Photovoltaic panel image segmentation and anomaly identification method

Abstract

The invention relates to the technical field of image processing and deep learning, in particular to a photovoltaic panel image segmentation and anomaly recognition method, which comprises the following steps: firstly, acquiring a visible light image of a photovoltaic station, carrying out four-vertex calibration on a photovoltaic panel in the visible light image, creating a photovoltaic panel anomaly recognition data set, and constructing a photovoltaic panel group anomaly recognition network; performing four-point perspective correction on a photovoltaic panel in the photovoltaic panel anomaly identification data set, training a photovoltaic panel anomaly identification model, making a photovoltaic panel group segmentation data set according to the obtained visible light picture, training a photovoltaic panel group segmentation model by using the photovoltaic panel group segmentation data set, performing photovoltaic panel group segmentation and correction on the input photovoltaic station image, and obtaining a photovoltaic panel anomaly identification result. The method comprises the following steps: firstly, acquiring a photovoltaic panel image, then segmenting the photovoltaic panel, sending the photovoltaic panel image to a photovoltaic panel abnormity identification model one by one to predict the state of the photovoltaic panel, finally analyzing a state sequence in a time dimension, and triggering an alarm if abnormity occurs. According to the technical scheme, the detection accuracy of the abnormal state of the photovoltaic panel is higher.

Description

Technical Field

[0001] The present invention relates to the field of image processing and deep learning technology, and in particular to a photovoltaic panel image segmentation and anomaly recognition method. Background Art

[0002] Photovoltaic power stations are usually built in remote mountainous areas, deserts, large ponds and other places. Due to the difficulty in maintenance, there are problems such as loss of photovoltaic panels, dust accumulation, snow accumulation, etc., which seriously affect the normal operation of photovoltaic power stations. Since the workload of manual inspections is very large, it is impossible to inspect at all times. Therefore, the existing technology uses target detection models to detect the status of photovoltaic panels, collect images of photovoltaic stations through cameras, and automatically analyze abnormal conditions of photovoltaic panels.

[0003] However, the method of using the target detection model to detect photovoltaic panels has the problem of missed detection and false detection, and the target detection usually obtains the bounding box of the target, which contains part of the environmental image, which is not conducive to the subsequent abnormal state analysis. On the other hand, due to the complex outdoor lighting conditions and the strong reflectivity of photovoltaic panels, the traditional image segmentation method and the existing depth segmentation method both have the problem of inaccurate segmentation.

[0004] In addition, the existing image classification scheme does not consider the texture characteristics of photovoltaic panels, which can easily lead to classification errors. In the actual environment, photovoltaic panels cannot be segmented when they are lost, dusty, or snowy, so it is necessary to consider using historical coordinate information for segmentation. In fact, there is a high probability that photovoltaic panels cannot be detected when they are dusty or snowy, resulting in false alarms of photovoltaic panel loss. Therefore, the problems of photovoltaic panel loss, dusty, or snowy need to be considered comprehensively.

[0005] In summary, the methods used in the prior art to detect the status of photovoltaic panels of photovoltaic stations have low accuracy in detecting abnormal conditions such as loss, dust accumulation, snow accumulation, etc. of photovoltaic panels of photovoltaic stations in outdoor environments. Summary of the invention

[0006] The purpose of the present invention is to provide a photovoltaic panel image segmentation and anomaly recognition method to solve the problem that the method used in the prior art to detect the status of photovoltaic panels of photovoltaic stations has low accuracy in detecting abnormal conditions such as loss, dust accumulation, snow accumulation, etc. of photovoltaic panels in photovoltaic stations in outdoor environments.

[0007] To achieve the above object, the present invention provides a photovoltaic panel image segmentation and anomaly recognition method, comprising the following steps:

[0008] Obtain visible light images of photovoltaic panels of the photovoltaic station and obtain visible light images of the surrounding environment of the photovoltaic station;

[0009] The four vertices of the photovoltaic panels in the photovoltaic panel visible light image are calibrated to create a photovoltaic panel anomaly recognition data set; a photovoltaic panel group anomaly recognition network is constructed to perform four-point perspective correction on the photovoltaic panels in the photovoltaic panel anomaly recognition data set, and a photovoltaic panel anomaly recognition model is trained;

[0010] Based on the obtained visible light images, mark the outline of each photovoltaic panel group and create a photovoltaic panel group segmentation dataset. Use the photovoltaic panel group segmentation dataset to train the photovoltaic panel group segmentation model;

[0011] Perform photovoltaic panel group segmentation on the input photovoltaic station image;

[0012] Correct each photovoltaic panel group and perform photovoltaic panel segmentation;

[0013] If there is no historical segmentation result, each photovoltaic panel image is sent to the photovoltaic panel anomaly recognition model to predict the photovoltaic panel status;

[0014] If there are historical segmentation results, and the photovoltaic panel group and photovoltaic panel segmentation results are consistent with the historical segmentation results, the photovoltaic panel images are sent to the photovoltaic panel anomaly recognition model one by one to predict the photovoltaic panel status;

[0015] If there are historical segmentation results, and the photovoltaic panel group or photovoltaic panel segmentation results are inconsistent with the historical segmentation results, the photovoltaic panel group image is intercepted using the saved coordinate information, and the photovoltaic panel image is intercepted using the saved coordinate information after correction, and the photovoltaic panel images are sent to the photovoltaic panel anomaly recognition model one by one to predict the photovoltaic panel status;

[0016] Finally, the state sequence in the time dimension is analyzed, and an alarm is triggered if an abnormality occurs.

[0017] Among them, in the step "constructing a photovoltaic panel group anomaly recognition network, performing four-point perspective correction on the photovoltaic panels in the photovoltaic panel anomaly recognition data set, and training the photovoltaic panel anomaly recognition model", the constructed photovoltaic panel group anomaly recognition network includes a ResNet backbone, a horizontal CNN backbone, and a vertical CNN backbone. The horizontal CNN backbone includes a horizontal convolution layer and a pooling layer, and the vertical CNN backbone includes a vertical convolution layer and a pooling layer; the outputs of the three backbones are connected to a fully connected layer, and then connected to 3 Softmax layers, and finally output 3 states, representing the loss, dust accumulation, and snow accumulation states respectively.

[0018] Among them, in the step of "segmenting the input photovoltaic station image into photovoltaic panel groups", the method of photovoltaic panel group segmentation is specifically as follows:

[0019] Use the photovoltaic panel group segmentation model to segment the photovoltaic panels in the input visible light image. Set the pixels belonging to the photovoltaic panel group in the segmentation result to 255 and other pixels to 0 to obtain the initial segmentation map. Detect all contours in the initial segmentation map, filter out small contours, and obtain all photovoltaic panel groups;

[0020] The area where each photovoltaic panel group is located is divided into multiple areas, and the average color of the central area of ​​each area is counted in the visible light image to obtain the color set of each photovoltaic panel group;

[0021] In the visible light image, a certain number of seed points are uniformly selected from the long axis direction of each photovoltaic panel group. Starting from these seed points, the region growing method is used to find the precise outline of the photovoltaic panel group. The rule of region growing is to compare the similarity between the color of the adjacent pixel points in the image and the color set of the photovoltaic panel group. If the similarity is large, it is set to 1, otherwise it is set to 0, thus obtaining a secondary segmentation map;

[0022] Fill the holes in the secondary segmentation map. If the neighborhood of black pixels in the secondary segmentation map is surrounded by white pixels, modify them to white pixels.

[0023] Detect the contours in the secondary segmentation image, filter out small contours, save the coordinate information of the remaining contours, and complete the segmentation of the photovoltaic panel group.

[0024] Among them, in the step "the rule of region growing is to compare the similarity between the colors of adjacent pixels in the image and the color set of the photovoltaic panel group", the method for calculating the similarity between the color of the pixel point and the color set of the photovoltaic panel group is to calculate the Cartesian distance between the color of the pixel point and each color of the color set of the photovoltaic panel group one by one, and take the minimum Cartesian distance / 442 as the similarity.

[0025] Among them, in the step of "correcting each photovoltaic panel group and performing photovoltaic panel segmentation", the specific method of photovoltaic panel segmentation is:

[0026] The rectified photovoltaic panel group image is grayed, and its average gray value A is counted. In the range of [A-60, A+60], with a step size of 10, multiple gray segmentation thresholds are obtained;

[0027] Dividing the photovoltaic panel group into a plurality of zones;

[0028] For each area, use the above grayscale segmentation threshold to perform binary segmentation one by one, segment the edges of the photovoltaic panels, and count the number and position of their rows and columns;

[0029] The segmentation result with the most uniform row and column positions and the most suitable number is selected as the photovoltaic panel segmentation coordinates.

[0030] Among them, in the step of "analyzing the state sequence on the time dimension", the specific method of analyzing the state sequence on the time dimension is:

[0031] If the state sequence of the photovoltaic panel abnormality recognition model changes from normal to snow accumulation, a photovoltaic panel snow accumulation alarm is generated; when the snow accumulation alarm is not triggered, if the state sequence of the photovoltaic panel abnormality recognition model changes from normal to dust accumulation, a photovoltaic panel dust accumulation alarm is generated; when the dust accumulation alarm is not triggered, if the state sequence of the photovoltaic panel abnormality recognition model changes from existence to loss, a photovoltaic panel lost alarm is generated.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present invention is more accurate and controllable in the segmentation of photovoltaic panel groups, ensuring the normal correction and segmentation of subsequent photovoltaic panel groups. The traditional image segmentation method and the deep image segmentation model are greatly affected by light in the segmentation of photovoltaic panel groups, and there are relatively many edge pixel mis-segments, which is not conducive to the subsequent correction of photovoltaic panel groups. The present invention first uses the deep image segmentation model to perform the initial segmentation of the photovoltaic panel group, and then divides the segmented photovoltaic panel group into sections and counts the color set of the central area, and uses the regional growth method to perform secondary segmentation according to the color similarity to obtain an accurate photovoltaic panel group segmentation map. Since there are many photovoltaic panel edges inside the photovoltaic panel group, the regional growth method will select multiple seed points, and the seed points are evenly distributed to ensure that each photovoltaic panel can be included. In addition, since the segmentation results of the photovoltaic panel are inconsistent with the historical results when it is lost, dusty, or snowy, it is even impossible to perform the photovoltaic panel group segmentation operation. Therefore, the segmentation coordinate information of the photovoltaic panel is recorded each time the segmentation is completed. When the segmentation result is inconsistent with the historical result, the photovoltaic panel group and the photovoltaic panel image are intercepted using the historical coordinate information for further state analysis.

[0034] The present invention uses partition multi-threshold binarization to segment the photovoltaic panel group image, thereby improving the segmentation accuracy of the photovoltaic panel. The present invention uses the color difference between the photovoltaic panel cell and the photovoltaic panel frame, selects a certain intermediate value of the grayscale of the photovoltaic panel cell and the photovoltaic panel frame to binarize the photovoltaic panel group, and obtains the photovoltaic panel frame information. There are differences in the colors of multiple partitions within a photovoltaic panel group, so the present invention divides the photovoltaic panel group into multiple zones and binarizes them separately; in addition, the optimal segmentation threshold for each zone is difficult to determine, so the present invention tries and selects multiple thresholds within the range near the grayscale mean.

[0035] The present invention effectively solves the problem of detecting abnormal conditions such as loss, dust accumulation, snow accumulation, etc. of photovoltaic panels in photovoltaic stations in outdoor environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 It is a flow chart of the photovoltaic panel image segmentation and anomaly recognition method provided by the present invention.

[0038] Figure 2 This is a flow chart of photovoltaic panel group segmentation provided by the present invention.

[0039] Figure 3 This is a flow chart of photovoltaic panel segmentation provided by the present invention.

[0040] Figure 4 This is a flow chart of the photovoltaic panel anomaly recognition model training provided by the present invention. DETAILED DESCRIPTION

[0041] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0042] Please see attached Figures 1 to 4 , a photovoltaic panel image segmentation and anomaly recognition method, the specific implementation plan is:

[0043] S1. Model training:

[0044] Obtain visible light images of the photovoltaic panels of the photovoltaic station, and obtain visible light images of the surrounding environment of the photovoltaic station.

[0045] Calibrate the four vertices of the photovoltaic panel in the visible light image of the photovoltaic panel and create a photovoltaic panel anomaly recognition dataset;

[0046] A photovoltaic panel group anomaly recognition network is constructed. The photovoltaic panel group anomaly recognition network includes a ResNet backbone, a horizontal CNN backbone, and a vertical CNN backbone. The horizontal CNN backbone includes a horizontal convolution layer and a pooling layer, and the vertical CNN backbone includes a vertical convolution layer and a pooling layer. The outputs of the three backbones are connected to a fully connected layer, and then connected to three Softmax layers. Finally, three states are output, indicating the loss, dust accumulation, and snow accumulation states respectively.

[0047] The photovoltaic panels in the photovoltaic panel anomaly recognition data set are cut out, and four-point perspective correction is performed, and then sent to the photovoltaic panel group anomaly recognition network for training to obtain the photovoltaic panel anomaly recognition model.

[0048] In this embodiment, the DeepLabV3 segmentation model is adopted, labelme is used to mark the outlines of all photovoltaic panel groups, a photovoltaic panel group segmentation data set is prepared, and the DeepLabV3 photovoltaic panel group segmentation model is trained.

[0049] S2. Segment the photovoltaic panel groups of the input photovoltaic station image:

[0050] Obtain a visible light image of the photovoltaic station, and use the photovoltaic panel group segmentation model to segment the photovoltaic panel groups in the input image. In this embodiment, the DeepLabV3 photovoltaic panel group segmentation model is used to segment the photovoltaic panel groups in the input image, and the pixels belonging to the photovoltaic panel group in the segmentation result are set to 255, and the other pixels are set to 0 to obtain an initial segmentation map, detect all contours in the initial segmentation map, filter out small contours, and obtain all photovoltaic panel groups.

[0051] The area where each photovoltaic panel group is located is divided into multiple areas, and the average color of the central area of ​​each area is counted in the visible light image to obtain the color set of each photovoltaic panel group.

[0052] In the visible light image, a certain number of seed points are uniformly selected from the long axis direction of each photovoltaic panel group. Starting from these seed points, the region growing method is used to find the precise outline of the photovoltaic panel group. The rule of region growing is to compare the similarity between the color of the adjacent pixel points in the image and the color set of the photovoltaic panel group. If the similarity is large, it is set to 1, otherwise it is set to 0, so as to obtain a secondary segmentation map. The similarity calculation method of the pixel color and the photovoltaic panel color set is to calculate the Cartesian distance between the pixel color and each color of the photovoltaic panel color set one by one, and take the minimum Cartesian distance / 442 as the similarity.

[0053] Fill the holes in the secondary segmentation map. If the neighborhood of black pixels in the secondary segmentation map is surrounded by white pixels, modify them to white pixels.

[0054] Detect the contours in the secondary segmentation image, filter out small contours, save the coordinate information of the remaining contours, and complete the segmentation of the photovoltaic panel group.

[0055] S3. Segment the photovoltaic panels in the photovoltaic panel group:

[0056] The screenshot of the photovoltaic panel group can be corrected by using the four-point perspective transformation method to correct the photovoltaic panel group to the horizontal direction.

[0057] The corrected photovoltaic panel group image is grayed, and its average gray value A is counted. In the range of [A-60, A+60], with a step size of 10, multiple gray segmentation thresholds are obtained.

[0058] The photovoltaic panel group is divided into multiple zones, and the number of zones is determined according to the length of the photovoltaic panel group. The longer the photovoltaic panel group is, the greater the color difference inside the panel group is, and the more zones there are. For each zone, the above-mentioned grayscale segmentation threshold is used for binary segmentation one by one to segment the edges of the photovoltaic panels, and the number and position of rows and columns are counted. The segmentation result with the most uniform row and column positions and the most suitable number is taken as the photovoltaic panel segmentation coordinate. In this embodiment, the number range of photovoltaic panels in the photovoltaic panel group is set in advance for comparison reference.

[0059] S4. Photovoltaic panel abnormal status identification:

[0060] If there is no historical segmentation result, each photovoltaic panel image will be sent to the photovoltaic panel anomaly recognition model to predict the photovoltaic panel status.

[0061] If there are historical segmentation results, and the photovoltaic panel group and photovoltaic panel segmentation results are consistent with the historical segmentation results, the photovoltaic panel images are sent to the photovoltaic panel anomaly recognition model one by one to predict the photovoltaic panel status.

[0062] If there are historical segmentation results, and the photovoltaic panel group or photovoltaic panel segmentation results are inconsistent with the historical segmentation results, the photovoltaic panel group image is intercepted using the saved coordinate information, and the photovoltaic panel image is intercepted using the saved coordinate information after correction, and the photovoltaic panel images are sent to the photovoltaic panel anomaly recognition model one by one to predict the photovoltaic panel status.

[0063] S5. Analysis of abnormal events of photovoltaic panels in the time dimension

[0064] The state sequence in the time dimension is analyzed, and an alarm is triggered if an abnormality occurs. Specifically, if the state sequence of the photovoltaic panel abnormality recognition model changes from normal to snow accumulation, a photovoltaic panel snow accumulation alarm is generated; when the snow accumulation alarm is not triggered, if the state sequence of the photovoltaic panel abnormality recognition model changes from normal to dust accumulation, a photovoltaic panel dust accumulation alarm is generated; when the dust accumulation alarm is not triggered, if the state sequence of the photovoltaic panel abnormality recognition model changes from existence to loss, a photovoltaic panel loss alarm is generated.

[0065] A photovoltaic panel image segmentation and abnormality recognition method of the present invention first obtains a visible light image of a photovoltaic panel of a photovoltaic station, obtains a visible light image of the surrounding environment of the photovoltaic station, calibrates the four vertices of the photovoltaic panel in the visible light image of the photovoltaic panel, creates a photovoltaic panel abnormality recognition data set, builds a photovoltaic panel group abnormality recognition network, performs four-point perspective correction on the photovoltaic panels in the photovoltaic panel abnormality recognition data set, and trains a photovoltaic panel abnormality recognition model. According to the obtained visible light image, the outline of each photovoltaic panel group is marked, and a photovoltaic panel group segmentation data set is made. The photovoltaic panel group segmentation model is trained using the photovoltaic panel group segmentation data set, and the input photovoltaic station image is segmented and corrected by the photovoltaic panel group, and then the photovoltaic panel is segmented, and the photovoltaic panel images are sent one by one to the photovoltaic panel abnormality recognition model to predict the photovoltaic panel state. Finally, the state sequence in the time dimension is analyzed, and an alarm is triggered if an abnormality occurs. The present invention effectively solves the photovoltaic panel segmentation of photovoltaic power station images under light interference and the abnormal state recognition problems such as photovoltaic panel loss, dust accumulation, and snow accumulation.

[0066] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A photovoltaic panel image segmentation and anomaly recognition method, characterized in that: The steps include: Obtain visible light images of photovoltaic panels of the photovoltaic station and obtain visible light images of the surrounding environment of the photovoltaic station; The four vertices of the photovoltaic panels in the photovoltaic panel visible light image are calibrated to create a photovoltaic panel anomaly recognition data set; a photovoltaic panel group anomaly recognition network is constructed to perform four-point perspective correction on the photovoltaic panels in the photovoltaic panel anomaly recognition data set, and a photovoltaic panel anomaly recognition model is trained; Based on the obtained visible light images, mark the outline of each photovoltaic panel group and create a photovoltaic panel group segmentation dataset. Use the photovoltaic panel group segmentation dataset to train the photovoltaic panel group segmentation model; Perform photovoltaic panel group segmentation on the input photovoltaic station image; Correct each photovoltaic panel group and perform photovoltaic panel segmentation; If there is no historical segmentation result, each photovoltaic panel image is sent to the photovoltaic panel anomaly recognition model to predict the photovoltaic panel status; If there are historical segmentation results, and the photovoltaic panel group and photovoltaic panel segmentation results are consistent with the historical segmentation results, the photovoltaic panel images are sent to the photovoltaic panel anomaly recognition model one by one to predict the photovoltaic panel status; If there are historical segmentation results, and the photovoltaic panel group or photovoltaic panel segmentation results are inconsistent with the historical segmentation results, the photovoltaic panel group image is intercepted using the saved coordinate information, and the photovoltaic panel image is intercepted using the saved coordinate information after correction, and the photovoltaic panel images are sent to the photovoltaic panel anomaly recognition model one by one to predict the photovoltaic panel status; Finally, the state sequence in the time dimension is analyzed, and an alarm is triggered if an abnormality occurs.

2. The photovoltaic panel image segmentation and anomaly recognition method according to claim 1, characterized in that: In the step "Building a photovoltaic panel group anomaly recognition network, performing four-point perspective correction on photovoltaic panels in the photovoltaic panel anomaly recognition dataset, and training a photovoltaic panel anomaly recognition model", the constructed photovoltaic panel group anomaly recognition network includes a ResNet backbone, a horizontal CNN backbone, and a vertical CNN backbone. The horizontal CNN backbone includes a horizontal convolution layer and a pooling layer, and the vertical CNN backbone includes a vertical convolution layer and a pooling layer; the outputs of the three backbones are connected to a fully connected layer, and then connected to three Softmax layers, and finally three states are output, representing the loss, dust accumulation, and snow accumulation states respectively.

3. The photovoltaic panel image segmentation and anomaly recognition method according to claim 1, characterized in that: In the step of "segmenting the input photovoltaic station image into photovoltaic panel groups", the method of photovoltaic panel group segmentation is specifically as follows: Use the photovoltaic panel group segmentation model to segment the photovoltaic panels in the input visible light image. Set the pixels belonging to the photovoltaic panel group in the segmentation result to 255 and other pixels to 0 to obtain the initial segmentation map. Detect all contours in the initial segmentation map, filter out small contours, and obtain all photovoltaic panel groups; The area where each photovoltaic panel group is located is divided into multiple areas, and the average color of the central area of ​​each area is counted in the visible light image to obtain the color set of each photovoltaic panel group; In the visible light image, a certain number of seed points are uniformly selected from the long axis direction of each photovoltaic panel group. Starting from these seed points, the region growing method is used to find the precise outline of the photovoltaic panel group. The rule of region growing is to compare the similarity between the color of the adjacent pixel points in the image and the color set of the photovoltaic panel group. If the similarity is large, it is set to 1, otherwise it is set to 0, thus obtaining a secondary segmentation map; Fill the holes in the secondary segmentation map. If the neighborhood of black pixels in the secondary segmentation map is surrounded by white pixels, modify them to white pixels. Detect the contours in the secondary segmentation image, filter out small contours, save the coordinate information of the remaining contours, and complete the segmentation of the photovoltaic panel group.

4. The photovoltaic panel image segmentation and anomaly recognition method according to claim 3, characterized in that: In the step "the rule of region growing is to compare the similarity between the colors of adjacent pixels in the image and the color set of the photovoltaic panel group", the method for calculating the similarity between the color of the pixel and the color set of the photovoltaic panel group is to calculate the Cartesian distance between the color of the pixel and each color of the color set of the photovoltaic panel group one by one, and take the minimum Cartesian distance / 442 as the similarity.

5. The photovoltaic panel image segmentation and anomaly recognition method according to claim 1, characterized in that: In the step "correcting each photovoltaic panel group and performing photovoltaic panel segmentation", the specific method of photovoltaic panel segmentation is as follows: The rectified photovoltaic panel group image is grayed, and its average gray value A is counted. In the range of [A-60, A+60], with a step size of 10, multiple gray segmentation thresholds are obtained; Dividing the photovoltaic panel group into a plurality of zones; For each area, use the above grayscale segmentation threshold to perform binary segmentation one by one, segment the edges of the photovoltaic panels, and count the number and position of their rows and columns; The segmentation result with the most uniform row and column positions and the most suitable number is selected as the photovoltaic panel segmentation coordinates.

6. The photovoltaic panel image segmentation and anomaly recognition method according to claim 1, characterized in that: In the step of "analyzing the state sequence on the time dimension", the specific method for analyzing the state sequence on the time dimension is: If the state sequence of the photovoltaic panel abnormality recognition model changes from normal to snow accumulation, a photovoltaic panel snow accumulation alarm is generated; when the snow accumulation alarm is not triggered, if the state sequence of the photovoltaic panel abnormality recognition model changes from normal to dust accumulation, a photovoltaic panel dust accumulation alarm is generated; when the dust accumulation alarm is not triggered, if the state sequence of the photovoltaic panel abnormality recognition model changes from existence to loss, a photovoltaic panel lost alarm is generated.

Images