Photovoltaic panel dusting visualization detection system based on deep learning and cloud monitoring platform

Abstract

The application belongs to the technical field of photovoltaic power generation, and discloses a photovoltaic panel dusting visual detection system and a cloud monitoring platform based on deep learning, which comprises a data acquisition and preprocessing module, a model construction module, a model training and evaluation module and a detection application module. The data acquisition and preprocessing module is used for acquiring clean photovoltaic panel images, and performing image correction, silver line removal and dusted picture synthesis processing to generate a training test data set. The model construction module is used for constructing a photovoltaic dusting detection model. The model training and evaluation module is used for training the photovoltaic dusting detection model based on the training test data set, setting evaluation indexes to evaluate the trained photovoltaic dusting detection model, and obtaining an optimal photovoltaic dusting detection model. The detection application module is used for inputting photovoltaic dusting images in actual working conditions into the optimal photovoltaic dusting detection model to monitor the dusting condition of the photovoltaic panel in real time. The application can accurately detect dust on the surface of the photovoltaic panel and realize dust distribution visualization, and realizes real-time monitoring and analysis of data through the cloud platform.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation technology, and in particular relates to a deep learning-based visualization detection system for dust accumulation on photovoltaic panels and a cloud monitoring platform. Background Technology

[0002] Photovoltaic panels are typically exposed to the outdoors, where the atmosphere is filled with dust particles. In practical applications, dust easily accumulates on the surface of photovoltaic panels, which can affect the power generation efficiency and safety of the modules to varying degrees. First, a dust layer covering the panel obstructs sunlight, reducing the light transmittance of the glass surface and thus decreasing the amount of solar radiation received by the panel. Second, dust absorbs heat, converting it into its own thermal energy, further affecting the heat dissipation performance of the glass covering the photovoltaic panel, thereby impacting overall operating efficiency.

[0003] The existing photovoltaic dust accumulation detection technologies include the following:

[0004] Direct observation: This method is the most direct, but it requires a lot of human intervention and is difficult to carry out in bad weather or at night. Furthermore, human observation is easily affected by subjective factors and cannot accurately judge subtle changes in dust.

[0005] Laser scanning method: This method can quickly and accurately obtain the dust coverage of photovoltaic panels, but it requires the use of high-precision laser scanning instruments, which is costly and has high requirements for equipment.

[0006] Image processing method: This method can detect dust coverage by processing and analyzing the surface image of the photovoltaic panel, but it is easily affected by factors such as light and shadow, and the processing speed is relatively slow.

[0007] In summary, existing technologies for detecting dust accumulation on photovoltaic panels suffer from drawbacks such as high equipment requirements, poor analytical accuracy, and low efficiency. Therefore, there is an urgent need to propose a deep learning-based visual detection system and cloud monitoring platform for detecting dust accumulation on photovoltaic panels. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a deep learning-based photovoltaic panel dust accumulation visualization detection system and cloud monitoring platform. This system is used to accurately detect dust on the surface of photovoltaic panels and visualize the dust distribution. Simultaneously, it enables real-time monitoring and analysis of data through a cloud platform, thereby solving the problems existing in the prior art.

[0009] To achieve the above objectives, the present invention provides a deep learning-based photovoltaic panel dust accumulation visualization detection system, comprising: a data acquisition and preprocessing module, a model building module, a model training and evaluation module, and a detection application module connected in sequence;

[0010] The data acquisition and preprocessing module is used to acquire images of clean photovoltaic panels, and to perform image correction, silver line removal, and composite dust accumulation images to generate training and testing datasets.

[0011] The model building module is used to build a photovoltaic dust accumulation detection model, which includes a feature extraction layer, four residual blocks, and a fully connected layer.

[0012] The model training and evaluation module is used to train the photovoltaic dust accumulation detection model based on the training and testing dataset, and to set evaluation indicators to evaluate the trained photovoltaic dust accumulation detection model in order to obtain the optimal photovoltaic dust accumulation detection model.

[0013] The detection application module is used to input the photovoltaic dust accumulation image in actual working conditions into the optimal photovoltaic dust accumulation detection model. After preprocessing, transmittance calculation and image postprocessing, it outputs the corresponding medium transmission map and average transmittance to monitor the dust accumulation of photovoltaic panels in real time.

[0014] Optionally, the data acquisition and preprocessing module includes an image correction unit, a silver line removal unit, and a gray image synthesis unit;

[0015] The image correction unit is used to select four corresponding points on the source image and target image of the clean photovoltaic panel, respectively, to construct and solve the perspective transformation equation system to achieve image correction.

[0016] The silver line removal unit is used to convert the corrected image from RGB color space to HSV color space, create a binary mask to identify the silver line based on the color characteristics of the silver line in HSV space, and then fill in the missing part based on morphological dilation operation and Navier-Stokes algorithm to remove the silver line.

[0017] The synthesized gray image unit is used to segment the clean photovoltaic panel image into gray-free blocks of preset pixels, and randomly generate transmittance. Based on the gray-free blocks, gray blocks are calculated through an atmospheric scattering model, thereby generating images with different degrees of gray accumulation and their corresponding transmittance data labels.

[0018] Optionally, the feature extraction layer employs multi-scale convolution with kernel sizes of 3×3, 5×5, and 7×7, and each scale uses a 32-channel filter.

[0019] The multi-scale convolution calculation process of the feature extraction layer includes: first performing a convolution operation, then sequentially performing batch normalization and ReLU activation function processing, and finally adding the channel outputs corresponding to the three different scale convolution kernels to obtain the output of the feature extraction layer.

[0020] Optionally, each residual block includes two 3×3 convolutional kernels, each followed by batch normalization and ReLU activation functions; each residual block also includes skip connections, residual mappings, and SEBlock modules;

[0021] The SEBlock module is used to adaptively recalibrate the feature responses across channels through Squeeze and Excitation operations.

[0022] Optionally, the fully connected layer includes an average pooling layer, two linear transformations, and a ReLU activation function;

[0023] The fully connected layer is used to compress the feature map extracted from the feature extraction layer and residual block into a scalar based on the average pooling layer. After two linear transformations and ReLU activation function processing, the average transmittance and medium transport map are finally output.

[0024] Optionally, the model training and evaluation module is used to set the parameters for model training, select the optimizer and loss function, and use the training and testing datasets to train and evaluate the model. Then, the model performance is quantified by mean squared error, mean absolute error and relative error, and the model with the smallest loss during training is saved as the optimal photovoltaic dust accumulation detection model.

[0025] The present invention also provides a cloud monitoring platform for deploying the aforementioned deep learning-based photovoltaic panel dust accumulation visualization detection system, comprising: a data transmission module, a data processing and analysis module, an early warning module, and a remote monitoring module;

[0026] The data transmission module is used to connect the detection equipment to the cloud monitoring platform through Internet of Things (IoT) technology to achieve real-time data transmission and storage.

[0027] The data processing and analysis module is equipped with the deep learning-based photovoltaic panel dust accumulation visualization detection system, which is used to process and analyze the collected data, assess the degree of dust accumulation on the photovoltaic panel, and generate dust distribution maps and trend maps.

[0028] The early warning module is used to automatically issue an early warning signal when the dust accumulation exceeds a set threshold.

[0029] The remote monitoring module is used to view the dust accumulation on photovoltaic panels and the system's operating status anytime, anywhere through the web or mobile application of the cloud monitoring platform.

[0030] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the system.

[0031] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the system described herein.

[0032] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the system described above.

[0033] Compared with the prior art, the present invention has the following advantages and technical effects:

[0034] 1. Existing technical solutions do not quantitatively assess the degree of dust accumulation in photovoltaic strings, while this invention can quantitatively assess the degree and specific distribution of dust accumulation on photovoltaic panels;

[0035] 2. Existing technical solutions utilize actual site resources to select actual control groups to measure the degree of dust accumulation in components, which is relatively complex and limits remote diagnostic methods. This invention utilizes deep learning methods to achieve remote analysis, which has good portability.

[0036] 3. Existing technologies utilize deep learning to identify the degree of dust accumulation on photovoltaic panels, requiring a large amount of labeled data for training. However, in real-world applications, obtaining sufficient labeled data, especially photovoltaic panel image data containing dust distribution information, is often very difficult. This reliance on large amounts of data not only increases the time and cost of data preparation but also, due to data scarcity, can lead to a decline in model performance and affect its generalization ability. This invention can obtain training set data by adding grayscale to clean images, reducing the dependence on large-scale labeled data.

[0037] 4. Existing technical solutions have limited monitoring range and cannot monitor in real time. This invention uses a cloud platform to achieve large-area, real-time monitoring, providing remote monitoring and early warning functions, and improving the efficiency of operation and maintenance management. Attached Figure Description

[0038] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0039] Figure 1 This is a flowchart illustrating the specific implementation of the monitoring system according to an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of the dataset for the photovoltaic dust accumulation detection model in an embodiment of the present invention;

[0041] Figure 3 This is a network structure diagram of the photovoltaic dust accumulation detection model according to an embodiment of the present invention;

[0042] Figure 4These are the input and output images of the photovoltaic dust accumulation detection model in this embodiment of the invention, where (a) and (e) are the original dust accumulation images, (b) and (f) are the images after adding contrast, (c) and (g) are the media transport images estimated by the model, and (d) and (h) are the result images after Gaussian smoothing of the estimated media transport images. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0045] Example 1

[0046] Photovoltaic panels are used outdoors, where the atmosphere is filled with dust particles of various shapes. In practical engineering, dust easily accumulates on the surface of photovoltaic panels, and this dust affects the power generation performance and safety of the panels to varying degrees. On the one hand, dust on the surface of photovoltaic panels blocks sunlight, reduces the transmittance of the glass, and decreases solar radiation on the surface of the cells; on the other hand, dust absorbs solar radiation and converts it into its own heat energy, hindering external heat dissipation from the photovoltaic panel's cover glass. Many factors influence dust accumulation on photovoltaic panels. While current dust accumulation detection methods can achieve real-time monitoring and are cost-effective, they struggle to achieve satisfactory detection results when faced with complex dust accumulation patterns and challenging environmental conditions.

[0047] like Figure 1 As shown, to address the aforementioned problems, this embodiment provides a deep learning-based photovoltaic panel dust accumulation visualization detection system for accurately detecting dust on the surface of photovoltaic panels and visualizing dust distribution. The model employs a lightweight design, reducing the complexity of training, optimization, and maintenance, and enabling real-time monitoring and early warning functions. The system specifically includes: a data acquisition and preprocessing module, a model building module, a model training and evaluation module, and a detection application module connected sequentially.

[0048] The data acquisition and preprocessing module is used to acquire images of clean photovoltaic panels, and to perform image correction, silver line removal, and composite dust accumulation images to generate training and testing datasets.

[0049] The model building module is used to build a photovoltaic dust accumulation detection model, which includes a feature extraction layer, four residual blocks, and a fully connected layer.

[0050] The model training and evaluation module is used to train the photovoltaic dust accumulation detection model based on the training and testing dataset, and to set evaluation indicators to evaluate the trained photovoltaic dust accumulation detection model in order to obtain the optimal photovoltaic dust accumulation detection model.

[0051] The detection application module is used to input the photovoltaic dust accumulation image in actual working conditions into the optimal photovoltaic dust accumulation detection model. After preprocessing, transmittance calculation and image postprocessing, it outputs the corresponding medium transmission map and average transmittance to monitor the dust accumulation of photovoltaic panels in real time.

[0052] As an implementable approach, the specific process includes the following:

[0053] Step S21: The data acquisition and preprocessing module is used to acquire visible light images of clean photovoltaic panels, perform image correction, remove silver lines, and synthesize a dust accumulation image based on the principle of photovoltaic panel light attenuation, which serves as the training and testing dataset for the model provided in this embodiment.

[0054] The data acquisition and preprocessing module, as implemented, includes an image correction unit, a silver line removal unit, and a gray image synthesis unit. The image correction unit selects four corresponding points on the source and target images of the clean photovoltaic panel, constructs and solves a set of perspective transformation equations to achieve image correction. The silver line removal unit converts the corrected image from the RGB color space to the HSV color space, creates a binary mask to identify the silver lines based on their color characteristics in the HSV space, and then fills in the missing parts based on morphological dilation and the Navier-Stokes algorithm to remove the silver lines. The gray image synthesis unit segments the clean photovoltaic panel image into gray-free blocks of preset pixels, randomly generates transmittance, and calculates gray blocks based on the gray-free blocks using an atmospheric scattering model, thereby generating images with different degrees of gray accumulation and their corresponding transmittance data labels.

[0055] Specifically, it includes:

[0056] Step S211: Use a TIR-10 camera manufactured by Pinling Company to acquire images of the cleaned photovoltaic panels. The TIR-10 camera has 2 megapixels, 10x optical autofocus, and outputs images with a resolution of 1920×1080. It transmits visible light images in real time via a network interface.

[0057] Step S212: Perform image preprocessing, including image correction and removal of silver lines:

[0058] (1) Image correction employs perspective transformation. By selecting four corresponding points on the source and target images, a system of perspective transformation equations containing eight unknowns is constructed and solved to obtain the transformation parameters, thereby correcting image distortion. Specifically:

[0059] Perspective transformation is used to project an image onto a new viewing plane to correct the image.

[0060]

[0061] (X,Y,Z) are the original image plane coordinates, and the corresponding transformed image plane coordinates are (X0,Y0,Z0).

[0062] Since we are dealing with a two-dimensional image, we can set z = 1 and divide the transformed image coordinates by Z to reduce the image from three dimensions to two dimensions, thus obtaining the following equation:

[0063]

[0064] Generally, in this embodiment, let m 33 =1, to easily obtain x and y, so that the denominator on the left side of equation (3) is 1, expand the above formula to get the case of a point:

[0065]

[0066] There are a total of 8 unknowns (m) in equation (3). i j) To solve for the unknown, eight sets of equations need to be set up, that is, four points are manually selected on the source image and the target image respectively, usually the four vertices of the image.

[0067] Select four coordinate points on the source image, namely A: (m0,n0), (m1,n1), (m2,n2), (m3,n3).

[0068] Select four coordinate points on the target image, namely B: (M0′,N0′), (M1′,N1′), (M2′,N2′), (M3′,N3′).

[0069] Substituting equation (3), we can obtain equation (4), as follows:

[0070]

[0071] (2) Remove the silver lines by creating a binary mask and using morphological methods:

[0072] Silver lines serve as electrodes or connecting wires in photovoltaic panels. During image feature extraction, these silver lines may be mistakenly identified as important features, interfering with true feature recognition. Furthermore, silver lines can introduce noise or artifacts, degrading the overall image quality. Therefore, removing silver lines is necessary to improve the image's visual appeal and ensure more reliable subsequent output.

[0073] The specific process for removing silver lines includes: converting the image from the RGB color space to the HSV color space, creating a binary mask to identify the silver lines based on their color characteristics in the HSV space, performing morphological dilation, and finally filling in the missing parts based on the Navier-Stokes algorithm and the fast traversal method.

[0074] The specific steps for removing silver lines are as follows:

[0075] Step 1: Convert the image from the RGB color space to the HSV color space;

[0076] Step 2: Based on the specific color characteristics of the silver line in the HSV space, create a binary mask to accurately identify the pixels within this color range ([0,0,150]-[255,250,255]);

[0077] Step 3: Perform morphological dilation to enhance the effect of the binary mask and ensure that it can completely cover the silver lines that you want to remove;

[0078] Step 4: Based on the Navier-Stokes algorithm, propagate information about the damaged area to the surrounding area. Gradually expand the known area using a fast-moving method to fill in the missing parts.

[0079] Step S213: Create a training and testing dataset, which includes images of photovoltaic panels with different levels of dust accumulation and corresponding transmittance data labels.

[0080] Furthermore, the dust accumulation image is obtained by performing pixel-level operations on the image of a clean photovoltaic panel based on the principle of light attenuation in photovoltaic panels.

[0081] Furthermore, the principle of light attenuation in photovoltaic panels is an improvement derived from an atmospheric scattering model combined with the actual situation of dust accumulation on the surface of photovoltaic panels.

[0082] The feasible method for synthesizing dusty images based on the principle of photovoltaic panel light attenuation involves dividing a clean photovoltaic panel image into 16×16 pixel blocks, randomly generating transmittance, and calculating the presence of dusty blocks using a formula, given no dusty blocks and global atmospheric light a (set to 1). This allows for the acquisition of photovoltaic panel dusty images with different degrees of dust accumulation and their corresponding transmittance data labels, thus constructing a training and testing dataset.

[0083] In practice, during the process of capturing images of dust-covered photovoltaic panels, the light reflected by the photovoltaic panels is scattered and absorbed due to the presence of tiny dust particles in the light transmission path. The attenuation of incident light caused by dust accumulation is similar to the effect of haze and dust in the atmosphere on light. The atmospheric scattering model describes the imaging mechanism under the combined effect of haze and light. Therefore, this embodiment uses the atmospheric scattering model in combination with the photovoltaic dust accumulation image to describe the attenuation process of light passing through the dust accumulation surface.

[0084] The mathematical expression for the atmospheric scattering model is:

[0085] I(m)=J(m)t(m)+a(1-t(m)) (5)

[0086] t(m)=e -θd(m) (6)

[0087] m represents the position of a pixel, I(m) is the pixel value of the dusty photovoltaic panel image at point m, J(m) is the pixel value of the clean photovoltaic panel image at point m, t(m) is the transmittance of the medium at point m, a is global atmospheric light, and θ is the scattering coefficient of the transmission medium. Transmittance t(m) indicates the degree to which light is scattered and absorbed when passing through the dust layer, while global atmospheric light a represents the overall brightness of the image.

[0088] In equation (6), when d(m) approaches infinity, t(m) approaches 0, and combining equation (5) we have:

[0089] a=I(m) d(m)→∞ (7)

[0090] In long-distance imaging, d(m) is not infinite, but a very large value. According to equation (6), this makes the transmittance t1 very small. Therefore, the global atmospheric light a can be accurately estimated based on the following expression:

[0091]

[0092] The principle of light attenuation in photovoltaic panels indicates that as long as the transmittance of the medium is obtained, the conversion between clean images and dusty images can be achieved.

[0093] Since image content is independent of medium transmittance, and medium transmittance is locally constant, that is, image pixels in a small patch tend to have similar medium transmittance.

[0094] Assume the transmittance of a single image patch is random. Given a gray-free patch J... p (m), global atmospheric light a and random transmittance t∈(0,1), with gray blocks as I p (m)=J p (m)t(m)+a(1-t(m)), where atmospheric light a is set to 1.

[0095] The specific operation involves segmenting the clean photovoltaic panel image into 16×16 pixel blocks, resulting in 45,522 images. A random number t(m) (0 ≤ t(m) ≤ 1) is generated, and each segmented photovoltaic panel image is then processed to add grayscale, generating grayscale images and their corresponding transmittance, thus obtaining the dataset images as shown below. Figure 2 As shown.

[0096] Step S22: The specific process of the model building module includes:

[0097] (1) The RGB color image of the photovoltaic panel with accumulated dust is used as the model input, and the dust concentration deposited on the photovoltaic surface is used as the model output in the form of average transmittance and medium transport map.

[0098] (2) The ash accumulation detection model consists of a feature extraction layer, four residual blocks, and a fully connected layer.

[0099] Furthermore, the main function of the feature extraction layer is to extract low- to mid-level feature information, such as edges, textures, and local gray-level distributions, from the input image of the dusty photovoltaic panel. This layer employs a multi-scale convolution mechanism, processing the input image simultaneously with convolution kernels of different scales to capture information from different spatial ranges. This fusion helps to form a more comprehensive and richer feature representation.

[0100] Specifically, multi-scale convolution is used in the first layer of the network structure, with convolution kernel sizes of 3×3, 5×5, and 7×7, and 32-channel filters are used for all three scales.

[0101] The multi-scale convolution calculation process of the feature extraction layer is as follows: first, a convolution operation is performed, then batch normalization and ReLU activation function processing are performed in sequence, and finally the channel outputs corresponding to the three convolution kernels at different scales are added together to obtain the output of the layer.

[0102] The specific calculation process is as follows:

[0103] Assuming the input feature map X, the output of the convolutional layer (without bias) is:

[0104] F k×k =W k×k *X (9)

[0105] Where Conv k×k These represent the kernel sizes for 3×3, 5×5, and 7×7 convolutions, respectively, with * indicating the convolution operation.

[0106] The output of the batch normalization is then:

[0107]

[0108] Here, ε is a very small positive number used to prevent division by zero. μ and σ 2 γ represents the mean and variance of the convolutional layer output, respectively. k×k ,β k×k , where is a learnable parameter, and ⊙ represents element-wise multiplication.

[0109] After passing through the ReLU activation layer, the output is:

[0110]

[0111] Finally, the outputs of the three channels are summed to obtain the first layer output:

[0112]

[0113] The second layer of the network consists of four residual blocks. In deep learning, increasing the number of layers beyond a certain point usually leads to worse training results than expected; this phenomenon is called network degradation. To address the network degradation problem, this embodiment introduces a residual structure:

[0114] The residual block uses two 3×3 convolutional kernels to extract features, each followed by batch normalization (BN) and a ReLU activation function; in addition, each residual block also contains skip connections, residual mappings, and an SEBlock module. The SEBlock attention mechanism adaptively recalibrates the feature responses across channels by explicitly modeling the interdependencies between channels.

[0115] This module can construct any given transformation F tr X→U

[0116] For simplicity, in the following symbols, F tr This is a standard convolution. Let V = [v1, v2, ..., v c ] represents the set of learned filter parameters, where v c This refers to the parameters of the c-th filter.

[0117] Then F can be tr The output is written as U = [u1, u2, ..., u c ],in Here, * represents convolution.

[0118] The SEBlock module introduces two steps after the convolution operation: Squeeze and Excitation. The Squeeze operation uses global average pooling across all channels to compress the W×H×C feature map into a 1×1×C feature vector. The Excitation operation employs a gating mechanism combined with the Sigmoid activation function, and recalibrates the cross-channel feature responses through parameterized gating mechanisms using two linear transformation layers. The specific operation flow is as follows:

[0119] 1) Squeeze operation:

[0120] F sq The operation involves using global average pooling of the channels to directly compress the W×H×C feature map containing global information into a 1×1×C feature vector. In other words, each two-dimensional channel is transformed into a value with a global receptive field. At this point, one pixel represents one channel, masking spatial distribution information and making better use of the correlation between channels.

[0121] Specifically, this is expressed as a statistic. It is obtained by reducing the spatial dimension H×W by U, where the formula for calculating the c-th element of z is:

[0122] 2) Excitation operation:

[0123] In order to take advantage of the information gathered in the Squeeze operation, a second operation is then used to fully capture channel dependencies.

[0124] To achieve this goal, a simple gating mechanism is chosen, and Sigmoid activation is used:

[0125] s = F ex (z,W)=σ(W2δ(W1z)) (14)

[0126] Where δ refers to the ReLU function. and

[0127] To limit model complexity and aid generalization, two linear transformations are used before and after the nonlinear activation function to parameterize the gating mechanism, namely the ReLU activation function of the dimensionality reduction layer with parameters W1 and r.

[0128] Then, a dimension-increasing layer with parameter W2 is applied. Finally, the transformation U is readjusted using the sigmoid activation function to obtain the final output.

[0129]

[0130] Where F scale (uc ,s c ) refers to the feature map and scalar s c Multiplication of channels between them.

[0131] The third layer of the network is a fully connected layer, whose function is to map the features extracted from the previous feature extraction layer and residual block into the final output, namely the average transmittance and the medium transport map.

[0132] The fully connected layer in this embodiment includes an average pooling layer, two linear transformations, and a ReLU activation function.

[0133] Average pooling is a pooling operation that reduces the size of an image or feature map by dividing it into different regions and calculating the average value within each region.

[0134] By performing adaptive average pooling on all pixels of each channel, the two-dimensional feature map of each channel is compressed into a scalar.

[0135] Average pooling is represented as follows:

[0136]

[0137] The number of channels C is 32, and H and W are the height and width, respectively.

[0138] Linear transformation 1 maps the pooled feature tensor from a 32-dimensional vector to a 16-dimensional vector. Assume the input is... The output is The specific calculation formula is as follows:

[0139] y = w1x + b1 (17)

[0140] in This is the weight matrix, connecting the input and the hidden layer. This is the bias vector.

[0141] The output result, after passing through the ReLU activation function, can be expressed as:

[0142] y′=ReLU(w1x+b1) (18)

[0143] Linear transformation 2 maps the 16-dimensional vector from the previous step to a scalar value (i.e., the output size is 1-dimensional). The calculation method is the same as that of linear transformation 1, and will not be repeated here.

[0144] Step S23: Model training and evaluation module, used to set the learning rate, batch size, number of iterations, adaptive moment estimator (Adam), loss function, training set and validation set, evaluation metric, and save the optimal model.

[0145] The optimal model includes: recording the training loss and validation loss for each epoch during the training of the detection model. The model weights are saved every 10 epochs, and the weights with the minimum loss are selected to obtain the optimal model. The photovoltaic dust accumulation image is input into the optimal model, and the average transmittance and its dielectric transport map are output to visualize the specific distribution of dust on the photovoltaic panel surface.

[0146] During model training, the complete dataset was divided into training and validation sets at a ratio of 80% and 20%, respectively. The filter weights of each layer of the model followed a Gaussian distribution. The learning rate was set to 0.000001. The model was trained for 500,000 iterations on a PC equipped with an Nvidia GeForce GTX780 GPU. The batch size was 32. The mean squared error (MSE) was used as the loss function.

[0147] The model learns the mapping function between RGB values ​​and medium transmittance, and is trained using hyperparameters and weights (Θ). p The loss function between (m) and the corresponding real medium transmittance t is minimized.

[0148] This embodiment demonstrates the use of mean squared error (MSE) as the loss function:

[0149]

[0150] The complete dataset is divided into a training set and a validation set, with the training set accounting for 80% of the total dataset.

[0151] The filter weights of each layer of the model follow a Gaussian distribution (mean μ = 0, standard deviation σ = 0.001), and the learning rate is set to 0.000001.

[0152] Training was performed on a PC equipped with an Nvidia GeForce GTX 780 GPU, with 500,000 iterations and a batch size of 32.

[0153] To analyze the effectiveness of the model, MSE, MAE, and RE were used as evaluation metrics.

[0154] MAE is the average of the absolute errors between predicted and actual values. It reflects the average deviation between the model's predictions and the actual results, and thus reflects the accuracy of the model's predictions. It is defined as follows:

[0155]

[0156] MSE is the average of the squared errors between the predicted and actual values. It is more sensitive to larger errors, thus helping to identify shortcomings in the model. It is defined as follows:

[0157]

[0158] RE is the ratio of the error between the predicted and actual values ​​to the actual value. It reflects the proportion of error to the actual value and provides information on the ratio of error to the actual value, facilitating error assessment at different scales. The definition is as follows:

[0159]

[0160] In the above formula, n represents the sample size, y i This represents the true value of the i-th sample. Let represent the predicted value of the i-th sample. The smaller the value on the left side of the equation, the higher the prediction accuracy of the model.

[0161] Step S24: The detection application module is used to input photovoltaic panels with different dust accumulation levels in actual working conditions into the optimal model, adjust the contrast of the input photovoltaic panel dust accumulation image, calculate the average transmittance of the dust accumulation image, and output the corresponding medium transmission map and perform Gaussian smoothing to reduce high-frequency noise and visualize the specific distribution of dust on the photovoltaic panel surface.

[0162] Furthermore, the process of adjusting the contrast of the tested gray image includes:

[0163] Contrast adjustment is a common operation in image processing. By adjusting the differences in brightness between different areas of an image, the visual effect of the image is made more vivid. When adjusting the contrast, the brightness values ​​of the image are stretched or compressed according to a contrast factor. The larger the contrast factor, the stronger the contrast of the image, and vice versa.

[0164] The basic idea behind adjusting contrast is to perform a non-linear transformation on each pixel value in the image. The adjusted value can be calculated using the following formula:

[0165] I′(x,y)=(I(x,y)-128)×f+128 (23)

[0166] I(x,y): The pixel value at position I(x,y) in the original image. I′(x,y) is the adjusted value, with a contrast factor f = 2.5, adjusting the contrast to 2.5 times the original.

[0167] The output medium transport map is Gaussian smoothed by weighted averaging of each pixel and its neighborhood to reduce noise, resulting in a smoother image.

[0168] Gaussian smoothing is achieved by applying a weighted average of each pixel in an image using a Gaussian function. The Gaussian function is based on a normal distribution and has the following form:

[0169]

[0170] Where (x,y) is the position of the current pixel, and σ is the standard deviation of the Gaussian function, which controls the degree of smoothing. In this embodiment, σ = 8.

[0171] This is accomplished by applying a Gaussian kernel (i.e., a matrix). The value of each pixel is multiplied by and summed with the weighted values ​​of its surrounding pixels (based on the value of the Gaussian kernel). In this embodiment, the convolution kernel size is 13×13. This eliminates high-frequency noise in the medium transport map, enhances the visual continuity of the image, and yields a dust distribution that more closely resembles reality.

[0172] The input and output images of this embodiment are further shown as follows: Figure 4 As shown, (a) and (e) are the original gray images, (b) and (f) are the images with increased contrast, (c) and (g) are the media transport maps estimated by the model, and (d) and (h) are the results after Gaussian smoothing of the estimated media transport maps.

[0173] To comprehensively evaluate the effectiveness of the proposed model, this embodiment conducted 10 comparative experiments using MobileNetV2, MobileNetV2(1.4), ResNet50, ShuffleNet(1.5), ShuffleNet(x2), and the model proposed in this embodiment. The specific network structure of the model is as follows: Figure 3 As shown.

[0174] All models used the Adaptive Moment Estimator (IME) optimizer with a learning rate of 0.000001, a batch size of 32, and the loss function MSE. Subsequently, this embodiment calculated the RE, MAE, and MSE metrics between the actual and estimated transmittance, and conducted a detailed comparative analysis of the experimental results of several algorithms. The experimental results are shown in Table 1.

[0175] Table 1

[0176] Model MAE MSE RE Parameters MobileNetV2 0.1756 0.0308 0.2277 3504872 MobileNetV2(1.4) 0.0697 0.0048 0.2849 6963136 ResNet50 0.2688 0.0722 0.1988 25557032 ShuffleNet(1.5) 0.0347 0.1864 0.2212 1364608 ShuffleNet(x2) 0.0238 0.0056 0.2879 2385184 The model 0.0211 0.0004 0.0468 83809

[0177] Specifically, ShuffleNet and MobileNetV2 (1.4) outperform some other models in terms of MAE. However, their MAE is still higher than that of the aforementioned models, indicating that the aforementioned models are more accurate in predicting dust levels.

[0178] Both versions of MobileNetV2, ShuffleNet(x2), and ResNet50 all achieved MSE scores below 0.1, indicating that these models can fit the data well to some extent and reduce the sum of squared errors. However, the relative errors of these models are much higher than those of the aforementioned models.

[0179] Both versions of ShuffleNet and MobileNetV2(1.4) performed well on MAE, but their RE metrics all exceeded 0.2, indicating that these models had large REs and low prediction stability and accuracy.

[0180] In contrast, the relative error of the model proposed in this embodiment is much lower than 0.2, and the prediction results are more accurate and stable.

[0181] In terms of MAE, MSE, and RE, the model proposed in this embodiment significantly outperforms other neural network models. Furthermore, it has a much lower number of parameters than other algorithms, resulting in improved computational speed.

[0182] The above demonstrates the effectiveness and superiority of the model proposed in this embodiment in the dust detection experiment of photovoltaic panels.

[0183] This embodiment also provides a cloud monitoring platform for deploying the deep learning-based photovoltaic panel dust accumulation visualization detection system, including: a data transmission module, a data processing and analysis module, an early warning module, and a remote monitoring module;

[0184] Step S31: Data transmission module, including: network interface, communication protocol (TCP / IP), data encryption, data buffer, and status monitoring.

[0185] This technology connects testing equipment to a cloud platform via IoT (Internet of Things) to enable real-time data transmission and storage, ensuring the stability and security of data transmission. Specifically:

[0186] It provides multiple network connection options to ensure stable data transmission to the cloud platform. It selects appropriate communication protocols based on the network environment to ensure efficient and stable data transmission. Data during transmission is encrypted to prevent leakage or tampering. A buffer is used to temporarily store data during transmission to prevent data loss due to network issues. The system monitors data transmission status in real time, automatically handles transmission interruptions or errors, and promptly notifies maintenance personnel.

[0187] Step S32: The data processing and analysis module includes: data storage, data cleaning, and model deployment.

[0188] This is used to process and analyze the collected data, assess the degree of dust accumulation on photovoltaic panels, and generate dust distribution maps and trend charts. Specifically:

[0189] The collected data is stored in a cloud database for easy processing and analysis. Noise and outliers are removed from the data to improve data quality and provide an accurate data foundation for subsequent analysis. The trained deep learning model is deployed to the cloud platform to analyze the real-time acquired grayscale images and output quantitative results of the grayscale level.

[0190] Step S33: The early warning module includes: threshold setting, early warning rules, notification system, and log recording.

[0191] This system is used to automatically issue an early warning signal when the dust accumulation exceeds a set threshold, reminding maintenance personnel to perform timely cleaning and maintenance. Details are as follows:

[0192] Based on the operating status and historical data of the photovoltaic panels, set reasonable thresholds for dust accumulation. Define early warning rules when dust accumulation exceeds the thresholds to ensure timely notification to maintenance personnel. Notify maintenance personnel promptly through various means such as SMS, email, and mobile app to ensure timely problem resolution. Record early warning events and their handling status for subsequent analysis and auditing, providing data support for operation and maintenance management.

[0193] Step S34: The remote monitoring module includes: web application, mobile application, user management, real-time monitoring, and historical data query.

[0194] Maintenance personnel can check the dust accumulation on photovoltaic panels and the system's operating status anytime, anywhere through the cloud platform's web or mobile application. Details are as follows:

[0195] It provides an intuitive web interface for maintenance personnel to view and manage data on their computers. A mobile application is also available, allowing maintenance personnel to view and manage data anytime, anywhere. Multi-user login is supported, enabling users with different permissions to view and manage different data, improving the flexibility of maintenance management. Real-time display of dust accumulation on photovoltaic panels and system operating status ensures maintenance personnel can promptly identify problems. Historical data query functionality facilitates data analysis and review, supporting maintenance decision-making.

[0196] Through the design and implementation of the above modules, the cloud monitoring platform can achieve efficient and accurate monitoring and analysis of dust accumulation on photovoltaic panels, providing strong technical support for the operation and maintenance management of photovoltaic power plants.

[0197] Example 2

[0198] This embodiment also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the system.

[0199] Example 3

[0200] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the system described above.

[0201] Example 4

[0202] This embodiment also provides a computer program product, including a computer program that, when executed by a processor, implements the system described above.

[0203] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based visual detection system for dust accumulation on photovoltaic panels, characterized in that, include: The data acquisition and preprocessing module, model building module, model training and evaluation module, and detection application module are connected sequentially. The data acquisition and preprocessing module is used to acquire images of clean photovoltaic panels, and to perform image correction, silver line removal, and composite dust accumulation images to generate training and testing datasets. The model building module is used to build a photovoltaic dust accumulation detection model, which includes a feature extraction layer, four residual blocks, and a fully connected layer. The model training and evaluation module is used to train the photovoltaic dust accumulation detection model based on the training and testing dataset, and to set evaluation indicators to evaluate the trained photovoltaic dust accumulation detection model in order to obtain the optimal photovoltaic dust accumulation detection model. The detection application module is used to input the photovoltaic dust accumulation image in actual working conditions into the optimal photovoltaic dust accumulation detection model. After preprocessing, transmittance calculation and image postprocessing, it outputs the corresponding medium transmission map and average transmittance to monitor the dust accumulation of photovoltaic panels in real time. The data acquisition and preprocessing module includes an image correction unit, a silver line removal unit, and a gray image synthesis unit. The image correction unit is used to select four corresponding points on the source image and target image of the clean photovoltaic panel, respectively, to construct and solve the perspective transformation equation system to achieve image correction. The silver line removal unit is used to convert the corrected image from RGB color space to HSV color space, create a binary mask to identify the silver line based on the color characteristics of the silver line in HSV space, and then fill in the missing part based on morphological dilation operation and Navier-Stokes algorithm to remove the silver line. The synthesized gray image unit is used to segment the clean photovoltaic panel image into gray-free blocks of preset pixels, and randomly generate transmittance. Based on the gray-free blocks, gray blocks are calculated through an atmospheric scattering model, thereby generating images with different degrees of gray accumulation and their corresponding transmittance data labels. The feature extraction layer employs multi-scale convolution with kernel sizes of 3×3, 5×5, and 7×7, and each scale uses a 32-channel filter. The multi-scale convolution calculation process of the feature extraction layer includes: first performing a convolution operation, then sequentially performing batch normalization and ReLU activation function processing, and finally adding the channel outputs corresponding to the three convolution kernels at different scales to obtain the output of the feature extraction layer. Each residual block includes two 3×3 convolutional kernels, each followed by batch normalization and ReLU activation functions; each residual block also contains skip connections, residual maps, and SEBlock modules; The SEBlock module is used to adaptively recalibrate the feature responses across channels through Squeeze and Excitation operations.

2. The system according to claim 1, characterized in that, The fully connected layer includes an average pooling layer, two linear transformations, and a ReLU activation function; The fully connected layer is used to compress the feature map extracted from the feature extraction layer and residual block into a scalar based on the average pooling layer. After two linear transformations and ReLU activation function processing, the average transmittance and medium transport map are finally output.

3. The system according to claim 1, characterized in that, The model training and evaluation module is used to set the parameters for model training, select the optimizer and loss function, and use the training and testing datasets to train and evaluate the model. Then, the model performance is quantified by mean squared error, mean absolute error and relative error, and the model with the smallest loss during training is saved as the optimal photovoltaic dust accumulation detection model.

4. A cloud monitoring platform, characterized in that, The photovoltaic panel dust accumulation visualization detection system based on deep learning as described in any one of claims 1-3 includes: a data transmission module, a data processing and analysis module, an early warning module, and a remote monitoring module; The data transmission module is used to connect the detection equipment to the cloud monitoring platform through Internet of Things (IoT) technology to achieve real-time data transmission and storage. The data processing and analysis module is equipped with the deep learning-based photovoltaic panel dust accumulation visualization detection system, which is used to process and analyze the collected data, assess the degree of dust accumulation on the photovoltaic panel, and generate dust distribution maps and trend maps. The early warning module is used to automatically issue an early warning signal when the dust accumulation exceeds a set threshold. The remote monitoring module is used to view the dust accumulation on photovoltaic panels and the system's operating status anytime, anywhere through the web or mobile application of the cloud monitoring platform.

5. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the system according to any one of claims 1-3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the system according to any one of claims 1-3.

7. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the system according to any one of claims 1-3.

Images