Step S130: Based on the adjusted flight path and flight control parameters, generate a second control command for the UAV and send the second control command to the UAV's flight control system. The flight control system responds to the second control command and controls the UAV to fly according to the adjusted flight path and flight control parameters.

[0135] In one alternative approach, the adjusted flight path waypoint coordinates and flight control parameters are imported into the UAV flight control system, and the UAV's autonomous flight mode is activated. The UAV will start from its current position and automatically plan its flight trajectory according to the optimized flight path. During the flight, it will strictly follow the adjusted control parameters to complete bolt image acquisition. The acquired images are transmitted in real time to the bolt detection system for secondary analysis, thereby realizing automated re-inspection and patrol of high-risk areas.

[0136] In another optional approach, the operator visualizes the adjusted flight path and control parameters on the UAV ground station and remotely controls the UAV via the UAV remote controller to fly along the optimized path. During flight, the operator fine-tunes the UAV's flight attitude and camera angle based on the real-time image transmission from the ground station to ensure strict adherence to the adjusted control parameters. For critical locations with a high density of loose bolts, the operator manually controls the UAV to perform fixed-point hovering and filming, enabling refined and precise re-inspection of high-risk areas.

[0137] In this embodiment, a linkage adjustment mechanism between bolt loosening rate and inspection strategy is established. When the bolt loosening rate in the target area exceeds the preset bolt loosening rate, the flight path and flight control parameters are automatically optimized, further improving the accuracy of bolt loosening detection results.

[0138] Based on the above embodiments of this application, in the sixth embodiment of this application, the content that is the same as or similar to the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, the bolt loosening detection method includes: Step S210: Based on the reconstruction error determined in each historical loosening detection of the target bolt, construct a time series of the reconstruction error of the target bolt, where the target bolt is any bolt at the bracket connection of any photovoltaic string; The target bolt is a single bolt at any photovoltaic string support connection point in a photovoltaic power station. Each target bolt is assigned a unique identifier, which enables the traceability of data from previous tests.

[0139] Each historical loosening detection refers to all past bolt loosening detection work carried out on the target bolt during the operation and maintenance of the photovoltaic power station, including detections at different inspection batches and at different time points. Each detection will record the corresponding detection time and reconstruction error value.

[0140] In one optional approach, an independent health status database is established for each target bolt, storing its unique identifier, timestamps of each loosening detection, reconstruction error values, detection batch, site environment, and other information. Through database retrieval commands, all reconstruction error values ​​of the target bolt are extracted according to the detection timestamp. The system automatically arranges the detection time and reconstruction error values ​​in a one-to-one correspondence to generate a time series of reconstruction errors.

[0141] In another optional approach, all historical reconstruction error data of the target bolts are first extracted from the bolt detection database. The data is then cleaned to remove abnormal reconstruction error values ​​caused by image blurring, model failure, or human error. For missing detection data caused by inspection omissions or equipment failures, linear interpolation is used to complete the data based on the reconstruction error values ​​and time intervals between two adjacent valid detections. The cleaned and completed detection time and reconstruction error value data are then arranged in chronological order according to the detection time to construct an accurate and complete reconstruction error time series.

[0142] Step S220: Based on the time series of reconstruction error, predict the loosening trend of the target bolt.

[0143] The time series of reconstruction error refers to the numerical sequence formed by arranging the reconstruction error values ​​of the target bolt in each historical loosening detection in chronological order of the detection time. Each data point in the sequence contains the correspondence between the detection time and the reconstruction error value, which can intuitively reflect the change law of the target bolt reconstruction error over time.

[0144] The loosening trend refers to the direction, magnitude, and rate of change of the reconstruction error of the target bolt over time. Since the reconstruction error is positively correlated with the degree of bolt loosening, the time change trend of the reconstruction error is the bolt loosening trend, which mainly includes types such as steady, slow rise, rapid rise, and fluctuating rise, and can reflect the evolution law of the bolt loosening state.

[0145] In one alternative approach, the reconstruction error time series of the target bolt is curve fitted to obtain the fitted curve and fitting formula of the reconstruction error changing with time; based on the slope and curvature of the fitted curve, the direction and rate of change of the reconstruction error of the target bolt are determined, and the loosening trend of the bolt in the future is predicted. For example, if the fitted curve is an upward curve with a gradually increasing slope, the predicted loosening trend of the bolt is a rapid increase.

[0146] In another alternative approach, the reconstruction error time series data of the target bolt is input into a trained time series prediction model. Through the model's learning and extrapolation of historical data, the reconstruction error values ​​at multiple future time points are predicted. The predicted reconstruction error values ​​are then arranged in chronological order to form a future reconstruction error change trend curve. Based on the change characteristics of this curve, the loosening trend of the target bolt is determined.

[0147] In this embodiment, by constructing a time series of reconstruction errors, discrete detection data is transformed into continuous data that reflects the changing patterns of bolt conditions, thus achieving accurate prediction of bolt loosening trends. The loosening trend prediction results provide photovoltaic power plant operation and maintenance personnel with a clear direction for bolt condition evolution, enabling them to grasp the development pattern of bolt loosening in advance and formulate targeted inspection and maintenance plans. This effectively avoids safety accidents such as photovoltaic panel displacement and collapse caused by sudden bolt loosening, thereby improving the operational safety of photovoltaic power plant equipment.

[0148] Based on the above embodiments of this application, in the seventh embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, the bolt loosening detection method includes: Step S310: Based on the reconstruction error determined in each historical loosening detection of the target bolt, construct a time series of the reconstruction error of the target bolt, where the target bolt is any bolt at the bracket connection of any photovoltaic string; Step S320: Based on the time series of reconstruction error, assess the loosening rate of the target bolt and the degradation risk level of the target bolt.

[0149] Loosening rate refers to how quickly the reconstruction error of a target bolt changes over time. It is an indicator of the speed at which bolt loosening develops. It is usually expressed as the increase in reconstruction error per unit time. The higher the loosening rate, the faster the bolt loosening state develops and the higher the safety risk.

[0150] The degradation risk level refers to the classification of the safety risk of bolt loosening and degradation based on factors such as the loosening rate of the target bolt, the current value of the reconstruction error, and the operating environment. It is a qualitative indicator that characterizes the degree of bolt loosening risk. For example, it can be divided into four levels: low, medium, high, and extremely high.

[0151] In one alternative approach, a linear regression fitting is performed on the reconstruction error time series to calculate the slope of the fitted curve. This slope represents the average loosening rate of the target bolt, indicating the average increase in reconstruction error per unit time. Simultaneously, a degradation risk level assessment system is constructed, using the loosening rate, the current value of the reconstruction error, and the severity of the environment in which the target bolt is located as assessment indicators. Different weight coefficients are assigned to each indicator, and a comprehensive risk assessment score is obtained through weighted calculation. Based on the range of the comprehensive score, the degradation risk level of the target bolt is divided into four levels: low, medium, high, and extremely high.

[0152] In this embodiment, the loosening change pattern of the target bolt is transformed into loosening rate and qualitative degradation risk level, so as to realize accurate quantitative assessment of bolt loosening risk and provide maintenance personnel with a clear basis for judging the degree of bolt risk.

[0153] Based on the above embodiments of this application, in the eighth embodiment of this application, the content that is the same as or similar to the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, the bolt loosening detection method includes: Step S410: Based on the reconstruction error determined in each historical loosening detection of the target bolt, construct a time series of the reconstruction error of the target bolt, where the target bolt is any bolt at the bracket connection of any photovoltaic string; Step S420: Based on the time series of reconstruction errors, generate maintenance recommendations for the target bolt.

[0154] Maintenance recommendations refer to the specific maintenance operation guidelines for the target bolt, such as inspection, tightening, and replacement, based on the reconstruction error time series analysis results of the target bolt.

[0155] In one alternative approach, the reconstruction error time series of the target bolt is trend-followed to predict the changes in reconstruction error over a future period. This is combined with information such as the bolt's model and specifications, operating environment, and design service life for comprehensive analysis. Based on the analysis results, personalized maintenance recommendations are developed for the target bolt. These recommendations not only specify the type, time, and frequency of maintenance operations but also provide specific operational requirements based on the bolt's loosening characteristics and environment. For example, for bolts with high loosening rates in windy areas, it is recommended to use anti-loosening nuts for tightening and increase the inspection frequency to once a month.

[0156] In this embodiment, maintenance recommendations based on the reconstruction error time series, combined with the historical loosening changes and future development trends of bolts, enable the operation and maintenance work to shift from passive response to proactive prediction. This allows for early maintenance measures to be taken for high-risk bolts, effectively avoiding various safety accidents caused by bolt loosening and improving the operational safety and stability of photovoltaic power plants.

[0157] In one embodiment, reference is made to Figure 6 , Figure 6 The flowchart of the overall technical solution of this application is as follows: (1) 3D modeling of photovoltaic power station, string aggregation and under-panel route planning, used to determine the flight path. (2) Under-panel image acquisition, used to obtain bolt images. (3) Two bolt loosening detection methods are provided: one is bolt detection and loosening rule determination, that is, the method of determining the loosening detection result based on bolt features mentioned above in this application, and the other is GAN detection based on normal samples, that is, the method of determining the loosening detection result through the target generative adversarial network model mentioned above in this application. (4) Multi-feature fusion determination, including combining the two methods in (3) above to determine the loosening detection result. (5) Dynamic route iteration, for loose bolts, re-planning the flight path and adjusting the flight control parameters. (6) Finally, the loosening evolution trend analysis and maintenance decision are performed.

[0158] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the bolt loosening detection method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0159] Based on the same inventive concept, this application provides a server, the server comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the bolt loosening detection method in the above embodiments.

[0160] like Figure 7As shown, the server may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for server operation. The processing device 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the server to communicate wirelessly or wiredly with other devices to exchange data. Although servers with various systems are shown in the figure, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0161] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0162] The server provided in this application, employing the bolt loosening detection method described in the above embodiments, can improve the detection rate of loose bolts in photovoltaic power plants. Compared with the prior art, the beneficial effects of the server provided in this application are the same as those of the bolt loosening detection method provided in the above embodiments, and other technical features of the server are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0163] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0164] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included 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 bolt loosening, characterized in that, The method includes: Obtain images of bolts at the support connections of photovoltaic strings in a photovoltaic power plant; The reconstructed image corresponding to the bolt image is determined by a target generative adversarial network model, and the reconstruction error between the bolt image and the reconstructed image is determined. The target generative adversarial network model is trained by fastening bolt image samples. The reconstruction error is used to characterize the degree of looseness of the bolt to be detected in the bolt image, and the reconstruction error is positively correlated with the degree of looseness. Based on the reconstruction error, the first loosening detection result of the bolt to be tested is determined.

2. The bolt loosening detection method as described in claim 1, characterized in that, The step of determining the reconstructed image corresponding to the bolt image through a target generative adversarial network model, and determining the reconstruction error between the bolt image and the reconstructed image, includes: The bolt image is mapped to a low-dimensional bolt feature vector through the encoding network of the target generative adversarial network model. The low-dimensional bolt feature vector is reconstructed by the decoding network of the target generative adversarial network model to obtain the reconstructed image. The mean square error between the bolt image and the reconstructed image is determined pixel by pixel and is used as the reconstruction error between the bolt image and the reconstructed image.

3. The bolt loosening detection method as described in claim 1, characterized in that, Before determining the reconstructed image corresponding to the bolt image through the target generative adversarial network model, and determining the reconstruction error between the bolt image and the reconstructed image, the method further includes: Identify the bolt region in the bolt image; The bolt image is cropped into multiple bolt region images, and each bolt region image is input into the target generative adversarial network model. The step of determining the reconstructed image corresponding to the bolt image through a target generative adversarial network model, and determining the reconstruction error between the bolt image and the reconstructed image, includes: The target generative adversarial network model is used to determine the reconstructed image corresponding to the bolt region image, and the reconstruction error between the bolt region image and the reconstructed image corresponding to the bolt region image is determined.

4. The bolt loosening detection method according to any one of claims 1 to 3, characterized in that, The step of determining the first loosening detection result of the bolt to be tested based on the reconstruction error includes: If the reconstruction error is greater than or equal to the preset error value, the first loosening detection result is determined to be that the bolt to be detected is loose. If the reconstruction error is less than the preset error value, the first loosening detection result is determined to be that the bolt to be tested is tight.

5. The bolt loosening detection method as described in claim 1, characterized in that, After obtaining the bolt images at the bracket connections of the photovoltaic strings in the photovoltaic power station, the method further includes: Extract the bolt features of the bolt to be detected from the bolt image, wherein the bolt features include a first feature and / or a second feature, the first feature includes the length of the bolt extending out of the nut, and the second feature includes at least one of the following: photovoltaic bracket gap, photovoltaic module alignment, and bolt posture; The second loosening detection result of the bolt to be tested is determined based on the first feature and / or at least one of the second features; Based on the first loosening detection result and / or the second loosening detection result, the final loosening detection result of the bolt to be tested is determined.

6. The bolt loosening detection method as described in claim 5, characterized in that, Based on the first feature, the second loosening detection result of the bolt to be tested includes: Obtain other bolts located in the same row as the bolt to be tested; Determine the difference between the length of the other bolts extending out of the nut and the length of the bolt to be tested extending out of the nut; If the difference is greater than a preset difference, the second loosening detection result is determined to be that the bolt to be tested is loose.

7. The bolt loosening detection method as described in claim 1, characterized in that, Applied to a server, before acquiring the bolt images at the bracket connections of photovoltaic strings in a photovoltaic power station, the process includes: Determine the flight path and flight control parameters of the UAV, wherein the flight control parameters include at least the target's top-down angle; Based on the flight path and the flight control parameters, a first control command for the UAV is generated and sent to the flight control system of the UAV. The flight control system responds to the first control command, controls the UAV to fly along the side of the photovoltaic string and follow the flight path, and during the flight of the UAV, controls the image acquisition module of the UAV to view the support connection below the photovoltaic string from the target top angle, acquire bolt images of the support connection below the photovoltaic string, and transmit the bolt images to the server.

8. The bolt loosening detection method as described in claim 7, characterized in that, After determining the loosening detection result of the bolt to be tested based on the reconstruction error, the method further includes: Determine the bolt loosening rate within the target area where the loose bolts are located; If the bolt loosening rate is greater than the preset bolt loosening rate, the flight path and the flight control parameters are adjusted. Based on the adjusted flight path and flight control parameters, a second control command for the UAV is generated and sent to the UAV's flight control system. The flight control system responds to the second control command and controls the UAV to fly according to the adjusted flight path and flight control parameters.

9. The bolt loosening detection method as described in claim 8, characterized in that, The adjustment of the flight path and the flight control parameters includes at least one of the following: Add a hovering point above the loose bolt; Adjust the drone's pitch angle to acquire bolt images from multiple angles; Lower the flight altitude of the drone; Improve the image acquisition resolution of the image acquisition module on the UAV.

10. The bolt loosening detection method as described in claim 1, characterized in that, The bolt loosening detection method further includes: Based on the reconstruction error determined in each historical loosening detection of the target bolt, a time series of the reconstruction error of the target bolt is constructed, wherein the target bolt is any bolt at the bracket connection of any photovoltaic string; Based on the time series of the reconstruction error, the loosening trend of the target bolt is predicted.

11. The bolt loosening detection method as described in claim 1, characterized in that, The bolt loosening detection method further includes: Based on the reconstruction error determined in each historical loosening detection of the target bolt, a time series of the reconstruction error of the target bolt is constructed, wherein the target bolt is any bolt at the bracket connection of any photovoltaic string; Based on the time series of the reconstruction error, assess the loosening rate of the target bolt and the degradation risk level of the target bolt; or, based on the time series of the reconstruction error, generate maintenance recommendations for the target bolt.

12. A server, characterized in that, The server includes: a memory, a processor, and a bolt loosening detection program stored in the memory and executable on the processor, the bolt loosening detection program being configured to implement the steps of the bolt loosening detection method as described in any one of claims 1 to 11.

Images