System and method for identifying dust deposition state of photovoltaic cell panel based on machine learning

Abstract

The embodiment of the invention provides a photovoltaic cell panel dust deposition state recognition system and method based on machine learning. The method comprises the steps that a photovoltaic panel RGB image collected by a camera is acquired; performing multi-scale image feature extraction based on a multi-color space on the image to obtain a shallow-layer fusion feature map, a middle-layer fusion feature map and a deep-layer fusion feature map of the multi-color space of the photovoltaic panel; a self-adaptive attention module is screened through gating on the feature maps to obtain respective adaptive enhanced photovoltaic panel multi-color space fusion feature maps; and based on the multi-color space fusion feature maps of the adaptive enhanced photovoltaic panels, determining the identification result of the dust deposition degree grade. A multi-color space and multi-scale color features of a photovoltaic panel surface image are analyzed by combining a machine learning algorithm to represent the dust state of the photovoltaic panel surface, so that the dust deposition degree and grade are automatically identified.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of intelligent identification technology, and specifically to a system and method for identifying the dust accumulation state of a photovoltaic panel based on machine learning. Background Art

[0002] With the increasing popularity of renewable energy, photovoltaic panels are key components of solar power generation, and their performance and efficiency directly affect the power generation capacity of the entire photovoltaic system. However, photovoltaic panels often face the problem of dust accumulation in practical applications. Dust coverage will significantly reduce the photoelectric conversion efficiency of the panels. Therefore, accurately identifying the dust accumulation on photovoltaic panels and cleaning them in time is crucial to maintaining the efficient operation of the photovoltaic system.

[0003] Traditional methods for identifying dust accumulation status rely on manual inspection, which is inefficient, time-consuming, labor-intensive, and unable to respond in a timely manner. With the rapid development of machine learning technology, especially the breakthrough of deep learning in the field of image processing, a new solution is provided for the automatic identification of the dust accumulation status of photovoltaic panels. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a photovoltaic panel dust accumulation status identification system and method based on machine learning.

[0005] In a first aspect, an embodiment of the present invention provides a method for identifying dust accumulation status of a photovoltaic panel based on machine learning, the method comprising:

[0006] Get the RGB image of the photovoltaic panel captured by the camera;

[0007] Performing multi-scale image feature extraction based on multi-color space on the photovoltaic panel RGB image to obtain a photovoltaic panel multi-color space shallow fusion feature map, a photovoltaic panel multi-color space middle fusion feature map and a photovoltaic panel multi-color space deep fusion feature map;

[0008] The photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map are gated and screened by an adaptive attention module to obtain an adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, an adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and an adaptive enhanced photovoltaic panel multi-color space deep fusion feature map;

[0009] Based on the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map, the recognition result of the dust accumulation degree level is determined.

[0010] In some possible embodiments, the performing multi-scale image feature extraction based on multi-color space on the photovoltaic panel RGB image to obtain a photovoltaic panel multi-color space shallow fusion feature map, a photovoltaic panel multi-color space middle fusion feature map and a photovoltaic panel multi-color space deep fusion feature map includes:

[0011] Performing color space conversion mapping on the photovoltaic panel RGB image to obtain a photovoltaic panel YCbCr image;

[0012] The photovoltaic panel RGB image and the photovoltaic panel YCbCr image are respectively passed through the Xception feature extraction network to obtain a photovoltaic panel RGB shallow feature map, a photovoltaic panel RGB middle feature map, a photovoltaic panel RGB deep feature map, a photovoltaic panel YCbCr shallow feature map, a photovoltaic panel YCbCr middle feature map and a photovoltaic panel YCbCr deep feature map;

[0013] The photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel RGB middle layer feature map and the photovoltaic panel YCbCr middle layer feature map, and the photovoltaic panel RGB deep layer feature map and the photovoltaic panel YCbCr deep layer feature map are fused respectively to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle layer fusion feature map and the photovoltaic panel multi-color space deep fusion feature map.

[0014] In some possible embodiments, performing color space conversion mapping on the photovoltaic panel RGB image to obtain a photovoltaic panel YCbCr image includes:

[0015] The photovoltaic panel RGB image is mapped to a YCbCr color space to obtain the photovoltaic panel YCbCr image.

[0016] In some possible embodiments, the photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel RGB middle feature map and the photovoltaic panel YCbCr middle feature map, and the photovoltaic panel RGB deep feature map and the photovoltaic panel YCbCr deep feature map are respectively fused to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map, including:

[0017] Cascading the photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map to obtain the photovoltaic panel multi-color space shallow fusion feature map;

[0018] Cascading the photovoltaic panel RGB middle-layer feature map and the photovoltaic panel YCbCr middle-layer feature map to obtain the photovoltaic panel multi-color space middle-layer fusion feature map;

[0019] The photovoltaic panel RGB deep feature map and the photovoltaic panel YCbCr deep feature map are cascaded to obtain the photovoltaic panel multi-color space deep fusion feature map.

[0020] In some possible embodiments, the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map are subjected to gated screening of the adaptive attention module to obtain the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map, including:

[0021] Performing maximum-based global pooling processing, random-value-based global pooling processing, and average-based global pooling processing on the photovoltaic panel multi-color space shallow fusion feature map to obtain a photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, a photovoltaic panel multi-color space shallow fusion global random value pooling feature vector, and a photovoltaic panel multi-color space shallow fusion global mean pooling feature vector;

[0022] Fusing the photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, the photovoltaic panel multi-color space shallow fusion global mean pooling feature vector and the photovoltaic panel multi-color space shallow fusion global random value pooling feature vector to obtain a photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector;

[0023] Performing channel-by-channel semantic feature interaction and feature activation on the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector to obtain a photovoltaic panel multi-color space shallow fusion channel semantic association feature vector;

[0024] Inputting the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector into a gating unit to obtain a gated screening photovoltaic panel multi-color space shallow fusion channel semantic association feature vector;

[0025] Normalizing the gated screening photovoltaic panel multi-color space shallow fusion channel semantic association feature vector to obtain the gated screening photovoltaic panel multi-color space shallow fusion channel semantic association weight feature vector;

[0026] The gated screening photovoltaic panel multi-color space shallow fusion channel semantic association weight feature vector is used as the weight feature vector, and its channel-by-channel point multiplication with the photovoltaic panel multi-color space shallow fusion feature map is calculated to obtain the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map.

[0027] In some possible embodiments, the fusion of the photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, the photovoltaic panel multi-color space shallow fusion global mean pooling feature vector and the photovoltaic panel multi-color space shallow fusion global random value pooling feature vector to obtain the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector includes:

[0028] Determine a first photovoltaic panel multi-color space shallow fusion semantic pooling representation factor, a second photovoltaic panel multi-color space shallow fusion semantic pooling representation factor, and a third photovoltaic panel multi-color space shallow fusion semantic pooling representation factor;

[0029] The photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector is obtained by calculating the dot product of the first photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the photovoltaic panel multi-color space shallow fusion global mean pooling feature vector, the dot product of the second photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, and the dot product of the third photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the photovoltaic panel multi-color space shallow fusion global random value pooling feature vector.

[0030] In some possible embodiments, the performing channel-by-channel semantic feature interaction and feature activation on the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector to obtain the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector includes:

[0031] The photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector is passed through a fully connected layer to perform channel-by-channel semantic feature interaction and feature activation on the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector to obtain the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector.

[0032] In some possible embodiments, the determination of the recognition result of the dust accumulation level based on the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map includes:

[0033] Cascading the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map into a photovoltaic panel multi-color space multi-scale fusion feature map;

[0034] The photovoltaic panel multi-color space multi-scale fusion feature map is passed through a dust accumulation state recognition module based on a classifier to obtain the recognition result, and the recognition result is used to represent the dust accumulation degree level label.

[0035] In a second aspect, an embodiment of the present invention provides a photovoltaic panel dust accumulation state recognition system based on machine learning, the system comprising:

[0036] An image acquisition module, used to acquire the RGB image of the photovoltaic panel acquired by the camera;

[0037] A multi-scale image feature extraction module is used to perform multi-scale image feature extraction based on a multi-color space on the photovoltaic panel RGB image to obtain a photovoltaic panel multi-color space shallow fusion feature map, a photovoltaic panel multi-color space middle fusion feature map and a photovoltaic panel multi-color space deep fusion feature map;

[0038] An adaptive enhancement module is used to obtain an adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, an adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and an adaptive enhanced photovoltaic panel multi-color space deep fusion feature map by gated screening of the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map through a gated screening adaptive attention module;

[0039] The recognition result determination module is used to determine the recognition result of the dust accumulation degree based on the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map.

[0040] In some possible embodiments, the multi-scale image feature extraction module has

[0041] Body is used for:

[0042] Performing color space conversion mapping on the photovoltaic panel RGB image to obtain a photovoltaic panel YCbCr image;

[0043] The photovoltaic panel RGB image and the photovoltaic panel YCbCr image are respectively passed through the Xception feature extraction network to obtain a photovoltaic panel RGB shallow feature map, a photovoltaic panel RGB middle feature map, a photovoltaic panel RGB deep feature map, a photovoltaic panel YCbCr shallow feature map, a photovoltaic panel YCbCr middle feature map and a photovoltaic panel YCbCr deep feature map;

[0044] The photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr are respectively fused

[0045] Shallow feature map, the photovoltaic panel RGB middle feature map and the photovoltaic panel YCbCr

[0046] The middle-layer feature map, the photovoltaic panel RGB deep-layer feature map, and the photovoltaic panel YCbCr deep-layer feature map are used to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle-layer fusion feature map, and the photovoltaic panel multi-color space deep fusion feature map.

[0047] The photovoltaic panel dust accumulation status identification system and method based on machine learning of the embodiments of the present invention characterizes the dust status on the photovoltaic panel surface by analyzing the multi-color space and multi-scale color features of the photovoltaic panel surface image in combination with a machine learning algorithm, thereby automatically identifying the degree and level of dust accumulation to replace traditional manual inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 is a flow chart of a method for identifying dust accumulation status of photovoltaic panels based on machine learning according to an embodiment of the present invention;

[0050] Figure 2 A system architecture diagram of a photovoltaic panel dust accumulation state identification method based on machine learning according to an embodiment of the present invention;

[0051] Figure 3 This is a flowchart of sub-step S2 of the photovoltaic panel dust accumulation state identification method based on machine learning according to an embodiment of the present invention;

[0052] Figure 4 4 is a block diagram of a photovoltaic panel dust accumulation state identification system based on machine learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0054] Unless otherwise specified, the technical terms or scientific terms used in the embodiments of the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. "Including" or "comprising" used in the embodiments of the present invention neither limit the shapes, numbers, steps, actions, operations, components, originals and / or their groups mentioned, nor exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, originals and / or their groups, or the addition of these. In addition, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number and order of the indicated technical features. Thus, the features defined as "first" and "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0055] Unless otherwise specifically stated, the relative arrangement of the components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be understood that for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship, and the techniques, methods and devices known to ordinary technicians in the relevant fields may not be discussed in detail, but where appropriate, the techniques, methods and devices shown should be considered as part of the authorized specification. In all examples shown and discussed here, any specific other examples may have different values. It should be noted that similar symbols and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0056] In the description of the embodiments of the present invention, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in the embodiments of the present invention and the features of the different embodiments or examples, without contradiction.

[0057] Traditional methods for identifying dust accumulation status rely on manual inspection, which is inefficient, time-consuming, labor-intensive, and unable to respond in a timely manner. With the rapid development of machine learning technology, especially the breakthrough of deep learning in the field of image processing, a new solution is provided for the automatic identification of the dust accumulation status of photovoltaic panels.

[0058] In the technical solution of the present invention, a method for identifying the dust accumulation state of photovoltaic panels based on machine learning is proposed. Figure 1 The present invention is a flowchart of a method for identifying dust accumulation status of photovoltaic panels based on machine learning according to an embodiment of the present invention. Figure 2 1 is a system architecture diagram of a photovoltaic panel dust accumulation state identification method based on machine learning according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, according to an embodiment of the present invention, a method for identifying the dust accumulation state of a photovoltaic panel based on machine learning includes the following steps: S1, obtaining an RGB image of a photovoltaic panel collected by a camera; S2, performing multi-scale image feature extraction based on a multi-color space on the RGB image of the photovoltaic panel to obtain a shallow fusion feature map of a multi-color space of a photovoltaic panel, a middle fusion feature map of a multi-color space of a photovoltaic panel, and a deep fusion feature map of a multi-color space of a photovoltaic panel; S3, filtering the shallow fusion feature map of a multi-color space of a photovoltaic panel, the middle fusion feature map of a multi-color space of a photovoltaic panel, and the deep fusion feature map of a multi-color space of a photovoltaic panel through a gated screening adaptive attention module to obtain an adaptive enhanced photovoltaic panel shallow fusion feature map of a multi-color space of a photovoltaic panel, an adaptive enhanced photovoltaic panel middle fusion feature map of a multi-color space of a photovoltaic panel, and an adaptive enhanced photovoltaic panel deep fusion feature map of a multi-color space; S4, determining the recognition result of the dust accumulation degree based on the adaptive enhanced photovoltaic panel shallow fusion feature map of a multi-color space of a photovoltaic panel, the adaptive enhanced photovoltaic panel middle fusion feature map of a multi-color space of a photovoltaic panel, and the adaptive enhanced photovoltaic panel deep fusion feature map of a multi-color space of a photovoltaic panel.

[0059] In particular, the S1 obtains the RGB image of the photovoltaic panel collected by the camera. Here, the RGB image of the photovoltaic panel directly reflects the actual situation of the surface of the photovoltaic panel, including the dust coverage and the surface color change, which provides an intuitive basis for the subsequent dust state identification. At the same time, modern camera technology has been developed and mature, and high-quality RGB images can be easily obtained. Moreover, the camera is easy to install and deploy at different positions of the photovoltaic panel, and can be installed and adjusted according to actual needs. Therefore, it is a better choice to collect the RGB image of the photovoltaic panel by the camera as the basic data for subsequent analysis.

[0060] In particular, the S2 performs multi-scale image feature extraction based on multi-color space on the photovoltaic panel RGB image to obtain a photovoltaic panel multi-color space shallow fusion feature map, a photovoltaic panel multi-color space middle fusion feature map and a photovoltaic panel multi-color space deep fusion feature map. In particular, in a specific example of the present invention, as Figure 3As shown, the S2 includes: S21, performing color space conversion mapping on the photovoltaic panel RGB image to obtain a photovoltaic panel YCbCr image; S22, respectively passing the photovoltaic panel RGB image and the photovoltaic panel YCbCr image through the Xception feature extraction network to obtain a photovoltaic panel RGB shallow feature map, a photovoltaic panel RGB middle feature map, a photovoltaic panel RGB deep feature map, a photovoltaic panel YCbCr shallow feature map, a photovoltaic panel YCbCr middle feature map and a photovoltaic panel YCbCr deep feature map; S23, respectively fusing the photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel RGB middle feature map and the photovoltaic panel YCbCr middle feature map, and the photovoltaic panel RGB deep feature map and the photovoltaic panel YCbCr deep feature map to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map.

[0061] Specifically, the S21 performs color space conversion mapping on the photovoltaic panel RGB image to obtain a photovoltaic panel YCbCr image. In a specific example of the present invention, the photovoltaic panel RGB image is mapped to the YCbCr color space to obtain the photovoltaic panel YCbCr image. Considering that in existing application scenarios, the YCbCr color space divides color information into two parts: brightness (Y) and chrominance (Cb, Cr), where the Y component represents brightness information, and the Cb and Cr components represent the color difference information of blue and red, respectively. This separation of color information helps to more effectively analyze and process color information in image processing. In the technical solution of the present invention, the photovoltaic panel RGB image is mapped to the YCbCr color space to make full use of brightness information and chrominance information to increase the sensitivity to dust recognition, thereby obtaining a photovoltaic panel YCbCr image.

[0062] Specifically, the S22, the photovoltaic panel RGB image and the photovoltaic panel YCbCr image are respectively passed through the Xception feature extraction network to obtain the photovoltaic panel RGB shallow feature map, the photovoltaic panel RGB middle feature map, the photovoltaic panel RGB deep feature map, the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel YCbCr middle feature map and the photovoltaic panel YCbCr deep feature map. Among them, Xception is a deep learning architecture that designs separable convolutions to reduce the amount of model calculation while maintaining the ability of feature extraction to adapt to real-time or large-scale image processing tasks. More importantly, the Xception feature extraction network can extract features of different levels from the input image, such as shallow, middle and deep features, which usually correspond to different levels of abstraction of the image, which helps to identify multi-scale distribution patterns and spatial structures in the image. Specifically, in the photovoltaic panel RGB image, shallow features such as edges, textures, and simple shape features can describe the preliminary outline of dust and the grid structure of the photovoltaic panel; mid-level features may include slightly complex textures and local neighborhood spatial structures, such as the spatial distribution pattern of dust; and deep features usually include more advanced abstract information, such as the overall pattern of dust coverage and the overall appearance changes of the photovoltaic panel. The same is true in the photovoltaic panel YCbCr image, where shallow features may include brightness changes and rough color differences; mid-level features may include more detailed color distribution and chromaticity changes; deep features may include high-level abstract representations of color and brightness, such as local state changes of dust and photovoltaic panels and the impact of dust coverage on the overall color and brightness of photovoltaic panels. These multi-scale features are of great significance for identifying dust states.

[0063] Specifically, the S23 is to fuse the photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel RGB middle feature map and the photovoltaic panel YCbCr middle feature map, and the photovoltaic panel RGB deep feature map and the photovoltaic panel YCbCr deep feature map to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map. That is, by fusing the color feature information under the RGB color space and the YCbCr color space, a more comprehensive image representation is provided.

[0064] In a specific example of the present invention, the photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel RGB middle-layer feature map and the photovoltaic panel YCbCr middle-layer feature map, and the photovoltaic panel RGB deep feature map and the photovoltaic panel YCbCr deep feature map are fused respectively to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle-layer fusion feature map and the photovoltaic panel multi-color space deep fusion feature map, including: cascading the photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map to obtain the photovoltaic panel multi-color space shallow fusion feature map; cascading the photovoltaic panel RGB middle-layer feature map and the photovoltaic panel YCbCr middle-layer feature map to obtain the photovoltaic panel multi-color space middle-layer fusion feature map; cascading the photovoltaic panel RGB deep feature map and the photovoltaic panel YCbCr deep feature map to obtain the photovoltaic panel multi-color space deep fusion feature map.

[0065] It is worth mentioning that in other specific examples of the present invention, the photovoltaic panel RGB image can also be subjected to multi-scale image feature extraction based on multi-color space in other ways to obtain a photovoltaic panel multi-color space shallow fusion feature map, a photovoltaic panel multi-color space middle fusion feature map and a photovoltaic panel multi-color space deep fusion feature map, for example: input the photovoltaic panel RGB image; perform multi-scale feature extraction on the image of each color space, and use convolution kernels or pooling operations of different sizes to capture information of different scales; at the shallow level, fuse the low-level features (such as edges, textures, etc.) extracted from different color spaces, and use simple feature cascade or splicing operations; apply appropriate convolution operations and nonlinear activation functions are used to extract higher-level feature representations; in the middle layer, the features extracted from each color space are processed more deeply, which may include more complex convolution structures or attention mechanisms to capture more abstract features; the features extracted from the middle layers of each color space are fused, and attention mechanisms, feature cascades, and other methods may be used; in the deep layer, the features fused from the middle layers of each color space are abstracted and represented more deeply, which may include more convolution layers, residual connections, and the like; the features extracted from the deep layers of each color space are fused to obtain the shallow fusion feature map of the photovoltaic panel multi-color space, the middle fusion feature map of the photovoltaic panel multi-color space, and the deep fusion feature map of the photovoltaic panel multi-color space.

[0066] In particular, in S3, the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map are gated and screened by an adaptive attention module to obtain an adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, an adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and an adaptive enhanced photovoltaic panel multi-color space deep fusion feature map. Considering that in the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map, not all channel features are equally important to the final target task, that is, dust state recognition. Specifically, if too much background or unimportant information is mixed in the color feature distribution pattern at different scales, such as the original photovoltaic panel RGB image may contain background or other objects other than the photovoltaic panel, these areas are not helpful for identifying the dust state on the photovoltaic panel, that is, these information will have an adverse effect on the feature expression and dust state recognition of the photovoltaic panel surface to a certain extent. To this end, in the technical solution of the present invention, the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map are further subjected to gated screening of the adaptive attention module to obtain the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map. In this way, the features most relevant to the dust state can be adaptively highlighted through the gated screening adaptive attention module.

[0067] More specifically, in a specific example of the present invention, the implementation process of obtaining an adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map by gated screening of the photovoltaic panel multi-color space shallow fusion feature map through an adaptive attention module includes: first, performing global pooling processing based on the maximum value, global pooling processing based on random values, and global pooling processing based on average values ​​on the photovoltaic panel multi-color space shallow fusion feature map to obtain a photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, a photovoltaic panel multi-color space shallow fusion global random value pooling feature vector, and a photovoltaic panel multi-color space shallow fusion global mean pooling feature vector. Here, in the technical solution of the present invention, different types of pooling operations are used to capture the expression of the diverse global feature information of the photovoltaic panel multi-color space shallow fusion feature map. Among them, the global pooling based on the maximum value can capture the maximum response in the local area, which may correspond to the significant features of each feature matrix along the channel dimension in the photovoltaic panel multi-color space shallow fusion feature map; while the global pooling based on the average value provides the average information of the local area, which can reflect the overall feature representation of the local area; in addition, the global pooling processing based on the average value can randomly select elements in the local area according to the probability value, which can avoid the excessive influence of interference values ​​on the representation of global feature information and help to retain the diversity features in the photovoltaic panel multi-color space shallow fusion feature map. Then, the photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, the photovoltaic panel multi-color space shallow fusion global mean pooling feature vector and the photovoltaic panel multi-color space shallow fusion global random value pooling feature vector are fused to obtain the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector.

[0068] In a specific example, the photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, the photovoltaic panel multi-color space shallow fusion global mean pooling feature vector and the photovoltaic panel multi-color space shallow fusion global random value pooling feature vector are fused to obtain the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector, including: determining a first photovoltaic panel multi-color space shallow fusion semantic pooling representation factor, a second photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and a third photovoltaic panel multi-color space shallow fusion semantic pooling representation factor; calculating the first The photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the photovoltaic panel multi-color space shallow fusion global mean pooling feature vector, the second photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, and the third photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the photovoltaic panel multi-color space shallow fusion global random value pooling feature vector are added element by element in position to obtain the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector. Then, the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector is subjected to channel-by-channel semantic feature interaction and feature activation to obtain the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector.

[0069] In a specific example, the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector is subjected to channel-by-channel semantic feature interaction and feature activation to obtain the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector, including: passing the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector through a fully connected layer to perform channel-by-channel semantic feature interaction and feature activation on the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector to obtain the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector. Here, the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector is subjected to channel-by-channel feature interaction and feature activation by using a fully connected layer, which helps to further extract and strengthen the correlation and importance between features. Subsequently, the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector is input into a gating unit to screen the features and retain the features with high relevance to the target task, thereby obtaining a gated screened photovoltaic panel multi-color space shallow fusion channel semantic association feature vector. After that, the gated screening photovoltaic panel multi-color space shallow fusion channel semantic association feature vector is normalized to ensure that each feature value is on the same scale, thereby obtaining the gated screening photovoltaic panel multi-color space shallow fusion channel semantic association weight feature vector. Finally, the gated screening photovoltaic panel multi-color space shallow fusion channel semantic association weight feature vector is used as the weight feature vector, and its channel-by-channel point multiplication with the photovoltaic panel multi-color space shallow fusion feature map is calculated to achieve adaptive enhancement guided by the feature distribution pattern of the photovoltaic panel multi-color space shallow fusion feature map itself, thereby obtaining the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map.

[0070] In summary, in the above embodiment, the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle-layer fusion feature map and the photovoltaic panel multi-color space deep fusion feature map are subjected to gated screening of the adaptive attention module to obtain the adaptively enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptively enhanced photovoltaic panel multi-color space middle-layer fusion feature map and the adaptively enhanced photovoltaic panel multi-color space deep fusion feature map, including: using the gated screening adaptive attention module to adaptively enhance the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle-layer fusion feature map and the photovoltaic panel multi-color space deep fusion feature map with the following adaptive enhancement formula to obtain the adaptively enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptively enhanced photovoltaic panel multi-color space middle-layer fusion feature map and the adaptively enhanced photovoltaic panel multi-color space deep fusion feature map; wherein the formula is:

[0071] V c =α·AvgPool(F)+β·MaxPool(F)+γ·StoPool(F)

[0072] A=σ(W a *V c +B a )

[0073] A m =mask(A)

[0074]

[0075] F′=A′⊙F

[0076] Wherein, F is the photovoltaic panel multi-color space shallow fusion feature map, AvgPool, MaxPool and StoPool represent global pooling processing based on average value, global pooling processing based on maximum value and global pooling processing based on random value respectively, α, β and γ are the first photovoltaic panel multi-color space shallow fusion semantic pooling representation factor, the second photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the third photovoltaic panel multi-color space shallow fusion semantic pooling representation factor respectively, V c is the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector, W a and B a are the weight and bias of the fully connected layer, σ is the sigmoid activation function, A is the semantic association feature vector of the shallow fusion channel of the photovoltaic panel multi-color space, mask is the masking process of the gated unit, and A m is the gated screening photovoltaic panel multi-color space shallow fusion channel semantic association feature vector, A mi is the i-th eigenvalue in the semantic association feature vector of the shallow fusion channel of multi-color space of the gated screening photovoltaic panel, A′ is the semantic association weight feature vector of the shallow fusion channel of multi-color space of the gated screening photovoltaic panel, ⊙ represents the channel-by-channel dot product of the feature vector and the feature map, and F′ is the shallow fusion feature map of multi-color space of the adaptive enhanced photovoltaic panel.

[0077] In particular, the S4 determines the recognition result of the dust accumulation level based on the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map. In particular, in a specific example of the present invention, after the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map are cascaded into a photovoltaic panel multi-color space multi-scale fusion feature map, the photovoltaic panel multi-color space multi-scale fusion feature map is passed through a classifier-based dust accumulation state recognition module to obtain a recognition result, and the recognition result is used to represent the dust accumulation level label. That is, the multi-scale expression of dust state features is integrated and interacted in a cascaded manner to provide a more comprehensive observation perspective. The classifier is then used to construct a dust accumulation state recognition module to map the photovoltaic panel multi-color space multi-scale fusion feature map to a determined dust accumulation level label, thereby providing technical support for determining the priority and resource allocation of cleaning work.

[0078] Here, considering the misalignment of image semantic feature distributions of the cross-scale and cross-depth image semantic feature representations of the photovoltaic panel RGB image and the photovoltaic panel YCbCr image among the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map, the photovoltaic panel multi-color space multi-scale fusion feature map obtained by further gated screening of adaptive attention enhancement and feature fusion will also have image semantic distribution knowledge offset relative to the semantic distribution of the source image, thereby causing uncertainty in its image semantic feature distribution relative to the class probability understanding of the classifier, reducing the speed of classification training and the accuracy of classification results.

[0079] In a preferred embodiment, the photovoltaic panel multi-color space multi-scale fusion feature map is input into a dust accumulation state recognition module based on a classifier to obtain a recognition result, including:

[0080] Determine n probability values ​​corresponding to n dust accumulation level labels respectively obtained by inputting the multi-color space multi-scale fusion feature map of the photovoltaic panel into a dust accumulation state recognition module based on a classifier;

[0081] Calculate the binary norm of the probability distribution vector composed of the n probability values, divide the binary norm by the square root of n to obtain a classification probability value, and subtract the classification probability value from one to obtain a classification probability difference;

[0082] Calculate each eigenvalue of the photovoltaic panel multi-color space multi-scale fusion feature map respectively, using the classification probability value as an exponent of the power function to obtain a photovoltaic panel multi-color space multi-scale fusion class feature map;

[0083] Calculate each eigenvalue of the photovoltaic panel multi-color space multi-scale fusion feature map respectively, using the classification probability difference as an exponent of the power function to obtain the photovoltaic panel multi-color space multi-scale fusion class difference feature map;

[0084] The classification probability value is point-multiplied with the photovoltaic panel multi-color space multi-scale fusion class difference feature map to obtain a first photovoltaic panel multi-color space multi-scale fusion intermediate class feature map, and the classification probability difference value is point-multiplied with the photovoltaic panel multi-color space multi-scale fusion class difference feature map to obtain a second photovoltaic panel multi-color space multi-scale fusion intermediate class feature map;

[0085] After performing a dot product of the first photovoltaic panel multi-color space multi-scale fusion intermediate class feature map and the second photovoltaic panel multi-color space multi-scale fusion intermediate class feature map, further perform a dot addition with the classification probability value and the dot product result of the photovoltaic panel multi-color space multi-scale fusion feature map to obtain an optimized photovoltaic panel multi-color space multi-scale fusion feature map;

[0086] The optimized photovoltaic panel multi-color space multi-scale fusion feature map is input into the classifier-based dust accumulation state recognition module to obtain a recognition result.

[0087] Based on this, when the inventor of the present invention inputs the photovoltaic panel multi-color space multi-scale fusion feature map into the classifier-based dust accumulation state recognition module for classification, in order to achieve unsupervised domain adaptation from the feature distribution domain of the photovoltaic panel multi-color space multi-scale fusion feature map to the probability distribution domain of the classification probability, the class probability joint representation under each risk level label obtained by the classifier-based dust accumulation state recognition module of the photovoltaic panel multi-color space multi-scale fusion feature map is used as a domain agent, and the moving average of the probability distribution of the photovoltaic panel multi-color space multi-scale fusion feature map is performed through the interaction of the class distribution of the power function as an exponent, and the knowledge transfer from the unlabeled features to be classified to the labeled probability distribution is achieved by superimposing the feature domain probability distribution knowledge of the photovoltaic panel multi-color space multi-scale fusion feature map itself, thereby promoting the classification operation of the photovoltaic panel multi-color space multi-scale fusion feature map, that is, improving the speed of classification training and the accuracy of classification results.

[0088] In summary, a method for identifying the dust accumulation status of photovoltaic panels based on machine learning according to an embodiment of the present invention is explained, which characterizes the dust status on the surface of the photovoltaic panels by combining machine learning algorithms to analyze the multi-color space and multi-scale color features of the photovoltaic panel surface image, thereby automatically identifying the degree and level of dust accumulation to replace traditional manual inspection.

[0089] Furthermore, a photovoltaic panel dust accumulation status identification system based on machine learning is also provided.

[0090] Figure 4 FIG. 1 is a block diagram of a photovoltaic panel dust accumulation state recognition system based on machine learning according to an embodiment of the present invention. Figure 4 As shown, according to an embodiment of the present invention, a photovoltaic panel dust accumulation state recognition system 300 based on machine learning includes: an image acquisition module 310, which is used to obtain an RGB image of a photovoltaic panel acquired by a camera; a multi-scale image feature extraction module 320, which is used to perform multi-scale image feature extraction based on a multi-color space on the RGB image of the photovoltaic panel to obtain a photovoltaic panel multi-color space shallow fusion feature map, a photovoltaic panel multi-color space middle fusion feature map and a photovoltaic panel multi-color space deep fusion feature map; an adaptive enhancement module 330, which is used to extract the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map. The middle-layer fusion feature map of the color space and the deep-layer fusion feature map of the photovoltaic panel multi-color space are gated and screened by an adaptive attention module to obtain an adaptive enhanced photovoltaic panel multi-color space shallow-layer fusion feature map, an adaptive enhanced photovoltaic panel multi-color space middle-layer fusion feature map and an adaptive enhanced photovoltaic panel multi-color space deep-layer fusion feature map; a recognition result determination module 340 is used to determine the recognition result of the dust accumulation level based on the adaptive enhanced photovoltaic panel multi-color space shallow-layer fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle-layer fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep-layer fusion feature map.

[0091] Specifically, the multi-scale image feature extraction module 320 is specifically used to:

[0092] The photovoltaic panel RGB image is subjected to color space conversion mapping to obtain a photovoltaic panel YCbCr image; the photovoltaic panel RGB image and the photovoltaic panel YCbCr image are respectively passed through the Xception feature extraction network to obtain a photovoltaic panel RGB shallow feature map, a photovoltaic panel RGB middle feature map, a photovoltaic panel RGB deep feature map, a photovoltaic panel YCbCr shallow feature map, a photovoltaic panel YCbCr middle feature map and a photovoltaic panel YCbCr deep feature map; the photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel RGB middle feature map and the photovoltaic panel YCbCr middle feature map, and the photovoltaic panel RGB deep feature map and the photovoltaic panel YCbCr deep feature map are respectively fused to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map.

[0093] As described above, the photovoltaic panel dust accumulation state recognition system 300 based on machine learning according to the embodiment of the present invention can be implemented in various wireless terminals, such as a server with a photovoltaic panel dust accumulation state recognition algorithm based on machine learning. In a possible implementation, the photovoltaic panel dust accumulation state recognition system 300 based on machine learning according to the embodiment of the present invention can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the photovoltaic panel dust accumulation state recognition system 300 based on machine learning can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the photovoltaic panel dust accumulation state recognition system 300 based on machine learning can also be one of the many hardware modules of the wireless terminal.

[0094] Alternatively, in another example, the photovoltaic panel dust accumulation status identification system 300 based on machine learning and the wireless terminal may also be separate devices, and the photovoltaic panel dust accumulation status identification system 300 based on machine learning may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0095] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for identifying the dust accumulation state of photovoltaic panels based on machine learning, characterized in that: The method comprises: Get the RGB image of the photovoltaic panel captured by the camera; Performing multi-scale image feature extraction based on multi-color space on the photovoltaic panel RGB image to obtain a photovoltaic panel multi-color space shallow fusion feature map, a photovoltaic panel multi-color space middle fusion feature map and a photovoltaic panel multi-color space deep fusion feature map; The photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map are gated and screened by an adaptive attention module to obtain an adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, an adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and an adaptive enhanced photovoltaic panel multi-color space deep fusion feature map; Based on the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map, the recognition result of the dust accumulation degree level is determined.

2. The photovoltaic panel dust accumulation state identification method based on machine learning according to claim 1 is characterized in that: The multi-scale image feature extraction based on multi-color space is performed on the photovoltaic panel RGB image to obtain a photovoltaic panel multi-color space shallow fusion feature map, a photovoltaic panel multi-color space middle fusion feature map and a photovoltaic panel multi-color space deep fusion feature map, including: Performing color space conversion mapping on the photovoltaic panel RGB image to obtain a photovoltaic panel YCbCr image; The photovoltaic panel RGB image and the photovoltaic panel YCbCr image are respectively passed through the Xception feature extraction network to obtain a photovoltaic panel RGB shallow feature map, a photovoltaic panel RGB middle feature map, a photovoltaic panel RGB deep feature map, a photovoltaic panel YCbCr shallow feature map, a photovoltaic panel YCbCr middle feature map and a photovoltaic panel YCbCr deep feature map; The photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel RGB middle layer feature map and the photovoltaic panel YCbCr middle layer feature map, and the photovoltaic panel RGB deep layer feature map and the photovoltaic panel YCbCr deep layer feature map are fused respectively to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle layer fusion feature map and the photovoltaic panel multi-color space deep fusion feature map.

3. The photovoltaic panel dust accumulation state identification method based on machine learning according to claim 2 is characterized in that: The step of performing color space conversion mapping on the photovoltaic panel RGB image to obtain a photovoltaic panel YCbCr image includes: The photovoltaic panel RGB image is mapped to a YCbCr color space to obtain the photovoltaic panel YCbCr image.

4. The photovoltaic panel dust accumulation state identification method based on machine learning according to claim 3 is characterized in that: The method of respectively fusing the photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel RGB middle feature map and the photovoltaic panel YCbCr middle feature map, and the photovoltaic panel RGB deep feature map and the photovoltaic panel YCbCr deep feature map to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map, includes: Cascading the photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map to obtain the photovoltaic panel multi-color space shallow fusion feature map; Cascading the photovoltaic panel RGB middle-layer feature map and the photovoltaic panel YCbCr middle-layer feature map to obtain the photovoltaic panel multi-color space middle-layer fusion feature map; The photovoltaic panel RGB deep feature map and the photovoltaic panel YCbCr deep feature map are cascaded to obtain the photovoltaic panel multi-color space deep fusion feature map.

5. The method for identifying dust accumulation status of photovoltaic panels based on machine learning according to claim 4 is characterized in that: The photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map are gated and screened by an adaptive attention module to obtain an adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, an adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and an adaptive enhanced photovoltaic panel multi-color space deep fusion feature map, including: Performing maximum-based global pooling processing, random-value-based global pooling processing, and average-based global pooling processing on the photovoltaic panel multi-color space shallow fusion feature map to obtain a photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, a photovoltaic panel multi-color space shallow fusion global random value pooling feature vector, and a photovoltaic panel multi-color space shallow fusion global mean pooling feature vector; Fusing the photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, the photovoltaic panel multi-color space shallow fusion global mean pooling feature vector and the photovoltaic panel multi-color space shallow fusion global random value pooling feature vector to obtain a photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector; Performing channel-by-channel semantic feature interaction and feature activation on the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector to obtain a photovoltaic panel multi-color space shallow fusion channel semantic association feature vector; Inputting the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector into a gating unit to obtain a gated screening photovoltaic panel multi-color space shallow fusion channel semantic association feature vector; Normalizing the gated screening photovoltaic panel multi-color space shallow fusion channel semantic association feature vector to obtain the gated screening photovoltaic panel multi-color space shallow fusion channel semantic association weight feature vector; The gated screening photovoltaic panel multi-color space shallow fusion channel semantic association weight feature vector is used as the weight feature vector, and its channel-by-channel point multiplication with the photovoltaic panel multi-color space shallow fusion feature map is calculated to obtain the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map.

6. The photovoltaic panel dust accumulation state identification method based on machine learning according to claim 5 is characterized in that: The fusion of the photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, the photovoltaic panel multi-color space shallow fusion global mean pooling feature vector and the photovoltaic panel multi-color space shallow fusion global random value pooling feature vector to obtain the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector includes: Determine a first photovoltaic panel multi-color space shallow fusion semantic pooling representation factor, a second photovoltaic panel multi-color space shallow fusion semantic pooling representation factor, and a third photovoltaic panel multi-color space shallow fusion semantic pooling representation factor; The photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector is obtained by calculating the dot product of the first photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the photovoltaic panel multi-color space shallow fusion global mean pooling feature vector, the dot product of the second photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the photovoltaic panel multi-color space shallow fusion global maximum pooling feature vector, and the dot product of the third photovoltaic panel multi-color space shallow fusion semantic pooling representation factor and the photovoltaic panel multi-color space shallow fusion global random value pooling feature vector.

7. The method for identifying dust accumulation status of photovoltaic panels based on machine learning according to claim 6, characterized in that: The step of performing channel-by-channel semantic feature interaction and feature activation on the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector to obtain the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector includes: The photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector is passed through a fully connected layer to perform channel-by-channel semantic feature interaction and feature activation on the photovoltaic panel multi-color space shallow fusion global semantic multi-scale representation vector to obtain the photovoltaic panel multi-color space shallow fusion channel semantic association feature vector.

8. The method for identifying dust accumulation status of photovoltaic panels based on machine learning according to claim 7, characterized in that: The method of determining the recognition result of the dust accumulation level based on the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map comprises: Cascading the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map into a photovoltaic panel multi-color space multi-scale fusion feature map; The photovoltaic panel multi-color space multi-scale fusion feature map is passed through a dust accumulation state recognition module based on a classifier to obtain the recognition result, and the recognition result is used to represent the dust accumulation degree level label.

9. A photovoltaic panel dust accumulation state recognition system based on machine learning, characterized in that: The system comprises: An image acquisition module, used to acquire the RGB image of the photovoltaic panel acquired by the camera; A multi-scale image feature extraction module is used to perform multi-scale image feature extraction based on a multi-color space on the photovoltaic panel RGB image to obtain a photovoltaic panel multi-color space shallow fusion feature map, a photovoltaic panel multi-color space middle fusion feature map and a photovoltaic panel multi-color space deep fusion feature map; An adaptive enhancement module is used to obtain an adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, an adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and an adaptive enhanced photovoltaic panel multi-color space deep fusion feature map by gated screening of the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle fusion feature map and the photovoltaic panel multi-color space deep fusion feature map through a gated screening adaptive attention module; The recognition result determination module is used to determine the recognition result of the dust accumulation degree based on the adaptive enhanced photovoltaic panel multi-color space shallow fusion feature map, the adaptive enhanced photovoltaic panel multi-color space middle fusion feature map and the adaptive enhanced photovoltaic panel multi-color space deep fusion feature map.

10. The photovoltaic panel dust accumulation state recognition system based on machine learning according to claim 9 is characterized in that: The multi-scale image feature extraction module is specifically used for: Performing color space conversion mapping on the photovoltaic panel RGB image to obtain a photovoltaic panel YCbCr image; The photovoltaic panel RGB image and the photovoltaic panel YCbCr image are respectively passed through the Xception feature extraction network to obtain a photovoltaic panel RGB shallow feature map, a photovoltaic panel RGB middle feature map, a photovoltaic panel RGB deep feature map, a photovoltaic panel YCbCr shallow feature map, a photovoltaic panel YCbCr middle feature map and a photovoltaic panel YCbCr deep feature map; The photovoltaic panel RGB shallow feature map and the photovoltaic panel YCbCr shallow feature map, the photovoltaic panel RGB middle layer feature map and the photovoltaic panel YCbCr middle layer feature map, and the photovoltaic panel RGB deep layer feature map and the photovoltaic panel YCbCr deep layer feature map are fused respectively to obtain the photovoltaic panel multi-color space shallow fusion feature map, the photovoltaic panel multi-color space middle layer fusion feature map and the photovoltaic panel multi-color space deep fusion feature map.

Images