A photovoltaic module point spot detection method and system based on tensor decomposition

Abstract

The application discloses a kind of photovoltaic module point-like hot spot detection method and system based on tensor decomposition, belong to infrared image processing field.The method includes: first, for the infrared image of input photovoltaic module is transformed into infrared local tensor set using sliding window method;Second, the initial background component is expanded in vectorization mode, and the improved Tucker rank of different weight is constructed to realize the feature representation of background component in tensor;Then, establish tensor component decomposition model, and the background, noise and target are separated in all tensors in infrared local tensor set;Then, according to the target component finally obtained, the point-like hot spot confidence map of original infrared image is reconstructed;Finally, calculate segmentation threshold to detect point-like hot spot.The application can quickly detect and identify point-like hot spot from the infrared image of photovoltaic power station inspection and shooting, solve the problem of low efficiency of maintenance photovoltaic module fault, and the detection and identification precision is high and fast.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to a method and system for detecting point hot spots in photovoltaic modules based on tensor decomposition. Background Technology

[0002] In recent years, the energy crisis has become a serious problem facing countries worldwide. The annual consumption of traditional energy sources has led to the continuous depletion of energy reserves, while also causing irreversible pollution to the ecological environment. Therefore, it is urgent to actively promote sustainable, environmentally friendly, and clean new energy sources to replace traditional energy sources, which has become a consensus among countries. Against this backdrop, photovoltaic power generation technology, as a new energy alternative that converts solar energy into electricity, has become an important component of the new energy strategy. However, with the continuous development of photovoltaic power generation, how to achieve efficient operation and maintenance of large-scale photovoltaic power plants and timely detect and eliminate faults and defects in photovoltaic modules has become an urgent problem for the photovoltaic industry to solve. Photovoltaic module failures can cause hot spot effects. Therefore, using drones equipped with infrared cameras for intelligent inspection to replace traditional manual inspections can significantly improve the operation and maintenance efficiency of photovoltaic power plants while reducing operation and maintenance costs. In actual photovoltaic power plant operation and maintenance scenarios, point hot spots on photovoltaic modules are the most common type of failure, and efficient and rapid detection of point hot spots has important practical significance.

[0003] In UAV inspection infrared images, point-like hot spot defects in photovoltaic modules appear as small, bright pixel areas with abnormally high temperatures. Most existing point-like hot spot detection methods for photovoltaic modules, both domestically and internationally, are based on electrical characteristics or handheld infrared detection, with little research on point-like hot spot detection technology specifically for UAV infrared images. The method for detecting point-like hot spots based on mirror gradient features proposed in patent CN114022482A—"Method, Apparatus, Equipment and Readable Storage Medium for Detecting Point-like Hot Spots in Photovoltaic Panels"—is difficult to adapt to inspection scenarios with low image signal-to-noise ratios and weak point-like hot spot characteristics. Because low infrared image signal-to-noise ratios, small point-like hot spot areas, and background clutter interference can affect point-like hot spot detection, how to correctly detect point-like hot spots while suppressing erroneous detections caused by background and noise during the detection process is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for detecting point hot spots in photovoltaic modules based on tensor decomposition. This method improves the accuracy of point hot spot detection in intelligent inspections of photovoltaic power plants using drones equipped with infrared sensors, and solves the problem of point hot spot detection failure caused by low signal-to-noise ratio and noise interference in images.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention proposes a method for detecting point hot spots in photovoltaic modules based on tensor decomposition, comprising the following steps:

[0007] (1) Input the infrared image of the photovoltaic module, construct the infrared local tensor according to the sliding window method, and traverse the entire infrared image to transform the infrared image into a set of infrared local tensors;

[0008] (2) Using the infrared local tensor as the data unit, the value of the infrared local tensor is used as the initial value of the background component of the infrared local tensor. Then, the background component is expanded by vectorization mode and an improved Tucker rank is constructed to realize the feature representation of the background component in the infrared local tensor.

[0009] (3) Using vectorized low-rank constraints and tensor sparsity constraints, an infrared local tensor component decomposition model is established. By solving the tensor decomposition optimal problem, background components, noise components and target components are decomposed in all infrared local tensors in the infrared local tensor set.

[0010] (4) Based on the target components finally obtained in step (3), the point hot spot confidence map of the original infrared image is reconstructed, and the segmentation threshold is calculated based on the point hot spot confidence map to detect the point hot spot.

[0011] Furthermore, the construction of the infrared local tensor based on the sliding window method specifically involves:

[0012]

[0013] in For containing I Ω neighborhood window, It is a local window of infrared image I, and Let (i) be the infrared local tensor, and (:,:,i) represent the i-th forward slice of the tensor.

[0014] Furthermore, the step of traversing the entire infrared image to transform it into a set of local infrared tensors specifically involves:

[0015] Slide the window sequentially from left to right and from top to bottom on the original infrared image according to the specified sliding window step size until the infrared image is traversed, and obtain the set of infrared local tensors obtained by infrared image transformation.

[0016] Furthermore, the vectorization pattern expansion of the background components is as follows:

[0017]

[0018]

[0019]

[0020] in, It is the background component of the infrared local tensor. The initial value is equal to the infrared local tensor. The vectorized expansion matrix obtained by expanding the slice vectorization mode 1. The vectorized expansion matrix obtained by expanding the slice vectorization mode 2. This is the vectorized expansion matrix obtained from slice vectorization mode 3. `vec(.)` represents the vectorization operation. This represents the c-th forward slice of the background components. This is the b-th lateral slice of the background component. This is the a-th level slice of the background component, and [.] indicates a splicing operation.

[0021] Furthermore, the improved Tucker rank with different weights is:

[0022]

[0023] Where λ i >0 represents the pattern weight, λ1>λ2, λ1>λ3, and Denotes the nuclear norm of a matrix. Background components of the infrared local tensor The improved Tucker rank, i.e., the characteristic representation of the background components of the infrared local tensor.

[0024] Furthermore, the tensor component decomposition model is as follows:

[0025]

[0026] in Indicates background components, Indicates the target component, Indicates noise components, For infrared local tensors, This means finding the function that minimizes the given information. ||·||1 represents L F Sparse norm, ||·|| F Let Frobenius norm be denoted, and α1, α2 > 0 be the constraint weight parameters.

[0027] Furthermore, the point-like hotspot confidence map of the original infrared image reconstructed from the target components finally obtained in step (3) is specifically as follows:

[0028] Based on the target components obtained in step (3), a reconstruction matrix with the same size as the original infrared image is constructed. In the reconstruction matrix, the target components are enhanced and the other components are suppressed. The resulting reconstruction matrix is ​​the confidence map for spot hotspot detection.

[0029] Furthermore, the segmentation threshold is

[0030] τ=(μ+β1×σ,β2×T max )

[0031] Where τ represents the segmentation threshold, μ represents the mean confidence level of the point hotspot detection confidence map T, and T max σ represents the maximum confidence level of the confidence map for point hot spot detection, σ represents the standard deviation of the confidence level of the confidence map for point hot spot detection, and β1,β2>0 represent empirical parameters.

[0032] The present invention also proposes a photovoltaic module point hot spot detection system for the above method, comprising:

[0033] The infrared image processing module is used to input infrared images of photovoltaic modules and convert the infrared images into sets of infrared local tensors.

[0034] The background component feature module is used to perform vectorized mode expansion of the infrared local tensor and construct an improved Tucker rank with different weights to realize the feature representation of the background components in the tensor.

[0035] The tensor component decomposition module is used to separate the background from the target in all infrared local tensors in the infrared local tensor set;

[0036] The detection module is used to obtain the point hotspot confidence map of the original infrared image based on the target tensor components obtained by the tensor component decomposition module, and to calculate the segmentation threshold based on the point hotspot confidence map to detect the point hotspots.

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] (1) Based on the local infrared tensor encoding of point hot spot features in infrared images, this invention can fully explore the correlation characteristics between point hot spots and the surrounding background, solve the interference of global background in infrared images on point hot spot detection, and improve the robustness of detection and recognition.

[0039] (2) The improved Tucker rank constraint of the present invention can fully characterize the similarity between local background regions. Normal components and ground objects in the infrared image of photovoltaic modules have similar temperatures and similar structures in local regions. The low-rank representation expanded by vectorization mode can maintain the local structural characteristics of the image, thereby realizing diversified and automated modeling of the background and improving the robustness of detection and recognition.

[0040] (3) The tensor component decomposition model of the present invention can effectively separate point hot spot targets from background and noise, effectively making up for the problem that traditional detection models are difficult to adapt to low signal-to-noise ratio images, and improving the accuracy and robustness of point hot spot detection and recognition. Attached Figure Description

[0041] Figure 1 This is a flowchart of the present invention;

[0042] Figure 2 The image is an infrared image taken by an unmanned aerial vehicle (UAV) during intelligent inspection of a power station according to the present invention.

[0043] Figure 3 This is a schematic diagram of the local infrared tensor group structure provided in an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the detection results of point hot spots provided in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention is described in detail below with reference to specific embodiments. Specific embodiments are described below to simplify the invention. However, it should be understood that the invention is not limited to the described embodiments, and various modifications are possible without departing from the basic principles; these equivalent forms also fall within the scope defined by the appended claims.

[0046] See Figure 1 The present invention provides a method for detecting point hot spots in photovoltaic modules based on tensor decomposition, comprising:

[0047] Step 1: Input the infrared image of the photovoltaic module, construct the infrared local tensor according to the sliding window method, and traverse the entire infrared image to transform the infrared image into a set of infrared local tensors.

[0048] In this embodiment, infrared images captured by an infrared camera mounted on a drone during an inspection of a photovoltaic power station are collected, such as... Figure 2 As shown, the infrared image size is 225×225. Considering that the pixel area of ​​the dot-shaped hot spot is small, its grayscale distribution on the infrared image has local differences from the neighboring normal components and background pixels. Therefore, a local window of the infrared image is selected as the processing unit to construct an infrared local tensor unit.

[0049] Specifically, normally operating photovoltaic (PV) modules maintain a constant temperature range, and the infrared characteristics of the PV module background remain similar. Point-like hot spots on PV modules are small, point-like areas of abnormal heating caused by faults such as short circuits, damage, or contamination; their temperature is generally higher than that of neighboring normal modules and the background. Utilizing this characteristic, we can focus on local regions of the input infrared image and construct an infrared local tensor structure using a specific local window, such as... Figure 3 As shown in (a).

[0050] The specific implementation of the infrared local tensor is shown in formulas (1) and (2), where, For containing I Ω neighborhood window, It is a local window of infrared image I, i = 1, ..., 9, and Let (:,:,i) be the infrared local tensor, and (:,:,i) represent the i-th forward slice of the tensor. Considering that the size of point hotspots is usually not less than 25×25, the local window scale w is set to 25 in this embodiment.

[0051]

[0052] Furthermore, the location of the point-like hotspots in the infrared image is random and unknown. Therefore, it is necessary to process the entire infrared image. A sliding window method is used to sequentially slide the window from left to right and from top to bottom on the original infrared image according to a specified sliding window step size *s*, until the entire infrared image is traversed. The resulting set of infrared local tensors from the input inspected infrared image is then transformed, as shown below. Figure 3 As shown in (b). Considering that the entire image region needs to be processed, and the local window size is 25, in this specific embodiment, the sliding window step size s is 25, and the number of tensors in the infrared local tensor is 81.

[0053] Step 2: Using the infrared local tensor as the data unit, the value of the infrared local tensor is used as the initial value of the background component of the infrared local tensor. Then, the background component is expanded into a vectorized pattern to construct an improved Tucker rank to realize the feature representation of the background component in the tensor.

[0054] Specifically, the infrared local tensor constructed in this embodiment contains the background information of the infrared image. The background information of the infrared image includes normal photovoltaic modules and image features. Normal photovoltaic modules and image features have structural similarities in the image, which is reflected in the low-rank nature of the infrared local tensor. Considering the diversity of tensor low-rank nature, the similarity of a slice of a certain dimension of the tensor can characterize the structural similarity of adjacent local background regions. The background components are sliced ​​and vectorized along different dimensions.

[0055] The specific implementation of the tensor slice vectorization mode expansion is shown in formulas (2), (3), and (4). yes Background components, initial value and equal, The vectorized expansion matrix obtained by expanding the slice vectorization mode 1. The vectorized expansion matrix obtained by expanding the slice vectorization mode 2. This is the vectorized expansion matrix obtained from slice vectorization mode 3. `vec(.)` represents the vectorization operation. This represents the c-th forward slice of the background components. This is the b-th lateral slice of the background component. The a-th horizontal slice of the background component, [.] indicates a stitching operation. The vectorized unfolded matrix obtained by unfolding slice vectorization modes 1, 2, and 3 is obtained by vectorizing the forward slice, lateral slice, and horizontal slice and then horizontally stitching them together. Specifically, in the embodiment, It is a 625×9 dimension matrix. and It is a 225×25 matrix.

[0056]

[0057]

[0058]

[0059] Furthermore, based on the vectorized unfolded matrices of slices obtained from the slice vectorization pattern, considering the structural similarity between normal photovoltaic modules and image features, which can be represented by the correlation of column vectors in the vectorized unfolded matrix, the tensor Tucker rank is used to instantiate the column vector correlation. Since the importance of structural similarity in the unfolded matrices of different slice dimensions varies, an improved Tucker rank with different weights is introduced for the unfolded matrices of different patterns.

[0060] The specific implementation of the improved Tucker rank is shown in formula (5), where λ i >0 represents the pattern weight, and Denotes the nuclear norm of a matrix. for The improved Tucker rank, i.e., the feature representation of the background components of the infrared local tensor. Since the mode 1 expansion matrix can characterize the image structure characteristics between neighboring windows, it needs to be assigned a larger weight. Therefore, in the specific embodiment, λ1 is 0.8, and λ2 and λ3 are 0.1.

[0061]

[0062] Step 3: Using vectorized low-rank constraints and tensor sparsity constraints, establish a tensor component decomposition model. By solving the tensor decomposition optimization problem, decompose all tensors in the infrared local tensor set into background components, noise components, and target components.

[0063] Specifically, to detect point hotspots in the entire infrared image, it is necessary to process the infrared local tensor set obtained by transforming the input infrared image. Furthermore, since infrared cameras typically generate Gaussian noise during imaging, any tensor unit contains target, background, and noise components, with the target component being the point hotspot. Therefore, establishing a tensor component decomposition model for the infrared local tensor unit can achieve the separation of point hotspots from the background and noise.

[0064] Furthermore, considering that point hotspots are collections of high-temperature pixels in small regions, and their tensors are sparsity, the L1 sparse norm is used for characterization. Normal components and image features can be characterized using the improved Tucker rank described above, while Gaussian noise is characterized using the Frobenius norm.

[0065] The specific implementation of the tensor component decomposition model is shown in formula (6), where Indicates background components, Indicates the target component, Indicates noise components, This means finding the function that minimizes the given information. ||·||1 represents the L1 sparse norm,||·|| F Let α1 and α2 represent the Frobenius norm, and α1 and α2 > 0 be the constraint weight parameters. In the specific embodiment, α1 and α2 take values ​​of 0.5 and 25, respectively.

[0066]

[0067] Step 4: Based on the target components obtained in Step 3, reconstruct the confidence map of the point hotspots in the original image. Based on the reconstructed confidence map of the point hotspots, design a segmentation threshold to detect the point hotspots.

[0068] Specifically, based on the tensor component decomposition model described above, it is possible to separate point hotspots from normal components, image features, and noise, wherein point hotspots are classified into target components. In the middle, normal components and image features are classified into the background component. In the middle, noise is divided into noise components. The tensor component decomposition model is used sequentially to decompose all infrared local tensors in the infrared local tensor set into components, and the target components are obtained accordingly. A reconstruction matrix with the same size as the original infrared image is constructed. In the reconstruction matrix, the target component is significantly enhanced and the other components are suppressed. The resulting reconstruction matrix is ​​the confidence map T for the detection of point hot spots. Based on this confidence map, a segmentation threshold is used for detection.

[0069] Furthermore, based on the obtained confidence map T for point hot spots, an adaptive detection threshold is constructed to achieve the final detection of point hot spots. The segmentation threshold is shown in formula (7), where τ represents the segmentation threshold, μ represents the mean confidence level of the detection confidence map T, and T max The confidence level represents the maximum confidence level of the confidence plot. β1,β2>0 indicates empirical parameters, and σ represents the standard deviation of the confidence level of the confidence plot. Specifically, in this embodiment, β1 and β2 take values ​​of 5 and 0.7, respectively. Figure 4 This is a diagram showing the detection results of this embodiment.

[0070] τ=(μ+β1×σ,β2×T max (7)

[0071] This embodiment also proposes a photovoltaic module point hot spot detection system for implementing the above method, which includes:

[0072] The infrared image processing module is used to input infrared images of photovoltaic modules and convert the infrared images into sets of infrared local tensors.

[0073] The background component feature module is used to perform vectorized mode expansion of the infrared local tensor and construct an improved Tucker rank with adaptive weights to realize the feature representation of the background components in the infrared local tensor.

[0074] The tensor component decomposition module is used to decompose all infrared local tensors in the infrared local tensor set into background components, noise components, and target components.

[0075] The detection module is used to obtain the point hotspot confidence map of the original infrared image based on the target tensor components obtained by the tensor component decomposition module, and to calculate the segmentation threshold based on the point hotspot confidence map to detect the point hotspots.

[0076] The accompanying drawings illustrate the objectives, technical solutions, and advantages of this invention, making them clearer and more understandable. It is particularly important to emphasize that these specific examples are for illustrative purposes only and are not intended to limit the scope of the invention. Any equivalent substitutions or improvements that conform to the methodological ideas and principles provided by this invention should be included within the scope of protection of this invention.

Claims

1. A method for detecting point-like hot spots in photovoltaic modules based on tensor decomposition, characterized in that, Includes the following steps: (1) Input the infrared image of the photovoltaic module, construct the infrared local tensor according to the sliding window method, and traverse the entire infrared image to transform the infrared image into a set of infrared local tensors; The specific steps for constructing the infrared local tensor based on the sliding window method are as follows: ； in, For inclusion neighborhood window, It is an infrared image. Local window, For infrared local tensors, This represents the i-th forward slice of a tensor; The process of traversing the entire infrared image to transform it into a set of local infrared tensors specifically involves: Slide the window sequentially from left to right and from top to bottom on the original infrared image according to the specified sliding window step size until the infrared image is traversed to obtain the set of infrared local tensors obtained by infrared image transformation. (2) Using the infrared local tensor as the data unit, the value of the infrared local tensor is used as the initial value of the background component of the infrared local tensor. Then, the background component is expanded by vectorization mode and an improved Tucker rank is constructed to realize the feature representation of the background component in the infrared local tensor. The improved Tucker rank is: ； in Expand the weights for the pattern. ,and , Denotes the nuclear norm of a matrix. Background components of the infrared local tensor The improved Tucker rank, i.e., the characteristic representation of the background component of the infrared local tensor; The specific steps of vectorizing and unfolding the background components are as follows: ； ； ； in, It is the background component of the infrared local tensor. The vectorized expansion matrix obtained by expanding the slice vectorization mode 1. The vectorized expansion matrix obtained by expanding the slice vectorization mode 2. The vectorized expansion matrix obtained by expanding the slice vectorization mode 3. For vectorization operations, This represents the c-th forward slice of the background components. This is the b-th lateral slice of the background component. This represents the a-th level slice of the background component, and [.] denotes a matrix concatenation operation; (3) Using vectorized low-rank constraints and tensor sparsity constraints, an infrared local tensor component decomposition model is established. By solving the tensor decomposition optimal problem, background components, noise components and target components are decomposed in all infrared local tensors in the infrared local tensor set. (4) Based on the target components finally obtained in step (3), the point hot spot confidence map of the original infrared image is reconstructed, and the segmentation threshold is calculated based on the point hot spot confidence map to detect the point hot spot; The specific steps for reconstructing the point hot spot confidence map of the original infrared image based on the target components finally obtained in step (3) are as follows: construct a reconstruction matrix with the same size as the original infrared image based on the target components obtained in step (3). In the reconstruction matrix, the target components are enhanced and the other components are suppressed. The resulting reconstruction matrix is ​​the point hot spot detection confidence map. The segmentation threshold is ； in Indicates the segmentation threshold. Confidence map for point-like hot spot detection The confidence level mean, This represents the maximum confidence level of the confidence map for point-like hot spot detection. This represents the standard deviation of the confidence level of the confidence plot for point-like hot spot detection. This represents an empirical parameter.

2. The method for detecting point hot spots in photovoltaic modules based on tensor decomposition according to claim 1, characterized in that, The tensor component decomposition model is as follows: ； in Indicates background components, Indicates the target component, Indicates noise components, For infrared local tensors, This means finding the function that minimizes the given information. , express Sparse norm, Denotes the Frobenius norm. These are the constraint weight parameters.

3. A photovoltaic module point hot spot detection system for implementing the method of claim 1, characterized in that, It includes: The infrared image processing module is used to input infrared images of photovoltaic modules and convert the infrared images into sets of infrared local tensors. The background component feature module is used to perform vectorized mode expansion of the infrared local tensor and construct an improved Tucker rank with different weights to realize the feature representation of the background components in the tensor. The tensor component decomposition module is used to decompose all infrared local tensors in the infrared local tensor set into background components, noise components, and target components. The detection module is used to obtain the point hotspot confidence map of the original infrared image based on the target tensor components obtained by the tensor component decomposition module, and to calculate the segmentation threshold based on the point hotspot confidence map to detect the point hotspots.

Images