Photovoltaic power station disconnecting link position identification method based on significance analysis and deep learning technology

Abstract

The invention discloses a photovoltaic power station disconnecting link position identification method based on significance analysis and a deep learning technology, and belongs to the technical field of photovoltaic power station digital analysis. According to the method, historical and current images of the disconnecting link of the photovoltaic power station are acquired through a camera, a saliency region containing the disconnecting link is analyzed and segmented by using a saliency feature extraction algorithm, and then the position of the disconnecting link in the historical images is trained by using a deep learning algorithm, such as YOLO-V5. And finally, quickly identifying the position of the disconnecting link in the current image by using the trained model. According to the method, the rapid target area positioning capability of the bionics of the visual saliency area is utilized to assist the deep learning algorithm to rapidly complete the identification work of the position of the disconnecting link of the photovoltaic power station, so that the intelligent identification work of the position of the disconnecting link of the photovoltaic power station is more effective and reliable; the accuracy and efficiency of photovoltaic power station disconnecting link position identification are effectively improved, safe production and operation and maintenance of a photovoltaic power station are ensured, and the method has important practical application value and popularization prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power station information processing and intelligent monitoring, and in particular to a photovoltaic power station switch position recognition method based on significance analysis and deep learning technology. Background Art

[0002] In recent years, with the continuous growth of global energy demand and the improvement of environmental protection awareness, new energy power generation projects have been rapidly developed around the world, especially photovoltaic power generation technology, which has attracted much attention due to its clean and renewable characteristics. In China, photovoltaic power generation has ushered in unprecedented development opportunities, and the number of photovoltaic power stations has increased rapidly, becoming an important force in optimizing the energy structure.

[0003] However, with the rapid increase in the number of photovoltaic power stations, safety issues in their operation process have become increasingly prominent. As a key power equipment, the position of the knife switch of the photovoltaic power station is directly related to the stable operation and safe production of the entire photovoltaic power generation system. Specifically, whether the opening and closing position of the knife switch is correct and in place directly affects the on-off state of the circuit, which is of great significance for preventing safety accidents such as short circuits and overloads.

[0004] In the traditional operation and maintenance of photovoltaic power stations, the identification of the knife switch position mainly relies on manual inspection. However, this method has many disadvantages: first, the efficiency of manual inspection is low, and it is difficult to achieve real-time monitoring of a large number of knife switch devices; second, due to the subtle differences in the visual characteristics of the knife switch position, especially when the switch is in place or not in place, manual identification is prone to errors, resulting in misjudgment or missed judgment; in addition, manual inspection is also restricted by various factors such as weather and environment, making it difficult to carry out work effectively under harsh conditions.

[0005] In order to overcome the shortcomings of traditional manual inspection methods and improve the accuracy and efficiency of photovoltaic power station switch position identification, the industry has begun to explore the use of new technologies and methods for intelligent identification. Among them, image recognition technology and deep learning algorithms have attracted much attention due to their powerful data processing and pattern recognition capabilities.

[0006] Specifically, by collecting image data of the photovoltaic power station switch through the camera and extracting key information from the image using image saliency region analysis technology, the initial location of the switch can be achieved. Subsequently, a high-precision recognition model can be constructed by learning and training a large amount of historical image data using a deep learning algorithm. Finally, the model is applied to the real-time collected image data to achieve fast and accurate recognition of the photovoltaic power station switch position.

[0007] Although some studies have attempted to apply image recognition technology and deep learning algorithms to the field of photovoltaic power station switch position recognition, the existing methods still have some shortcomings. For example, some methods are not effective in dealing with complex backgrounds and target occlusion; some methods have high computational complexity when dealing with large-scale data sets and are difficult to meet real-time requirements. Therefore, developing a more efficient and accurate method for rapid recognition of photovoltaic power station switch positions has important practical significance and application value.

[0008] In summary, the present invention aims at the deficiencies in the prior art and proposes a method for identifying the location of the switch in a photovoltaic power station based on saliency analysis and deep learning technology. This method realizes the rapid and accurate identification of the location of the switch in a photovoltaic power station by comprehensively utilizing the advantages of image saliency region analysis and deep learning algorithms, providing strong technical support for the safe production of photovoltaic power stations. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide a method for identifying the position of knife switches in photovoltaic power stations based on significance analysis and deep learning technology, so as to solve the problem of difficulty in identifying the position of knife switches in photovoltaic power stations, especially in the context of the increasing number of photovoltaic power generation systems. The traditional manual detection method is prone to misjudgment due to the similar appearance of knife switches, which in turn causes safety accidents.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for identifying the position of a photovoltaic power station switch based on significance analysis and deep learning technology, comprising the following steps: Step 1: Use a camera to obtain historical images and current images of the opening and closing positions of the photovoltaic power station switch; Step 2: Use the significant feature extraction algorithm to analyze the significant areas containing knife gates in the historical image and the current image obtained in Step 1 and segment them to obtain the target area; Step 3: Use the deep learning algorithm to learn and train the knife switch location area in the historical image processed in Step 2; Step 4: Use the trained deep learning algorithm model to quickly identify the location of the photovoltaic power station switch in the current image.

[0011] In the preferred solution, in Step 1, a camera is used to obtain historical images and current images of the opening and closing positions of the photovoltaic power station knife switch. The process is to fix the camera at a position where it can capture a panoramic view of the photovoltaic power station knife switch and collect images of the photovoltaic power station knife switch in real time.

[0012] In a preferred solution, the specific steps of Step 2 include: Step 2.1: Extract the RGB (R red, G green, B blue) model information of the image and use the central pixel of the image as the knife gate saliency seed node; Step 2.2: Perform convex hull operation on the knife gate significant seed nodes to expand the significant area, calculate the critical value of the significant seed nodes, and select the nodes that meet the conditions as the knife gate significant diffusion seed nodes; Step 2.3: Propagate and diffuse the significant feature information of the above seed node area, and use the algorithm of local feature information reconstruction to complete the complete detection of the knife switch area.

[0013] In a preferred solution, the significant feature extraction algorithm in Step 2 further includes: Extract the image RGB model information and calculate the Euclidean distance between nodes in the image to determine the critical value of the significant seed node; By setting the saliency diffusion seed node feature vector and using the adjacency matrix and undirected graph to propagate feature information, the knife switch area detection is finally completed using the algorithm of local feature information reconstruction.

[0014] In a preferred solution, the specific process of Step 2.2 includes: Step 2.2.1: Calculate and define the convex hull area containing the Harris node. At this time, it is assumed that there is nodes in the convex hull area and The RGB model attribute vectors in the two regions are respectively and Two sets are performed; then Indicates from a The distance to the RGB model attribute vector of other nodes in the image object is solved as follows: (1) in, for arrive The Euclidean gap; Step 2.2.2: Calculate the critical value of the knife gate significance seed node of the image through this distance : (2) According to the critical value, the area defined by the nodes in the defined convex hull is The nodes that meet the conditions are selected as the knife gate significance diffusion seed nodes.

[0015] In the preferred solution, the specific process of Step 2.3 includes: Step 2.3.1: Set the knife gate significant diffusion seed node feature vector: (3) in is the total number of the above seed nodes (number of pixels); Depend on Derived a feature undirected graph , using the adjacency matrix To verify the undirected graph ,use Used to represent undirected graphs Midpoint Distance and The weight of is the adjacency matrix An element in , so the maximum weight Means node and Due to similar significant features, communication can be carried out.

[0016] Step 2.3.2: Use the algorithm for reconstructing local feature information to propagate the information of significant features to diffuse the final fault area. The process is as follows: The formula for propagating significant features using the algorithm reconstructed using local feature information is as follows: satisfy (4) in is the saliency diffusion seed node described in Step 2.3.1, , and for and exist The node element of the position, In the original image, The calculated Laplace operator is is the adjacency matrix of the current image, is a diagonal matrix whose diagonal elements are .

[0017] Step 2.3.3: Complete the rapid detection of the knife switch based on the propagation results.

[0018] In a preferred solution, the specific steps of Step 3 include: Step 3.1: Label the historical images into the positions of knife switch closing, opening and not closing; Step 3.2: Divide the annotated historical images into training images, test images, and verification images in a ratio of 8:1:1; Step 3.3: Use the YOLO-V5 (You Only Look Once version 5) algorithm for training, set parameters including num_classes, img_size, batch_size, momentum, learning rate, box, cls and obj, ensure that the training model converges, and then select the optimal model as the formal recognition model.

[0019] In the preferred solution, the specific parameter settings of Step 3.3 include: num_classes is the number of classes of objects corresponding to the training set, and its general value is 3, which corresponds to three types of switch positions; img_size is the size of the images in the training set, which is set according to needs, and can be 640; batch_size is a larger value within the memory range, and its general value is 8; momentum is the learning rate momentum in deep learning, and its initial value is set according to the situation of the data set; learning rate is the learning rate in deep learning, which is understood as taking empirical values, and its initial value is set according to the situation of the data set; box is the coefficient of GIou loss, which is set to 2; cls is the coefficient of classification loss, and obj is the coefficient of object loss. The initial values ​​of the two are also set according to the situation of the data set.

[0020] In the preferred scheme, the specific method of ensuring that the training model converges in Step 3.3 and selecting the optimal model as the formal recognition model is: through the loss situation during training, ensure that the training model has reached convergence, and then select the model with the lowest loss value as the test model, perform test feedback through the test data set, debug the parameters, and finally select the parameter model with the best result as the formal recognition model.

[0021] The photovoltaic power station switch position identification method based on significance analysis and deep learning technology provided by the present invention has the following beneficial effects: 1. The present invention solves the problem that it is difficult to identify the position of the knife switch in the photovoltaic power station, and in the context of the increasing number of photovoltaic power generation systems, the traditional manual detection method is prone to misjudgment due to the similar appearance of the knife switches, thereby causing safety accidents; 2. The present invention realizes accurate and rapid identification of the location of the photovoltaic power station switch through intelligent image processing and machine learning technology; 3. The present invention uses a camera to capture images of all states of the switch position of a photovoltaic power station, and then uses an image saliency analysis method to obtain the accurate target area of ​​the switch, and then uses a deep learning algorithm to learn and evaluate the existing image data, and finally forms an effective data learning model, and then quickly identifies the photovoltaic power switch position for the current target area extracted by the image saliency analysis method; 4. The present invention uses a camera fixed at a position where a panoramic view of the photovoltaic power station knife switch can be captured through image acquisition and preprocessing, and real-time acquisition of historical images and current images of the opening and closing positions of the knife switch. This step ensures that the data source for subsequent analysis is accurate and real-time; 5. The significant region analysis and target segmentation of the present invention uses a significant feature extraction algorithm to process the collected image, and extracts the significant region containing the knife gate by analyzing the RGB model (R red, G green, B blue) information in the image. Further, the complete knife gate region is accurately segmented through convex hull calculation and significant feature information propagation and diffusion algorithm. This process significantly improves the efficiency and accuracy of subsequent recognition work; 6. The present invention uses deep learning model training to learn and train the knife switch position area in the historical image using a deep learning algorithm, such as YOLO-V5. By marking and classifying the image, the closed position, open position, and incomplete closed position, a training data set is constructed, and the algorithm parameters are optimized to achieve the best training effect. This step establishes an accurate knife switch position recognition model, which provides a reliable foundation for subsequent real-time recognition. 7. The present invention uses a trained deep learning algorithm model to quickly identify the location of the photovoltaic power station switch in the current image through real-time recognition and feedback, and feeds back the recognition result in real time. This step realizes the intelligent and automated monitoring of the location of the photovoltaic power station switch, and significantly improves the level of safety production management; 8. The present invention proposes an innovative and effective method for quickly identifying the position of knife switches in photovoltaic power stations by combining image salient area analysis and deep learning algorithms. This method not only overcomes the limitations of traditional manual detection methods, but also significantly improves the recognition efficiency and accuracy, providing a strong guarantee for the safe production of photovoltaic power stations.

[0022] 9. The present invention has been put into practice and has achieved good recognition results. It has solved the problem of identifying the position status of the switch in the photovoltaic power station, greatly improved the efficiency and accuracy of operation and maintenance, and reduced the risk of human misjudgment. At the same time, its intelligent and automated characteristics have also laid a solid foundation for the intelligent upgrade of photovoltaic power stations. 10. The present invention accurately identifies the position status of the knife switch, which helps to timely discover potential safety hazards and ensure the safe production of photovoltaic power stations; 11. The method of the present invention also has a strong generalization ability and can adapt to the identification of knife switches under different lighting conditions, angles and occlusion conditions, providing a solid technical support for the intelligent operation and maintenance of photovoltaic power stations; 12. The present invention has good scalability and further integrates a real-time monitoring and early warning system, which can immediately notify the operation and maintenance personnel when an abnormal switch position is found, realizing full-process automated management from identification to response, and further promoting the intelligent process of photovoltaic power station operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will be further described below in conjunction with the accompanying drawings and implementation examples: Figure 1 The figure is a flow chart of the rapid identification method of the present invention. DETAILED DESCRIPTION

[0024] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments: Example 1 like Figure 1 As shown, the photovoltaic power station switch position recognition method based on significance analysis and deep learning technology includes the following steps: Step 1: Use a camera to obtain historical images and current images of the opening and closing positions of the photovoltaic power station switch; Step 2: Use the significant feature extraction algorithm to analyze the significant areas containing knife gates in the historical image and the current image obtained in Step 1 and segment them to obtain the target area; Step 3: Use the deep learning algorithm to learn and train the knife switch location area in the historical image processed in Step 2; Step 4: Use the trained deep learning algorithm model to quickly identify the location of the photovoltaic power station switch in the current image.

[0025] In this embodiment, in Step 1, a camera is used to obtain historical images and current images of the opening and closing positions of the photovoltaic power station knife switch. The process is to fix the camera at a position where it can capture a panoramic view of the photovoltaic power station knife switch and collect images of the photovoltaic power station knife switch in real time.

[0026] Furthermore, the specific steps of Step 2 include: Step 2.1: Extract the RGB model information of the image and use the central pixel of the image as the gate saliency seed node; Step 2.2: Perform convex hull operation on the knife gate significant seed nodes to expand the significant area, calculate the critical value of the significant seed nodes, and select the nodes that meet the conditions as the knife gate significant diffusion seed nodes; Step 2.3: Propagate and diffuse the significant feature information of the above seed node area, and use the algorithm of local feature information reconstruction to complete the complete detection of the knife switch area.

[0027] Furthermore, the significant feature extraction algorithm in Step 2 further includes: Extract the image RGB model information and calculate the Euclidean distance between nodes in the image to determine the critical value of the significant seed node; By setting the saliency diffusion seed node feature vector and using the adjacency matrix and undirected graph to propagate feature information, the knife switch area detection is finally completed using the algorithm of local feature information reconstruction.

[0028] Furthermore, the specific process of Step 2.2 includes: Step 2.2.1: Calculate and define the convex hull area containing the Harris node. At this time, assume that there is nodes in the convex hull area and The RGB model attribute vectors in the two regions are respectively and Two sets are performed; then Indicates from a The distance to the RGB model attribute vector of other nodes in the image object is solved as follows: (1) in, for arrive The Euclidean gap; Step 2.2.2: Calculate the critical value of the knife gate significance seed node of the image through this distance : (2) According to the critical value, the area defined by the nodes in the defined convex hull is The nodes that meet the conditions are selected as the knife gate significance diffusion seed nodes.

[0029] Furthermore, the specific process of Step 2.3 includes: Step 2.3.1: Set the knife gate significant diffusion seed node feature vector: (3) in is the total number of the above seed nodes (number of pixels); Depend on Derived a feature undirected graph , using the adjacency matrix To verify the undirected graph ,use Used to represent undirected graphs Midpoint Distance and The weight of is the adjacency matrix An element in , so the maximum weight Means node and Due to similar significant features, communication can be carried out.

[0030] Step 2.3.2: Use the algorithm for reconstructing local feature information to propagate the information of significant features to diffuse the final fault area. The process is as follows: The formula for propagating significant features using the algorithm reconstructed using local feature information is as follows: satisfy (4) in is the saliency diffusion seed node described in Step 2.3.1, , and for and exist The node element of the position, In the original image, The calculated Laplace operator is is the adjacency matrix of the current image, is a diagonal matrix whose diagonal elements are .

[0031] Step 2.3.3: Complete the rapid detection of the knife switch based on the propagation results.

[0032] Furthermore, the specific steps of Step 3 include: Step 3.1: Label the historical images into the positions of knife switch closing, opening and not closing; Step 3.2: Divide the annotated historical images into training images, test images, and verification images in a ratio of 8:1:1; Step 3.3: Use the YOLO-V5 algorithm for training, set parameters including num_classes, img_size, batch_size, momentum, learning rate, box, cls and obj, ensure that the training model converges, and then select the optimal model as the formal recognition model.

[0033] Furthermore, the specific parameter settings of Step 3.3 include: num_classes is the number of classes of objects corresponding to the training set, and its general value is 3, which corresponds to three types of switch positions; img_size is the size of the images in the training set, which is set according to needs, and can be 640; batch_size is a larger value within the memory range, and its general value is 8; momentum is the learning rate momentum in deep learning, and its initial value is set according to the situation of the data set; learning rate is the learning rate in deep learning, which is understood as taking empirical values, and its initial value is set according to the situation of the data set; box is the coefficient of GIou loss, which is set to 2; cls is the coefficient of classification loss, and obj is the coefficient of object loss. The initial values ​​of the two are also set according to the situation of the data set.

[0034] Furthermore, the specific method for ensuring that the training model converges in Step 3.3 and selecting the optimal model as the formal recognition model is: through the loss situation during training, ensure that the training model has reached convergence, and then select the model with the lowest loss value as the test model, perform test feedback through the test data set, debug the parameters, and finally select the parameter model with the best result as the formal recognition model.

[0035] Example 2 In another preferred embodiment, based on the above embodiment 1, Figure 1 As shown, through a specific example of a photovoltaic power plant, the implementation method of the photovoltaic power station switch position identification method based on significance analysis and deep learning technology proposed by the present invention is described in detail.

[0036] Step 1: Image acquisition First, fix the camera in a suitable position to capture the panoramic view of the photovoltaic power station knife switch, ensuring that the camera can capture the opening and closing position image of the knife switch in real time and clearly; in this step, the camera continuously captures and saves the historical images and current images of the photovoltaic power station knife switch, providing a data basis for subsequent processing.

[0037] Step 2: Salient Region Analysis and Segmentation After acquiring the image, the image is processed using a significant feature extraction algorithm to analyze and segment the significant area containing the knife gate; the specific processing process is as follows: Extract salient seed nodes: Extract the central pixel points from the image according to the RGB model (red: R, green: G, blue: B) as the salient seed nodes of the knife gate; this step is based on the assumption that the central area of ​​the image often contains important information, which effectively narrows the scope of subsequent processing.

[0038] Convex hull operation: Perform convex hull operation on the significant seed nodes to further expand the feature information of the target area. In this process, the convex hull area containing the Harris node is calculated, and the RGB model attribute vector distance between the nodes is calculated. The solution formula is as follows: (1) in, for arrive The Euclidean gap; Determine the critical value of significant seed nodes : (2) Thus, nodes that meet the conditions are selected as saliency diffusion seed nodes.

[0039] Propagation and diffusion of salient feature information: Setting the feature vector of the salient diffusion seed node, (3) in is the total number of the above seed nodes (number of pixels); And construct a feature undirected graph; use the adjacency matrix to verify the undirected graph, and use the algorithm of local feature information reconstruction to propagate and diffuse the information of significant features, and finally obtain the complete knife switch area.

[0040] The formula for propagating significant features using the algorithm reconstructed using local feature information is as follows: satisfy (4) in is the saliency diffusion seed node described in Step 2.3.1, , and for and exist The node element of the position, In the original image, The calculated Laplace operator is is the adjacency matrix of the current image, is a diagonal matrix whose diagonal elements are .

[0041] Step 3: Deep learning algorithm training Use deep learning algorithms to learn and train the knife switch location area in historical images. Taking the YOLO-V5 algorithm as an example, the specific process is as follows: Data labeling and division: The historical images that have been analyzed and processed for significant regions are labeled and divided into the positions of the knife switch closed, open, and incompletely closed. The images are then divided into training images, test images, and verification images in a ratio of 8:1:1.

[0042] Parameter setting and training: According to the actual situation of the data set, set various parameters of the YOLO-V5 algorithm, including num_classes (number of training set object categories, 3 in this case), img_size (image size), batch_size (batch size), momentum (learning rate momentum), learning rate (learning rate), box (GIou loss coefficient), cls (classification loss coefficient) and obj (object loss coefficient), etc. Then start data training until the model converges.

[0043] Model selection and optimization: Select the model with the lowest loss value during training as the test model, conduct test feedback through the test data set, debug the parameters, and finally select the parameter model with the best results as the formal recognition model.

[0044] Step 4: Quick Identification Use the trained deep learning algorithm model to quickly identify the location of the photovoltaic power station switch in the current image. Input the current image into the model, the model automatically completes the detection and identification of the switch location, and outputs the identification result.

[0045] During the implementation process, by collecting images of various states of a knife switch in a photovoltaic power plant, and then performing image saliency analysis to obtain the target area, and then using it as training data for deep learning model training, the image data of the knife switch taken in real time in the area on a certain day in June 2024 is used as training data and as a test data source. Through the above 4 steps, after reverse parameter setting, the recognition model of the photovoltaic power generation knife switch state is obtained, and then tested in the test data source, and the results show that a higher recognition result is obtained; the present invention can efficiently and accurately identify the position state of the photovoltaic power station knife switch, providing a strong guarantee for the safe production of photovoltaic power stations.

[0046] In the preferred scheme, in Step 1, a camera is used to obtain historical images and current images of the opening and closing positions of the photovoltaic power station knife switch. The process is to fix the camera at a position where it can capture a panoramic view of the photovoltaic power station knife switch, and collect images of the photovoltaic power station knife switch in real time. The above settings ensure the comprehensiveness and accuracy of the image data. Subsequently, through image recognition technology, the historical and current images are automatically compared to accurately determine the opening and closing status of the knife switch, thereby providing timely and reliable monitoring information for power station operation and maintenance.

[0047] In the preferred scheme, the specific method of ensuring that the training model converges and selecting the optimal model as the formal recognition model in Step 3.3 is: through the loss situation during training, ensure that the training model has reached convergence, and then select the model with the lowest loss value as the test model, test feedback through the test data set, debug the parameters, and finally select the parameter model with the best result as the formal recognition model; the above settings can ensure that the selected model has the best performance in the recognition task, while reducing the risk of overfitting and improving the generalization ability of the model; this process ensures the scientificity and effectiveness of model selection.

[0048] In summary, the present invention provides a method for identifying the position of knife switches in photovoltaic power stations based on saliency analysis and deep learning technology, which solves the key problems in the identification of knife switches in photovoltaic power stations, realizes efficient and accurate automatic identification, and provides a strong guarantee for the safe production of photovoltaic power stations; the present invention innovatively combines the image saliency region analysis with the deep learning algorithm, and effectively improves the training efficiency and recognition accuracy of the deep learning model by quickly locating the saliency region of the knife switch; it collects images in real time through a fixed camera, realizes the automation and real-time identification of the position of the knife switch in the photovoltaic power station, and improves the operation and maintenance efficiency; the present invention utilizes the rapid positioning ability of the visual saliency region, combines the principle of bionics, and introduces a saliency feature extraction algorithm for the first time in the identification of the knife switch position in the photovoltaic power station, realizes the accurate segmentation of the knife switch area in the image, and improves the recognition efficiency of the target area; the present invention combines the deep learning algorithm with the image saliency region analysis, the method not only utilizes the powerful ability of deep learning in image recognition, but also quickly locates the target area through saliency analysis, reduces the search range of the deep learning algorithm, and improves The recognition speed and accuracy are improved; the present invention realizes the intelligent and real-time position recognition of the photovoltaic power station knife switch by real-time acquisition of the image of the photovoltaic power station knife switch and uses the method for rapid recognition, and provides strong technical support for the safe production and operation and maintenance management of the photovoltaic power station; the present invention has significant innovation in the technical route, breaking through the limitations of traditional recognition methods in speed and accuracy. In the process of significant feature extraction, algorithms such as the center method and convex hull operation are adopted. By accurately calculating the distance and critical value of the RGB model attribute vector, the accurate segmentation of the knife switch significant area is realized; at the same time, in the selection of deep learning algorithm, taking the YOLO-V5 algorithm as an example, the recognition accuracy and stability are further improved through fine parameter setting and model training. The method shows a high recognition rate and real-time performance in practical applications, and can effectively solve the misjudgment problem in the position recognition of the photovoltaic power station knife switch, and provides a reliable technical guarantee for the safe production and operation and maintenance management of the photovoltaic power station; at the same time, the method also has broad application prospects and can be extended to the state monitoring and recognition fields of other similar equipment.

Claims

1. A photovoltaic power station switch position identification method based on saliency analysis and deep learning technology, characterized in that: The method comprises the following steps: Step 1: Use a camera to obtain historical images and current images of the opening and closing positions of the photovoltaic power station switch; Step 2: Use the significant feature extraction algorithm to analyze the significant areas containing knife gates in the historical image and the current image obtained in Step 1 and segment them to obtain the target area; Step 3: Use the deep learning algorithm to learn and train the knife switch location area in the historical image processed in Step 2; Step 4: Use the trained deep learning algorithm model to quickly identify the location of the photovoltaic power station switch in the current image.

2. The photovoltaic power station switch position identification method based on significance analysis and deep learning technology according to claim 1 is characterized by: In Step 1, a camera is used to obtain historical images and current images of the opening and closing positions of the photovoltaic power station knife switch. The process is to fix the camera at a position where the panoramic view of the photovoltaic power station knife switch can be captured, and collect images of the photovoltaic power station knife switch in real time.

3. The photovoltaic power station switch position identification method based on significance analysis and deep learning technology according to claim 1 is characterized in that: The specific steps of Step 2 include: Step 2.1: Extract the RGB model information of the image and use the central pixel of the image as the gate saliency seed node; Step 2.2: Perform convex hull operation on the knife gate significant seed nodes to expand the significant area, calculate the critical value of the significant seed nodes, and select the nodes that meet the conditions as the knife gate significant diffusion seed nodes; Step 2.3: Propagate and diffuse the significant feature information of the above seed node area, and use the algorithm of local feature information reconstruction to complete the complete detection of the knife switch area.

4. According to claim 3, a method for identifying the location of a switch in a photovoltaic power station based on significance analysis and deep learning technology is characterized in that: The significant feature extraction algorithm in Step 2 further includes: Extract the image RGB model information and calculate the Euclidean distance between nodes in the image to determine the critical value of the significant seed node; By setting the saliency diffusion seed node feature vector and using the adjacency matrix and undirected graph to propagate feature information, the knife switch area detection is finally completed using the algorithm of local feature information reconstruction.

5. A method for identifying the location of a switch in a photovoltaic power station based on significance analysis and deep learning technology according to claim 4, characterized in that: The specific process of Step 2.2 includes: Step 2.2.1: Calculate and define the convex hull area containing the Harris node. At this time, it is assumed that there is nodes in the convex hull area and The RGB model attribute vectors in the two regions are respectively and Two sets are performed; then Indicates from a The distance to the RGB model attribute vector of other nodes in the image object is solved as follows: （1） in, for arrive The Euclidean gap; Step 2.2.2: Calculate the critical value of the knife gate significance seed node of the image through this distance : （2） According to the critical value, the area defined by the nodes in the defined convex hull is The nodes that meet the conditions are selected as the knife gate significance diffusion seed nodes.

6. A method for identifying the location of a switch in a photovoltaic power station based on significance analysis and deep learning technology according to claim 5, characterized in that: The specific process of Step 2.3 includes: Step 2.3.1: Set the knife gate significant diffusion seed node feature vector: （3） in is the total number of the above seed nodes (number of pixels); Depend on Derived a feature undirected graph , using the adjacency matrix To verify the undirected graph ,use Used to represent undirected graphs Midpoint Distance and The weight of is the adjacency matrix An element in , so the maximum weight Means node and Due to similar significant features, communication can be carried out. Step 2.3.2: Use the algorithm for reconstructing local feature information to propagate the information of significant features to diffuse the final fault area. The process is as follows: The formula for propagating significant features using the algorithm reconstructed using local feature information is as follows: satisfy (4) in is the saliency diffusion seed node described in Step 2.3.1, , and for and exist The node element of the position, In the original image, The calculated Laplace operator is is the adjacency matrix of the current image, is a diagonal matrix whose diagonal elements are . Step 2.3.3: Complete the rapid detection of the knife switch based on the propagation results.

7. The photovoltaic power station switch position identification method based on significance analysis and deep learning technology according to claim 1 is characterized in that: The specific steps of Step 3 include: Step 3.1: Label the historical images into the positions of knife switch closing, opening and not closing; Step 3.2: Divide the annotated historical images into training images, test images, and verification images in a ratio of 8:1:1; Step 3.3: Use the YOLO-V5 algorithm for training, set parameters including num_classes, img_size, batch_size, momentum, learning rate, box, cls and obj, ensure that the training model converges, and then select the optimal model as the formal recognition model.

8. The photovoltaic power station switch position identification method based on significance analysis and deep learning technology according to claim 7 is characterized in that: The specific parameter settings of Step 3.3 include: num_classes is the number of classes of objects corresponding to the training set, and its general value is 3, which corresponds to three types of switch positions; img_size is the size of the images in the training set, which is set according to needs. In this example, it is 640; batch_size is a larger value within the memory range, and its general value is 8; momentum is the learning rate momentum in deep learning, and its initial value is set according to the situation of the dataset; learning rate is the learning rate in deep learning, which is understood as taking empirical values, and its initial value is set according to the situation of the dataset; box is the coefficient of GIou loss, which is set to 2; cls is the coefficient of classification loss, and obj is the coefficient of object loss. The initial values ​​of the two are also set according to the situation of the dataset.

9. The photovoltaic power station switch position identification method based on significance analysis and deep learning technology according to claim 7 is characterized in that: The specific method of ensuring that the training model converges in Step 3.3 and selecting the optimal model as the formal recognition model is as follows: ensure that the training model has reached convergence through the loss situation during training, and then select the model with the lowest loss value as the test model, perform test feedback through the test data set, debug the parameters, and finally select the parameter model with the best result as the formal recognition model.

Images