Photovoltaic panel fault detection method, device and equipment

Abstract

The invention provides a photovoltaic panel fault detection method, device and equipment, and relates to the technical field of photovoltaic power generation. According to the invention, real-time acquisition of the surface image of the photovoltaic panel is realized by deploying the high-speed visible light lens, and the timeliness of fault monitoring is ensured; characterization is carried out based on the extracted vector angle and color space multi-dimensional fusion features, and the characterization capability of fault features and the robustness of resisting illumination, stains and other environmental interferences are enhanced. A rough matching-fine matching hierarchical matching strategy taking a local fault feature library as a reference is adopted, the detection speed and the recognition precision are both considered, and the real-time requirement of online monitoring of a large-scale photovoltaic array is met; and finally, the fault type, the position coordinates and the grade information are output, and an alarm is given according to the information, so that an end-to-end automatic process from image acquisition to fault positioning and grading alarm is realized, and the intelligent level and the fault response efficiency of operation and maintenance of the photovoltaic power station are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method, apparatus and equipment for detecting photovoltaic panel faults. Background Technology

[0002] With the acceleration of the global energy transition, photovoltaic power generation, as a major clean energy source, has seen continuous growth in installed capacity. To date, the global cumulative installed photovoltaic capacity has exceeded 1.5TW. However, module failures caused by long-term outdoor operation (such as hot spots, microcracks, and snail trails) seriously affect power generation efficiency and system lifespan. Fault detection technology has become crucial to ensuring the profitability of power plants.

[0003] Traditional manual inspection methods have gradually become a fundamental bottleneck restricting the efficiency and safety of power plant operation and maintenance. This method heavily relies on the personal experience and subjective judgment of inspectors, resulting in inconsistent inspection results, a high risk of missed inspections and misjudgments, long inspection cycles, and high labor costs, making it difficult to achieve timely and comprehensive inspections of large-scale photovoltaic arrays. Especially in harsh environments or high-altitude work scenarios, manual inspections also pose safety hazards. This inefficient and unscalable operation and maintenance model can no longer meet the practical needs of intelligent and lean management of photovoltaic power plants. Summary of the Invention

[0004] This invention provides a method, apparatus, and equipment for detecting photovoltaic panel faults, which solves the technical problem of low efficiency in traditional manual inspection methods.

[0005] In a first aspect, the present invention provides a method for detecting photovoltaic panel faults, the method comprising: acquiring surface images of the photovoltaic panel in real time using a high-speed visible light lens deployed near the photovoltaic panel; extracting features based on the surface images of the photovoltaic panel to obtain fault features of the photovoltaic panel, the fault features including vector angular features and color space features; obtaining matching results based on the fault features and a local fault feature library using a hierarchical matching strategy of coarse matching and fine matching; determining fault information of the photovoltaic panel based on the matching results, the fault information including fault type, location coordinates and fault level; and issuing a fault alarm based on the fault information.

[0006] Secondly, embodiments of the present invention provide a photovoltaic panel fault detection device. This device includes a communication module and a processing module. The communication module is used to acquire real-time surface images of the photovoltaic panel using a high-speed visible light lens deployed near the photovoltaic panel. The processing module is used to extract features based on the surface images of the photovoltaic panel to obtain fault features of the photovoltaic panel, including vector angle features and color space features. Based on the fault features and a local fault feature library, a hierarchical matching strategy of coarse matching and fine matching is adopted to obtain matching results. Based on the matching results, fault information of the photovoltaic panel is determined, including fault type, location coordinates, and fault level. Based on the fault information, a fault alarm is issued.

[0007] Thirdly, embodiments of the present invention provide an electronic device including a memory and a processor. The memory stores a computer program, and the processor is configured to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0008] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0009] This invention provides a method, apparatus, and equipment for photovoltaic panel fault detection. By deploying a high-speed visible light lens, the invention achieves real-time image acquisition of the photovoltaic panel surface, ensuring timely fault monitoring. It uses extracted vector angles and multi-dimensional fusion features in color space for characterization, enhancing the characterization capability of fault features and robustness against environmental interference such as light and dirt. A coarse-fine matching hierarchical matching strategy based on a local fault feature library is adopted, balancing detection speed and recognition accuracy to meet the real-time requirements of large-scale photovoltaic array online monitoring. Finally, the fault type, location coordinates, and level information are output, and alarms are triggered accordingly. This realizes an end-to-end automated process from image acquisition to fault location and hierarchical alarm, significantly improving the intelligence level and fault response efficiency of photovoltaic power plant operation and maintenance. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1This is a schematic flowchart of a photovoltaic panel fault detection method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a photovoltaic panel fault detection device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0013] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0014] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include other steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or device.

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0016] As described in the background section, traditional detection methods have significant limitations: manual inspection relies on experience-based judgment, resulting in low efficiency and strong subjectivity; while infrared thermal imaging can identify hot spots, its sensitivity to subtle defects such as early-stage microcracks is insufficient; electroluminescence detection offers high accuracy but requires power-off operation, making large-scale application difficult. Furthermore, existing methods mostly remain at the offline detection stage, failing to achieve real-time early warning and full lifecycle management, thus hindering the improvement of intelligent operation and maintenance levels in photovoltaic power plants. Therefore, developing efficient, accurate, and low-cost online fault detection technology is of significant practical importance for promoting the high-quality development of the photovoltaic industry.

[0017] This invention studies a fault detection algorithm that can be deployed on the device side, enabling real-time data acquisition and edge computing analysis. This avoids the latency issues of traditional centralized detection and improves fault response speed. Its advantages include: no need for large-scale data transmission, reducing network bandwidth pressure and costs; support for 24 / 7 online monitoring, enabling timely detection of early potential problems; and, combined with edge intelligent algorithms, localized data processing to ensure system security and privacy, providing underlying technical support for the intelligent operation and maintenance of photovoltaic power plants.

[0018] This invention proposes a photovoltaic panel fault detection method with a closed-loop "offline modeling-online detection-iterative optimization" system. The core relies on the fusion of multi-dimensional features in vector angle and color space and a lightweight intelligent algorithm to achieve accurate and real-time identification of photovoltaic panel faults and dynamic model adaptation.

[0019] This invention is supported by both offline and online data sources: In the offline stage, data is acquired from the enterprise's self-built image database. Data cleaning is completed through vector angle calculation, data quality screening, and invalid threshold filtering. DBSCAN unsupervised classification (combined with contour coefficient optimization parameters) is used to achieve fault data clustering. A lightweight CNN with adaptive convolution kernel selection is used to extract 64-dimensional core fault features. Finally, a nonlinear mapping model between features and fault types is constructed based on SVM to form a fault model knowledge base. In the online stage, photovoltaic panel images are acquired in real time using a high-speed visible light lens. After localized rapid preprocessing, feature matching is performed with the local knowledge base through a hierarchical strategy of "coarse matching (hash mapping + Euclidean distance) - fine matching (improved cosine similarity)". The fault type, location, and level are quickly output and an alarm is pushed. In the iterative optimization stage, unknown fault samples are manually reviewed in the cloud. The feature library and model parameters are incrementally updated. The method can adapt to new faults without reconstructing the model, and the performance of the method can be continuously upgraded.

[0020] This invention enhances anti-interference capabilities through multi-feature fusion, reduces equipment deployment costs through lightweight algorithm design, and balances detection efficiency and generalization ability through hierarchical matching and incremental update mechanisms. It can effectively identify 12 common and new types of photovoltaic panel faults, such as hot spots, cracks, and desoldering, meeting the real-time and accuracy requirements of online detection in large-scale photovoltaic power plants.

[0021] like Figure 1 As shown, the present invention provides a method for detecting photovoltaic panel faults. The method includes steps S101-S105.

[0022] S101. Real-time acquisition of surface images of photovoltaic panels using high-speed visible light lenses deployed near the photovoltaic panels.

[0023] In some embodiments, the present invention includes two types of data sources: offline data sources and online data sources. Offline data sources are primarily used for offline analysis and modeling of the method. The main implementation process is as follows: (a) Data Acquisition. In this invention, an offline analysis is performed using a self-built image database. By parsing the target data indicators of the images in the self-built image database, an offline fault database for this patent is established.

[0024] (b) Data cleaning. Based on the data acquisition in the previous step, this invention identifies faulty photovoltaic panels based on the vector angle and color space of the image. The specific implementation method is as follows: 1. Establish an image fault database relationship model, calculate the vector angle and color space parameters for the specified position of the image and store them in the database. This includes the following process: Establish an image fault database relationship model, that is, define the database table and the data table parameter structure. In this invention, the table name is defined as follows: solar_fault. The parameter structure includes the unique identifier ID of the image, the coordinate value of the upper left corner of the image target area, the target value of the lower right corner of the image target area, the horizontal vector angle sequence of the image target area, the vertical vector angle sequence of the image target area, the color space feature vector of the image target area, etc. The calculation process of the vector angle sequence of the image target area is as follows: a) Define the interpolation interval pol, b) Interpolate from the x-coordinate of the upper left corner of the target area to the x-coordinate of the lower right corner, calculate a vector angle every k pixels, c) Obtain the complete vector angle sequence of the target area. The formula for calculating the vector angle is shown in formula (1): ; Where a is the color space feature vector corresponding to white at position (x1, y1) on the photovoltaic panel, and b is the color space feature vector corresponding to the actual color at position (x1, y1) on the photovoltaic panel. The values ​​in the calculated vector angle sequence are as follows. The eigenvector is obtained and calculated as follows: First, the value of each color component in the RGB color gamut at position (x, y1) on the photovoltaic panel is calculated. Second, the value of each component in the HSV color gamut is calculated. Then, the Z-zero normalization method is used to standardize and normalize the color component values ​​in the RGB color gamut. Finally, the Z-zero normalization is combined with the HSV color gamut components to form the color vector at the current position. The Z-zero normalization calculation method is shown in the following formula: ; in, For each value in the feature vector of the color space, Let σ be the mean of the eigenvectors, σ be the standard deviation of the eigenvectors, and Z be the standardized value. Finally, the calculated vector angle sequence is weighted and combined with the color space eigenvectors to obtain the total eigenvector at position (x1, y1) on the photovoltaic panel. The weighted calculation formula is as follows: ; in, This is the total eigenvector. These are the weighting coefficients for the vector angular characteristics. These are the weighting coefficients for the feature vectors in the color space.

[0025] For example, embodiments of the present invention can define a custom data quality coefficient to distinguish between high-quality and low-quality data sources, laying the foundation for subsequent inference of photovoltaic panel fault types. The definition process and method are as follows: a) Obtain images of fault-free photovoltaic panels taken under good lighting conditions, and establish a fault-free image data standard based on these images.

[0026] b) Calculate the cosine of the angle between the vector of the image data taken under good lighting conditions in the database and the vector of the data standard. The calculation formula is as follows: ; Where 'a' is the standard feature vector of fault-free image data, and 'b' is the feature vector of image data captured under good lighting conditions. The angle between the cosines of the vectors.

[0027] c) Define the data quality coefficient as the cosine value of the angle between the vectors, ranging from 0 to 1.

[0028] d) The calculation method for image data captured under other conditions / fault conditions is similar.

[0029] e) If the data quality coefficient is less than 0.6, the current record is considered to be of too low quality and cannot be used, and is deleted from the database record.

[0030] For example, in this embodiment of the invention, an invalid data threshold can be defined. When the parameter value in the database does not meet the threshold range, the corresponding record will be deleted from the database. The specific process is as follows: a) Obtain images of fault-free photovoltaic panels taken under good lighting conditions to establish a standard for valid fault-free image data.

[0031] b) Calculate the color space complexity in the data standard. Complexity is defined as the number of color features in the color space domain. The calculation method is as follows: This is implemented using the KMeans classification algorithm. The core principle formula of KMeans is as follows: ; Where K is the preset number of clusters, For the k-th cluster, Let the cluster center of the k-th cluster be... It is the i-th sample in the k-th cluster.

[0032] c) Using the color space complexity * 0.8 and color space complexity * 1.2 in the data standard as the upper and lower limits of the threshold, delete the database records outside the threshold range.

[0033] For example, embodiments of the present invention can perform data analysis. The steps are as follows: a) Select the vector angle and color space at the same location in the database parameters to build a classification model for the research data elements. The initial value of the minimum number of samples for each class is set to 8, and the neighborhood radius is also set to 8. The classification algorithm is DBSCAN, and its core principle formula is as follows: ; in, For a single sample in the dataset, For the radius of the domain, The number of samples (including samples) ), then at this time the sample The core point. Then... Neighborhood The calculation formula is shown below: ; Where D is the sample set, q is the other sample points in the sample set except p, and dist() represents the distance between two samples.

[0034] Furthermore, when considering samples from non-core points... ,like At any core point of Within the domain, then the sample These are called boundary points, and their specific formulas are as follows: ; For samples that are not in the core area like Not at any core point of Within the neighborhood, it is called For noise points, the formula is as follows: ; Therefore, as long as any two sample points are density-directed or density-reachable, then the two sample points are classified into the same cluster.

[0035] b) Input the data elements into the classification model and perform unsupervised classification to obtain the initial classification results.

[0036] c) Use the standard vector angles and color space parameters in the database to perform category matching. Calculate the similarity between each object in each current category set and the standard parameters in the database. The similarity calculation method is cosine similarity. If the similarity is less than 0.6, the classification is restarted. The minimum number of samples in the classification model is 6, and the neighborhood radius is 6. The calculation formula for cosine similarity is shown in Equation 4 above.

[0037] d) Repeat the above process. When the similarity is higher than 0.6 and less than 0.8, continue the classification process, but reduce the decreasing interval of the minimum number of samples to 1, and make the same modification to the neighborhood radius.

[0038] e) When the similarity is greater than or equal to 0.8, stop the above loop process, and finally obtain the classification model parameters suitable for the data involved in this patent and save them to a local file.

[0039] S102. Based on the surface image of the photovoltaic panel, feature extraction is performed to obtain the fault characteristics of the photovoltaic panel.

[0040] In this embodiment of the application, the fault features include vector angle features and color space features.

[0041] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1024.

[0042] S1021. Divide the surface image into regions to obtain the color space vector of each pixel in each region.

[0043] S1022. Calculate the vector angle between the color space vector of each pixel in each region and the white standard color vector to obtain the vector angle feature.

[0044] S1023. Standardize the color space vectors of each region and convert them into color space features with uniform dimensions.

[0045] S1024. The vector angular features and color space features are weighted and fused to obtain the fault features.

[0046] For example, in this embodiment of the invention, feature extraction algorithms (CNNs) can be used to extract features for each type of fault data, and then the extracted features are saved locally for subsequent analysis. The specific implementation process and method are as follows: a) Based on the classification results, extract the corresponding vector angle and color space data from the database.

[0047] b) Calculate the smaller data dimension in each type of data. For example, if the data is a two-dimensional matrix of R*L and R < L, then select R as the convolution kernel size for feature extraction; specifically, when R = L, the classified data is square, and the convolution kernel size can be defined according to the following method: 1. Define a sequence of convolution kernel sizes. Commonly used convolution kernel sizes are 3, 5, and 7. Use different convolution kernels for feature extraction, defined as SIG; 3. Use SIG to perform the inverse process of feature extraction to obtain SIG'; 4. Compare and calculate the difference degree between SIG and SIG'. The difference degree calculation method is as follows: Calculate whether each vector angle data in SIG' is the same as the original data; 5. Select the convolution kernel size corresponding to the set with the smallest difference degree as the convolution kernel size in this patent. The padding processing method at the data edge is similar to that of the convolution kernel. The calculation process of convolution is shown in the following formula: ; where, is the size of the matrix after convolution, is the size of the input matrix for the convolution operation, is the convolution kernel size, is the edge padding, is the stride.

[0048] The calculation formula of the transposed convolution is shown as follows: ; c) Use the convolution kernel and the classified data of R*L for feature extraction, and finally obtain the feature vectors SIG of the vector angle and color space data.

[0049] (d) Data modeling. Taking the fault features saved locally as the input, use the fault detection algorithm designed in this patent to perform modeling operations on the features. The main process is to establish a non-linear correspondence between the features and the fault judgment, and form a fault model knowledge base that can be saved and recognized locally.

[0050] Exemplarily, the online data source of the present invention is mainly used for online analysis at the device end. The main implementation process is as follows: (a) Data acquisition. Use a high-speed visible light induction lens to capture photovoltaic panel image data within a certain range. After passing through the data acquisition and data cleaning modules in the algorithm model, convert the image data into recognizable digital resources and store them in the database. In order to reduce the pressure during model deployment, the image data is parsed and then uploaded to the cloud, and the model deletes the data locally.

[0051] (b) Data analysis. Use the model to perform online analysis on the data in the database, and perform feature matching in the local feature library and model library (c) Data utilization. Alarm push is implemented based on feature matching results, and the recognition results are added to the local feature and data model library to achieve iterative optimization of the algorithm model.

[0052] S103. Based on fault characteristics and the local fault characteristic library, a hierarchical matching strategy of coarse matching and fine matching is adopted to obtain the matching results.

[0053] As one possible implementation, step S103 can be specifically implemented as steps S1031-S1035.

[0054] S1031. Based on the fault features and the local fault feature library, coarse matching is performed through similarity mapping to obtain candidate fault features with a coarse matching similarity greater than the first threshold.

[0055] S1032. Based on the candidate fault features, an improved cosine similarity method is used for fine matching, which integrates the consistency of vector angle direction and the similarity of color space distribution to obtain the fine matching similarity.

[0056] S1033. If the similarity of the fine match is greater than the second threshold, then the match is considered successful.

[0057] S1034. If the fine matching similarity is less than or equal to the second threshold, or if there are no candidate fault features, then the fault features are marked as unknown fault samples.

[0058] S1035, Output the matching results.

[0059] In some embodiments, the matching results include the candidate fault feature with the highest fine-match similarity; the fault type corresponding to the candidate fault feature; and the fine-match similarity.

[0060] S104. Based on the matching results, determine the fault information of the photovoltaic panel.

[0061] In this embodiment, the fault information includes fault type, location coordinates, and fault level.

[0062] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1044.

[0063] S1041. Determine the fault type of the photovoltaic panel based on the successfully matched fault type; S1042. Calculate the location coordinates of the fault on the photovoltaic panel based on the location coordinates of the fault features in the surface image; S1043. Determine the fault level of the photovoltaic panel based on the fine matching similarity. S1044. Integrate fault type, location coordinates, and fault level into fault information.

[0064] S105. Based on the fault information, issue a fault alarm.

[0065] As one possible implementation, step S105 can be specifically implemented as steps S1051-S1053.

[0066] S1051. Generate alarm messages based on fault type and fault level.

[0067] S1052. Display alarm information on the device's display screen in real time and push it to the cloud-based operation and maintenance management platform.

[0068] S1053. Store the fault characteristics and fault information detected in this test into the local fault characteristic database.

[0069] For example, the data source for online detection is an online data source. Based on the fault model knowledge base built offline, it completes real-time parsing, feature matching, and fault judgment of photovoltaic panel image data. The core objective is to meet the real-time and accuracy requirements of online detection for large-scale photovoltaic arrays. The specific steps are as follows: (1) Online real-time data preprocessing: The photovoltaic panel image data (resolution 1920*1080) acquired by the high-speed visible light sensor first enters the local lightweight preprocessing module, reusing the data cleaning logic of the offline stage: calculating the vector angle sequence and color space feature vector of the target area of ​​the image, and filtering high-quality data through the data quality coefficient (threshold 0.6) and invalid threshold. The data quality coefficient is calculated by multi-dimensional feature weighted fusion, and the formula is as follows: ; in, This is the data quality coefficient (value range [0, 1]). For the stability weights of the vector angle sequence ( =0.6), Weights for the effectiveness of color space features ( =0.4, + =1); The stability coefficient of the vector angle sequence is calculated as follows: ; in The standard deviation of the vector angle sequence. This represents the maximum possible value of the vector angle in the photovoltaic panel image (unit: radians). The color space feature effectiveness coefficient is obtained by normalizing the variance of the color feature vector, and the formula is as follows: ; in, Let be the variance of the eigenvectors in the color space. This represents the maximum variance of color features obtained offline. Z-Score standardization is performed on the filtered data to ensure consistency with the offline feature library data format. The standardization formula is as follows: ; in, These are the standardized eigenvalues. For a single component in the original feature vector, This represents the mean of the corresponding feature dimension in the offline feature library. The standard deviation of the corresponding feature dimension in the offline feature library needs to be calculated separately for each dimension of the 64-dimensional feature vector. To reduce the computational burden on the device, the preprocessing process uses fixed-point arithmetic optimization, and a scaling factor is used to convert floating-point data to 16-bit integer data. The conversion formula is as follows: ; in, This is the scaling factor. This is a rounding function. After optimization, the processing time is controlled within 0.3s; at the same time, the preprocessed raw image data is uploaded to the cloud storage according to the requirements of 2.1.2(a), and only the standardized feature data is retained locally, saving device storage resources.

[0070] (2) Layered feature matching and fault identification: This step is the core of online detection. A layered strategy of “coarse matching-fine matching” is adopted to achieve rapid and accurate matching of fault features.

[0071] 1) Coarse matching stage: The 64-dimensional feature vector obtained from online preprocessing is quickly searched against the feature vectors in the local fault feature database. The initial similarity is calculated using Euclidean distance, the formula for which is: ; in, The Euclidean distance between two feature vectors is used (the smaller the distance, the higher the similarity). The 64-dimensional feature vector (n=64) is obtained through online preprocessing. This is a 64-dimensional feature vector from a local fault feature database. To accurately quantify similarity, two normalization mapping rules are used: Rule 1 (linear mapping) is... ; Rule 2 (non-linear mapping, improving discrimination) is: ; in, , All are similarities after mapping (values ​​range [0, 1]). This represents the maximum Euclidean distance between all feature vector pairs in the offline feature library. Nonlinear coefficients ( =0.8, controlling the slope of the mapping curve). This represents the average Euclidean distance between feature vector pairs in the offline statistical feature library. In practical applications, any rule can be selected to filter out candidate fault features with a similarity ≥ 0.7 (reducing the computational load of subsequent fine matching). This process is optimized through hash mapping, and the retrieval time is controlled within 0.2 seconds.

[0072] 2) Fine-matching stage: For candidate fault characteristics, an improved cosine similarity calculation method is introduced (fusing vector angular direction consistency and color space distribution similarity). First, the standard cosine similarity is calculated as a basis. The standard cosine similarity formula is: ; in, Cosine similarity (values ​​range from -1 to 1; the closer to 1, the stronger the directional consistency). The dot product of two eigenvectors. , Let be the magnitudes of the two eigenvectors. =64 is the feature vector dimension. Based on this, the improved cosine similarity calculation formula is: ; in, The improved overall similarity (judgment threshold ≥ 0.85) is used. The vector angular direction consistency weighting coefficient ( =0.7 (adjustable) The similarity weighting coefficient for color space distribution ( =0.3, + =1), To measure the similarity of color space feature distributions, Bach distance quantization (suitable for color histogram features) is used, with the following formula: ; in, The distance is the Bach distance (value range [0, 1]). , , respectively, are the color histogram feature vectors of the online image and the candidate fault, where m is the number of bins in the color histogram (m=256). and These are the frequency normalized values ​​of the i-th bin in the two histograms, respectively. The color similarity conversion formula based on Bach's distance is: ; The online features and candidate features are calculated using an improved cosine similarity method. If the similarity is ≥0.85, the match is considered successful and the corresponding fault type is output. If the similarity is between 0.7 and 0.85, the offline-trained CNN feature extraction network is called to re-extract features and perform matching verification again. If the similarity is <0.7, the fault is temporarily identified as an unknown fault, marked, and uploaded to the cloud.

[0073] (3) Fault Alarm and Result Feedback: After matching, the system generates alarm information based on the fault type, including fault location (based on the coordinates of the target area in the image), fault level (classified according to similarity quantification), and handling suggestions (e.g., severe hot spots require immediate shutdown for maintenance, minor cracks require regular monitoring). The formula for quantifying the fault level is as follows: ; in, The fault level is indicated by the severity level (level 3 is the most severe, level 0 is unknown). To improve the overall similarity, alarm information is displayed in real time on the local device screen and simultaneously pushed to the operation and maintenance management platform; at the same time, the fault identification results (feature vector + fault label) are stored in a local temporary database to provide data support for subsequent model iteration and optimization.

[0074] This invention provides a photovoltaic panel fault detection method. It achieves real-time image acquisition of the photovoltaic panel surface by deploying a high-speed visible light lens, ensuring timely fault monitoring. Based on extracted vector angles and multi-dimensional fusion features of the color space, the method enhances the characterization capability of fault features and the robustness against environmental interference such as light and dirt. A coarse-fine matching hierarchical matching strategy based on a local fault feature library is adopted, balancing detection speed and recognition accuracy to meet the real-time requirements of large-scale photovoltaic array online monitoring. Finally, the method outputs fault type, location coordinates, and level information, and issues alarms accordingly. This realizes an end-to-end automated process from image acquisition to fault location and hierarchical alarm, significantly improving the intelligence level and fault response efficiency of photovoltaic power plant operation and maintenance.

[0075] Optionally, the photovoltaic panel fault detection method provided in this embodiment of the invention further includes steps S201-S206 before step S103.

[0076] S201. Obtain historical images of the photovoltaic panel.

[0077] S202. Based on historical images, determine vector angular features and color space features, and calculate the data quality coefficient and invalid threshold for each historical image.

[0078] S203. Based on the data quality coefficient and invalid threshold of each historical image, select high-quality images with a data quality coefficient greater than the set value.

[0079] S204. Based on the vector angular features and color space features of high-quality images, the DBSCAN algorithm is used for cluster analysis to obtain clustering results.

[0080] S205. Based on the clustering results, a lightweight CNN network with adaptive convolution kernel selection is used to extract core fault features. Fault samples are generated by combining the fault types corresponding to each fault feature and storing them in the local fault feature library.

[0081] S206. Based on multiple fault samples in the local fault feature library, a fault type identification model is constructed using support vector machine, and the model parameters are optimized using artificial bee colony algorithm to form a fault detection model.

[0082] For example, embodiments of the present invention can use offline data sources as input to complete the entire process from data preprocessing to model training and validation. The core objective is to build a stable and reliable fault feature library and model knowledge base to provide core support for online detection. The specific steps are as follows: (1) Enhanced Data Preprocessing: Building upon data cleaning, a data standardization step is added. For the selected high-quality vector angle sequences and color space feature vectors, the Min-Max standardization method is used to eliminate the magnitude differences between different dimensions of the data, ensuring that the features are on the same order of magnitude. This avoids the feature weight imbalance caused by the difference between the vector angle numerical range (0°) and the color space numerical range (0~255). The Min-Max standardization formula is: ; in, These are the standardized eigenvalues. These are the original eigenvalues. and These are the minimum and maximum values ​​of this feature dimension, respectively. , This refers to the standardized target range.

[0083] (2) Iterative optimization of DBSCAN classification parameters: Based on the initially determined range of DBSCAN parameters, the silhouette coefficient is introduced as a classification performance evaluation metric to further optimize the neighborhood radius (Eps) and minimum sample size (MinPts). The formula for calculating the silhouette coefficient is: ; in, For the first The silhouette coefficient of each sample (within the range of [-1, 1], the closer to 1, the better the classification effect). For the first The average distance between a sample and other samples within its own cluster. For the first The average distance between a sample and its nearest other samples within the same cluster. The total number of samples, The average silhouette coefficient (an overall classification performance indicator) for all samples.

[0084] The specific process is as follows: within the parameter range obtained from the initial iteration (Eps: 6~8, MinPts: 6~8), traverse Eps with a step size of 0.5 and traverse MinPts with a step size of 1, calculate the silhouette coefficient corresponding to each set of parameters, and select the parameter combination corresponding to the maximum silhouette coefficient (approaching 1) as the optimal classification parameter.

[0085] (3) Training and optimization of CNN feature extraction model: Based on a defined modular architecture, a lightweight CNN feature extraction network is constructed with the goal of balancing feature extraction accuracy and model complexity. In the network structure, the input layer is a standardized vector angle-color space fusion data matrix (dimension: [dimension not specified]). ).

[0086] Convolutional layer feature calculation: The convolutional layer uses a predetermined optimal number of kernels (3 / 5 / 7). The core formula for feature calculation in convolution operations is: ; in, To output feature maps in Location feature value For the input feature map, The kernel size (e.g.) (corresponding to a 3×3 convolution kernel) The convolution stride (taken here) ), For the convolution kernel in The weighting parameters of the location, This is the bias term for the convolutional layer.

[0087] The formula for calculating the size of the output feature map after convolution is: ; in For the height / width of the input feature map, To fill in the number, take To maintain consistent feature map size.

[0088] Activation and Pooling: The ReLU activation function is chosen to solve the gradient vanishing problem. The ReLU formula is: ; The pooling layer uses max pooling with a pooling size of 2×2 and a stride of 2 to achieve feature dimensionality reduction and key information preservation. The fully connected layer maps the pooled features to feature vectors of fixed dimensions (the dimension is optimized to 64 dimensions according to the data complexity).

[0089] Output Layer and Loss Function: The output layer performs preliminary classification of the feature vectors through Softmax activation. The formula for the Softmax activation function is: ; in, For the first Predicted probability of class The first output of the fully connected layer Class score, This represents the total number of fault categories.

[0090] During training, the cross-entropy loss function is used to calculate the error between the predicted value and the true label. The formula for cross-entropy loss is: ; in, This represents the cross-entropy loss value. The number of training samples, For the first The sample corresponds to the first The actual label of the class (one-hot encoded, either 0 or 1). For the first The sample corresponds to the first The predicted probability of a class.

[0091] The initial learning rate was set to 0.001. The confidence level of the predicted value was calculated every 10 iterations until the loss function converged (error less than 0.001). After training, the 64-dimensional feature vector output by the network was used as the core fault features and stored in the local fault feature library.

[0092] (4) Construction of the fault model knowledge base: Using the trained CNN feature vectors as input, a fault identification model based on support vector machine (SVM) is constructed to achieve a nonlinear mapping between feature vectors and fault types.

[0093] The radial basis function (RBF) is used to map the 64-dimensional feature vector to a high-dimensional feature space. The formula for the radial basis function is: ; in, For feature vectors and In the inner product of higher-dimensional space, This is the bandwidth parameter of the RBF kernel (which controls the locality of the kernel function). for and The Euclidean distance. The decision function of SVM is: ; in, For the set of support vectors, For the Lagrange multipliers corresponding to the support vectors, For the true labels of support vectors, The bias term for the classification hyperplane.

[0094] Introducing the artificial bee colony optimization algorithm to regularize the classification hyperplane parameter (penalty coefficient) ) and RBF kernel parameters ( Cross-validation and optimization are performed (optimization range is 0.1~1.0). The fitness function of the artificial bee colony optimization targets the classification accuracy of the SVM with 5-fold cross-validation, and the formula is: ; in, For the first In cross-validation, the corresponding parameters SVM classification accuracy (value range [0, 1]). A higher value indicates better performance of the parameter combination.

[0095] Finally, the optimized SVM model parameters, CNN feature extraction network parameters, and fault feature library are integrated to form a locally callable fault model knowledge base. At the same time, a lightweight model file (ppb format) is exported to prepare for online deployment on the device.

[0096] Thus, this invention, through a systematic offline modeling process, introduces data quality coefficients and invalid threshold screening during the historical image data cleaning stage, effectively improving the purity and representativeness of the training data; it employs DBSCAN unsupervised clustering combined with contour coefficient optimization to enhance the separability of fault categories; furthermore, it extracts robust features through adaptive convolutional kernel CNN and optimizes SVM model parameters using artificial bee colony algorithm, constructing a high-precision fault identification model with strong generalization ability, providing a reliable feature library and discrimination basis for online detection, and significantly improving the detection accuracy and stability of the overall method.

[0097] Optionally, the photovoltaic panel fault detection method provided in this embodiment of the invention further includes steps S301-S304 after step S105.

[0098] S301. If an unknown fault sample with a similarity lower than the set threshold is detected, the unknown fault sample will be manually reviewed, and the manual review result will be obtained.

[0099] In some embodiments, the results of manual review include the fault type of unknown fault samples.

[0100] S302. Based on unknown fault samples and the local fault feature library, generate incremental samples.

[0101] S303. Based on incremental samples, perform incremental training on the fault detection model to obtain an updated fault detection model.

[0102] S304. Based on the updated fault detection model, perform fault detection on the photovoltaic panel.

[0103] Thus, this invention introduces an iterative optimization mechanism based on unknown fault samples. When low-similarity samples appear in online detection, their fault types are confirmed through manual review, and incremental samples are generated. These incremental samples are then used to incrementally train and update the fault detection model. This mechanism enables the system to continuously learn and adaptively evolve, gradually expanding the scope of new fault identification without reconstructing the model. This effectively improves the system's generalization performance and practicality in long-term operation, achieving dynamic enhancement of detection capabilities.

[0104] Optionally, the photovoltaic panel fault detection method provided by the present invention further includes steps S401-S404.

[0105] S401. Based on fault characteristics and fault detection models, determine the model identification results.

[0106] In some embodiments, the model identification results include the fault type, location coordinates, and fault level predicted by the model.

[0107] S402. Based on the model recognition results and fault information, auxiliary verification is performed to obtain the verification results.

[0108] S403. If the verification result is successful, a fault alarm will be issued based on the fault information.

[0109] S404. If the verification result is that the verification fails, a prompt message is generated to inform the user that the fault detection results are inconsistent.

[0110] Thus, this invention adds a dual-path verification step involving both model recognition results and feature matching results. By comprehensively comparing the fault information generated through matching with the recognition results directly predicted by the model, an auxiliary decision-making mechanism is formed. This design can effectively reduce false alarms or missed alarms that may be caused by a single judgment path, improve the reliability and confidence of fault identification results, and promptly prompt users to intervene and verify when the two are inconsistent, thereby enhancing the robustness and reliability of the system in actual deployment.

[0111] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0112] Figure 2 A schematic diagram of a photovoltaic panel fault detection device provided in an embodiment of the present invention is shown. The detection device 500 includes a communication module 501 and a processing module 502.

[0113] The communication module 501 is used to acquire real-time images of the photovoltaic panel surface through a high-speed visible light lens deployed near the photovoltaic panel; The processing module 502 is used to extract features based on the surface image of the photovoltaic panel to obtain the fault features of the photovoltaic panel, including vector angle features and color space features; based on the fault features and the local fault feature library, a hierarchical matching strategy of coarse matching and fine matching is adopted to obtain the matching results; based on the matching results, the fault information of the photovoltaic panel is determined, including the fault type, location coordinates and fault level; based on the fault information, a fault alarm is generated.

[0114] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 600 includes: a processor 601, a memory 602, and a computer program 603 stored in the memory 602 and executable on the processor 601. When the processor 601 executes the computer program 603, it implements the steps in the above-described method embodiments. Alternatively, when the processor 601 executes the computer program 603, it implements the functions of each module / unit in the above-described device embodiments.

[0115] For example, the computer program 603 may be divided into one or more modules / units, which are stored in the memory 602 and executed by the processor 601 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 603 in the electronic device 600.

[0116] The processor 601 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0117] The memory 602 can be an internal storage unit of the electronic device 600, such as a hard disk or memory of the electronic device 600. The memory 602 can also be an external storage device of the electronic device 600, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD) card, flash card, etc., equipped on the electronic device 600. Furthermore, the memory 602 can include both internal and external storage units of the electronic device 600. The memory 602 is used to store the computer program and other programs and data required by the terminal. The memory 602 can also be used to temporarily store data that has been output or will be output.

[0118] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for detecting faults in photovoltaic panels, characterized in that, include: Real-time images of the photovoltaic panel's surface are captured using high-speed visible light lenses deployed near the panel. Based on the surface image of the photovoltaic panel, feature extraction is performed to obtain the fault features of the photovoltaic panel, which include vector angle features and color space features; Based on the aforementioned fault characteristics and the local fault characteristic library, a hierarchical matching strategy of coarse matching and fine matching is adopted to obtain the matching results. Based on the matching results, the fault information of the photovoltaic panel is determined, including the fault type, location coordinates, and fault level. Based on the fault information, a fault alarm is issued.

2. The photovoltaic panel fault detection method according to claim 1, characterized in that, Before obtaining the matching result using a hierarchical matching strategy of coarse and fine matching based on the fault characteristics and a preset fault detection model, the process further includes: Acquire historical images of photovoltaic panels; Based on the historical images, determine the vector angular features and color space features, and calculate the data quality coefficient and invalidity threshold for each historical image; Based on the data quality coefficient and invalidity threshold of each historical image, high-quality images with a data quality coefficient greater than a set value are selected. Based on the vector angular features and color space features of the high-quality images, the DBSCAN algorithm is used for cluster analysis to obtain the clustering results. Based on the clustering results, a lightweight CNN network with adaptive convolution kernel selection is used to extract core fault features. Fault samples are generated by combining the fault types corresponding to each fault feature and storing them in a local fault feature library. Based on multiple fault samples in the local fault feature library, a fault type identification model is constructed using support vector machine, and the model parameters are optimized using artificial bee colony algorithm to form a fault detection model.

3. The photovoltaic panel fault detection method according to claim 1, characterized in that, After issuing an alarm based on fault type, location, and level, the following is also included: If an unknown fault sample with a similarity lower than a set threshold is detected, the unknown fault sample is manually reviewed, and the manual review result is obtained; the manual review result includes the fault type of the unknown fault sample. Based on the unknown fault samples and the local fault feature library, incremental samples are generated. Based on the incremental samples, the fault detection model is incrementally trained to obtain an updated fault detection model. Based on the updated fault detection model, fault detection is performed on photovoltaic panels.

4. The photovoltaic panel fault detection method according to claim 1, characterized in that, The step of extracting features from the surface image of the photovoltaic panel to obtain the fault features of the photovoltaic panel includes: The surface image is divided into regions to obtain the color space vector of each pixel in each region; Calculate the vector angle between the color space vector of each pixel in each region and the white standard color vector to obtain the vector angle feature; The color space vectors of each region are standardized and converted into color space features with uniform dimensions. The fault features are obtained by weighted fusion of vector angular features and color space features.

5. The photovoltaic panel fault detection method according to claim 1, characterized in that, Based on the fault characteristics and a preset fault detection model, a hierarchical matching strategy of coarse matching and fine matching is adopted to obtain the matching results, including: Based on the aforementioned fault features and the local fault feature library, coarse matching is performed through similarity mapping to obtain candidate fault features with a coarse matching similarity greater than a first threshold. Based on the candidate fault features, an improved cosine similarity method is used for fine matching, which integrates the consistency of vector angle direction and the similarity of color space distribution to obtain the fine matching similarity. If the similarity of the fine match is greater than the second threshold, the match is considered successful. If the fine matching similarity is less than or equal to the second threshold, or if there are no candidate fault features, then the fault features are marked as unknown fault samples. Output the matching results, which include the candidate fault feature with the highest fine-match similarity; the fault type corresponding to the candidate fault feature; and the fine-match similarity.

6. The photovoltaic panel fault detection method according to claim 1, characterized in that, The process of determining the fault information of the photovoltaic panel based on the matching results includes: Based on the successfully matched fault types, determine the fault type of the photovoltaic panel; Based on the location coordinates of the fault features in the surface image, calculate the location coordinates of the fault on the photovoltaic panel; The fault level of the photovoltaic panel is determined based on the precision matching similarity. The fault type, location coordinates, and fault level are integrated into fault information.

7. The photovoltaic panel fault detection method according to claim 1, characterized in that, The step of issuing a fault alarm based on the fault information includes: Generate alarm messages based on fault type and fault level; The alarm information is displayed on the device's screen in real time and pushed to the cloud-based operation and maintenance management platform; The fault characteristics and fault information detected in this test will be stored in the local fault characteristic database.

8. The photovoltaic panel fault detection method according to claim 1, characterized in that, The method further includes: Based on the fault characteristics and fault detection model, the model identification result is determined, which includes the fault type, location coordinates and fault level predicted by the model. Based on the model identification results and the fault information, auxiliary verification is performed to obtain the verification results; If the verification result is successful, a fault alarm will be issued based on the fault information. If the verification result is that the verification fails, a prompt message is generated to inform the user that the fault detection results are inconsistent.

9. A photovoltaic panel fault detection device, characterized in that, include: The communication module is used to acquire real-time images of the photovoltaic panel surface via a high-speed visible light lens deployed near the photovoltaic panel; The processing module is used to extract features based on the surface image of the photovoltaic panel to obtain the fault features of the photovoltaic panel, the fault features including vector angle features and color space features; Based on the aforementioned fault characteristics and the local fault characteristic library, a hierarchical matching strategy of coarse matching and fine matching is adopted to obtain the matching results. Based on the matching results, the fault information of the photovoltaic panel is determined, including the fault type, location coordinates, and fault level. Based on the fault information, a fault alarm is issued.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor being configured to invoke and run the computer program stored in the memory to perform the method as described in any one of claims 1 to 8.

Images