A solar photovoltaic module fault classification method based on class balancing strategy

Abstract

The present invention discloses a solar photovoltaic module fault classification method based on a category balance strategy. The present invention first collects and labels a data set and performs data enhancement on the sample images of the training set; secondly, the enhanced training set is sampled in parallel by random sampling and category balance sampling; then, a multi-scale classification network model is constructed and a feature space category balance strategy is constructed, and the classification network model is trained by unbalanced feature mixed samples to obtain an optimal classification model; finally, the sample images to be classified are input into the optimal classification model after preprocessing for fault classification; and a prediction matrix of the classification model is output. The classification network model of the present invention has fewer network parameters and computational complexity, which greatly improves the accuracy and speed of photovoltaic module image classification. The present invention also solves the problem that the network is biased towards the majority class in a category imbalanced data set and ignores the minority class, resulting in low classification accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of fault diagnosis, and in particular relates to a solar photovoltaic module fault classification method based on a category balancing strategy. Background Art

[0002] As energy demand continues to rise, demand for solar photovoltaic power generation is rapidly increasing due to its widespread availability, cleanliness, and sustainability. Currently, solar photovoltaic power generation has become a major form of energy production. Therefore, the rapid and reliable identification and classification of photovoltaic module faults is crucial to reduce energy losses and improve economic efficiency.

[0003] Currently, the most commonly used fault diagnosis method is photovoltaic system fault diagnosis based on infrared imaging. Specifically, an infrared thermal imager is embedded in a drone, which captures thermal images of photovoltaic modules along a preset trajectory. The abnormal heating of the photovoltaic cells within the array when a fault occurs is used to identify the fault and accurately locate the fault using the drone's GPS. This method primarily uses a convolutional neural network (CNN) to classify sample images of photovoltaic module faults. The CNN learns the correlations between different fault characteristics and accurately classifies them into corresponding fault categories.

[0004] In the infrared image mode, the background of photovoltaic module fault samples is single and lacks effective information. It mainly relies on information such as the size, lines, and structure of the fault to distinguish. However, due to the small gap between fault classes and the variability of fault types, existing methods have difficulty distinguishing the characteristics of different faults, resulting in low classification accuracy. In response to this, the present application designs a multi-scale convolutional neural network with three convolution branches. The network uses three convolution branches (3×3, 5×5, 7×7) to more effectively capture the size, lines, structure and other features of the fault. By splicing and fusing these features, the network can perceive multi-scale fault information more comprehensively, thereby more comprehensively understanding the structure and details of the image. This design significantly improves the network's classification performance for photovoltaic faults and improves classification accuracy.

[0005] Furthermore, solar photovoltaic modules operate in harsh outdoor environments. Due to various external factors, they can experience a variety of faults, including hotspots, dirt, offline modules, shadows, and cracks. These faults can have varying degrees of impact on the performance and efficiency of photovoltaic systems. However, the probability of these faults occurring varies, leading to a high class imbalance in the fault samples. When training convolutional neural networks on class-imbalanced infrared imaging datasets, the majority class, with abundant samples, typically dominates the network, resulting in excellent performance on these classes but poor performance on the minority class, with fewer samples. This means that some rare but important faults may go unnoticed due to their low frequency and limited sample size. However, timely detection of these faults is crucial for ensuring the reliability and efficiency of photovoltaic systems. Furthermore, the class imbalance of the samples leads to low overall fault classification accuracy. To address this, this application designs a feature space class balancing strategy. By embedding minority class feature information into the feature map of the majority class, this strategy achieves collaborative learning of the majority and minority classes. The majority class acts as a guide, helping the network better understand and learn the features of the minority class. This effectively alleviates the significant class imbalance in photovoltaic module fault data and improves network classification accuracy. Summary of the Invention

[0006] In order to solve the problems that the existing technology cannot effectively capture the characteristics of faults and the network tends to favor the majority class and ignore the minority class in the class-imbalanced infrared imaging dataset, resulting in low classification accuracy, this paper proposes a solar photovoltaic module fault classification method based on a class balance strategy.

[0007] In a first aspect, the present invention provides a solar photovoltaic module fault classification method based on a category balancing strategy, comprising the following steps:

[0008] Step S1: Collect and label the dataset;

[0009] Step S2: perform data augmentation on the sample images of the training set;

[0010] Step S3: Parallel sampling is performed on the enhanced training set using random sampling and class-balanced sampling. Specifically, the enhanced training set is sampled in parallel using a method in which each sample has the same sampling probability and a method in which the minority class has a higher sampling probability, thereby obtaining a random sample set that is biased towards the majority class and a class-balanced sample set that is biased towards the minority class.

[0011] Step S4: constructing a multi-scale classification network model, wherein the classification network model introduces three convolution branches of different scales to capture the size, line, and structural features of the fault;

[0012] Step S5: Construct a feature space class balancing strategy, by embedding the feature information of the minority class in the feature map of the majority class and obtaining the corresponding label according to the ratio of the majority class to the minority class in the sample, thereby obtaining an unbalanced feature mixed sample;

[0013] Step S6: Use the unbalanced feature mixed sample obtained in step S5 to train the classification network model to obtain the optimal classification model;

[0014] Step S7: normalize the sample image to be classified, adjust the resolution, and finally input it into the optimal classification model for fault classification;

[0015] Step S8: Output the prediction matrix of the classification model, find the index value of the maximum value in each row, map the index value to the actual fault type, and obtain the fault classification result corresponding to the sample image.

[0016] In a second aspect, the present invention provides a solar photovoltaic module fault classification device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the above-mentioned solar photovoltaic module fault classification method based on the category balancing strategy is implemented.

[0017] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program for executing a solar photovoltaic module fault classification method based on a class balancing strategy.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] This paper proposes a multi-scale classification network model with fewer network parameters and computational complexity, significantly improving the accuracy and speed of photovoltaic module image classification. It also proposes a feature space class balancing strategy to address the problem of low classification accuracy caused by the network favoring the majority class and ignoring the minority class in class-imbalanced datasets. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is the overall flow chart of the present invention;

[0021] Figure 2 is the parallel sampling strategy of the present invention;

[0022] Figure 3 This is the fault classification network structure diagram of the present invention. DETAILED DESCRIPTION

[0023] The technical solution provided by the present invention will be further described below with reference to the accompanying drawings.

[0024] The embodiment of the present application provides a solar photovoltaic module fault classification method based on a category balancing strategy. The overall process of the method is as follows: Figure 1 As shown, the following steps are included:

[0025] Step S1: Dataset collection and labeling.

[0026] In one embodiment, this step includes the following sub-steps:

[0027] S1.1 uses drones equipped with infrared thermal imaging and GPS positioning to photograph photovoltaic modules according to fixed inspection routes.

[0028] S1.2 classifies and labels the fault sample images according to hot spots, dirt, offline modules, shadows, and cracks.

[0029] S1.3 normalizes the fault sample image.

[0030] S1.4 adjusts the resolution of the normalized sample images to 128×128×3.

[0031] S1.5 divides the dataset into training set and validation set in a ratio of 8:2.

[0032] Step S2: Perform data augmentation on the sample images in the training set. All sample images in the training set are expanded in parallel by horizontal flipping, vertical flipping, and rotation. After augmentation, the number of sample images in the training set increases threefold.

[0033] Through these data augmentation methods, more training sample variations are obtained, which not only preserves image feature information but also increases data diversity. This effectively enhances the generalization ability of the network, making it better adaptable to the changes and noise that may exist in various practical application scenarios.

[0034] Step S3: Parallel sampling is performed on the enhanced training set using random sampling and class-balanced sampling. The parallel sampling strategy flow chart is as follows: Figure 2 As shown, the enhanced training set is sampled in parallel in a way that each sample has the same sampling probability and the minority class has a higher sampling probability, obtaining a random sample set biased towards the majority class and a class-balanced sample set biased towards the minority class.

[0035] In one embodiment, this step includes the following sub-steps:

[0036] Step S3.1: Construct a randomly distributed sample set.

[0037] An unbiased data selection method is used to randomly select sample images from the enhanced training set. In this process, each sample has an equal probability of being selected, and each sample is selected independently. Selecting one sample will not affect the selection results of other samples.

[0038] Step S3.2: Construct a class-balanced sample set.

[0039] For all c classes, let n i is the number of samples in the i-th category, and the total number of samples is Then the sampling probability of class i is defined as:

[0040]

[0041] The sampling factor r = 1 / 2, the sampling weight of the i-th class is equal to the number of samples n of the i-th class i Inversely proportional to the square root of .

[0042] Sampling is performed from the enhanced training set according to the sampling probability of each class. The number of samples in each class after final sampling is defined as:

[0043]

[0044] Where β = (N-1) / N.

[0045] By utilizing this parallel strategy, data diversity and minority class weighting are combined to increase the weight of minority classes during network training.

[0046] Step S4: Construct a multi-scale classification network model f(x), the network structure is as follows Figure 3 shown.

[0047] The input image size is 128×128 and the number of channels is 3.

[0048] The convolution kernel size used in the first layer of the network is 7×7, the stride is 2, and the edge padding is 3. This layer outputs a feature map with 64 channels and a size of 64×64.

[0049] The second layer uses a maximum pooling layer with a pooling window size of 3×3 and a stride of 2. This layer outputs a feature map with 256 channels and a size of 32×32.

[0050] The convolution kernel size used in the third layer is 5×5, the stride is 1, and the edge padding is 2. This layer outputs a feature map with 256 channels and a size of 32×32.

[0051] The fourth layer uses a multi-scale convolution block with three convolution branches, using 3×3, 5×5 and 7×7 convolution kernels respectively, with a stride of 1 and edge padding of 1, 2 and 3 respectively. Each branch outputs a 256-channel feature map with a size of 32×32.

[0052] The fifth layer also uses the same multi-scale convolutional block and finally outputs three parallel feature maps with 256 channels and a size of 32×32.

[0053] The output feature maps of the three branches are then concatenated along the channel dimension, stacking the output of each branch along the channel dimension to form a deeper feature map. This concatenation operation enriches the feature map and improves the model's representational capabilities. The final output is a feature map with 768 channels and a size of 32×32.

[0054] The sixth layer uses a maximum pooling layer with a pooling window size of 3×3 and a step size of 2. This layer outputs a feature map with 512 channels and a size of 16×16.

[0055] The convolution kernel size used in the seventh layer is 3×3, the stride is 1, and the edge padding is 1. This layer outputs a feature map with 512 channels and a size of 16×16.

[0056] The convolution kernel size used in the eighth layer is 3×3, the stride is 1, and the edge padding is 1. This layer outputs a feature map with 512 channels and a size of 16×16.

[0057] The ninth layer uses an average pooling layer with a pooling window size of 3×3 and a stride of 2. This layer outputs a feature map with 1024 channels and a size of 8×8.

[0058] The tenth layer is a fully connected layer, which is used to convert the multi-dimensional feature map of the upper layer into a one-dimensional feature map for output.

[0059] The eleventh layer is the softmax classification layer, which is used to calculate the probability that the output belongs to each category and obtain the classification result.

[0060] As the network goes deeper, the number of channels gradually increases. This is because as the size of the feature map decreases, more channels are required to learn multi-level and multi-scale features. Furthermore, a batch normalization layer is added after each convolutional layer. This helps accelerate network training and improve the model's generalization capabilities. It normalizes each mini-batch of data, making the distribution of the input data more stable, facilitating gradient propagation, and reducing the likelihood of gradient vanishing.

[0061] The classification network model f(x) designed in step S4 of the present embodiment introduces three convolution branches of different scales (3×3, 5×5, and 7×7) to more effectively capture the size, lines, structure, and other characteristics of the fault. This step enables the network to more comprehensively perceive multi-scale fault information by splicing and fusing these multi-scale features, thereby more comprehensively understanding the structure and details of the image. This design helps improve the classification performance of photovoltaic faults, enabling the model to more accurately identify different types of faults, thereby improving the network classification accuracy.

[0062] Step S5: Construct a feature space category balancing strategy.

[0063] In one embodiment, this step includes the following sub-steps:

[0064] Step S5.1: Traverse the structural information of the network f(x) and save the structural information in the list variable to obtain the hidden layer set net_list of the network.

[0065] Step S5.2: Sampling from the random sample set and the class-balanced sample set obtained in step S3, select samples (x b ,y b ), and select samples from the class-balanced sample set that is biased towards the minority class (x p ,y p ), and are input into the network f(x) in parallel for feature extraction. Subsequently, the middle hidden layer k is randomly selected from the network hidden layer set net_list.

[0066] Step S5.3: Draw samples (x b ,y b ) The feature representation of the kth hidden layer (g k (x b ),y b ), and the sample (x p ,y p ) The feature representation of the kth hidden layer (g k (x p ),y p ). where g k (x b ), g k (x p ) represents a feature map of size W×H, y b 、y p For the corresponding label, k (x b ),y b ) and (g k (x p ),y p) for block-level feature mixing.

[0067] Step S5.4: Create a binary mask that determines which features in the feature map will be mixed and which features will remain unchanged.

[0068] Furthermore, the step S5.4 is specifically as follows:

[0069] First, create a tensor Q with the same size as the feature map of the middle hidden layer of the kth layer, and all its features f ij The initial value is set to 1. The tensor Q is adjusted by the adjustment factor γ adj Multiply them together to get the probability tensor P.

[0070] Then, Bernoulli(P) is used to perform Bernoulli sampling on each feature in the probability tensor P. If the feature f ij If the sampling result is 0, then set the corresponding mask M ij =1, if feature f ij The sampling result is 1, then select the corresponding M in the mask ij is the center and the side length is b s The entire square area, let the mask M ij =0, indicating that the corresponding feature block will be modified.

[0071] The adjustment factor γ adj It is determined by the size of the feature map, the size of the feature block, and the size of the effective modification area. adj The formula is as follows:

[0072]

[0073] where f area =W×H represents the area of ​​the feature map, Represents the area of ​​the feature block. γ is an artificially set adjustment factor, v area =(Wb s +1) 2 Indicates the valid area in the feature map where features can be modified.

[0074] After obtaining the mask, the unmixed part of the feature map can be calculated k (x b ), and the mixed feature block Patchmix and its label Y Patchmix .

[0075] unchanged k (x b )=(1-Mask)⊙g k (x b )

[0076] Patchmix=Mix α (Patch b ,Patch p )

[0077] Y Patchmix =Mix α (y b ,y p )

[0078] Where ⊙ is element-wise multiplication. b 、Patch p Respectively represent the mask from the feature map g k (x b ), g k (x p ) feature blocks extracted.

[0079] The mixing ratio β between the two feature blocks is sampled from the Beta distribution Beta(2,2). The mixing operation is defined as:

[0080] Mix β (a,b)=β×a+(1-β)×b

[0081] Where a and b represent any two vectors, and the blending operation can make the blending between the two vectors smoother.

[0082] Finally, we get an unbalanced feature mixed sample Its characteristic diagram And the label Y is represented as:

[0083]

[0084] Y=Mix pc (Y Patchmix ,y b )

[0085] In the formula, pc represents the proportion of the features involved in the mixture to the total features in the feature map.

[0086] Step S5 of the present invention primarily addresses the significant class imbalance in photovoltaic module failure data. This step embeds the minority class's feature information within the majority class's feature map and derives corresponding labels based on the majority-to-minority ratio within the sample, thereby generating a mixed sample with unbalanced features. This enables collaborative learning of the majority and minority classes. The majority class acts as a guide, helping the network better understand and learn the minority class's features and increasing the network's focus on the minority class.

[0087] Step S6: Use the unbalanced feature mixed sample obtained in step S5 To train the network.

[0088] Specifically, the mixed feature samples As the input of the k+1th middle hidden layer of the network, it continues to pass through the network until the last layer. Then, the following loss function is minimized:

[0089]

[0090] in Representation feature map Mapping to output, Indicates mapping the feature map With different labels Y Patchmix 、y b , Y are input into the cross entropy loss results, L(f) loss function is based on the sample The ratio of the majority class to the minority class is designed.

[0091] The multi-scale classification network model uses a cross-entropy loss function and an Adamw optimizer with a fixed step-size decay strategy. During training, network weights and bias parameters are updated to generate a classifier. Simultaneously, validation samples are used to continuously validate the model classifier. After each iteration of the model classifier, the validation sample set is input into the model classifier to determine the recognition accuracy. Finally, the classifier hyperparameters with the highest recognition accuracy are saved as the optimal classification model.

[0092] Step S7: normalize the sample images to be classified, then adjust their resolution to 128×128×3, and finally input them into the optimal classification model for fault classification.

[0093] Step S8: Output the prediction matrix of the classification model, find the index value of the maximum value in each row, map the index value to the actual fault type, and obtain the fault classification result corresponding to the sample image.

[0094] An embodiment of the present application also discloses a solar photovoltaic module fault classification device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the above-mentioned solar photovoltaic module fault classification method based on the category balancing strategy is implemented.

[0095] An embodiment of the present application further discloses a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program is used to execute the above-mentioned solar photovoltaic module fault classification method based on the category balancing strategy.

[0096] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0097] The processor in the present application may include one or more processing cores. The processor calls the data stored in the memory by running or executing the instructions, programs, code sets or instruction sets stored in the memory, performs the various functions of the present application and processes data. The processor can be at least one of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller and a microprocessor. It is understandable that for different devices, the electronic device for realizing the above-mentioned processor function can also be other, and the embodiments of the present application are not specifically limited.

[0098] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only.

[0099] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such process, method, commodity, or apparatus.

[0100] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A solar photovoltaic module fault classification method based on a class balancing strategy, characterized by The method comprises the following steps: Step S1: Collect and label the dataset; Step S2: perform data augmentation on the sample images of the training set; Step S3: Parallel sampling is performed on the enhanced training set using random sampling and class-balanced sampling. Specifically, the enhanced training set is sampled in parallel using a method in which each sample has the same sampling probability and a method in which the minority class has a higher sampling probability, thereby obtaining a random sample set that is biased towards the majority class and a class-balanced sample set that is biased towards the minority class. Step S4: constructing a multi-scale classification network model, wherein the classification network model introduces three convolution branches of different scales to capture the size, line, and structural features of the fault; Step S5: Construct a feature space class balancing strategy, by embedding the feature information of the minority class in the feature map of the majority class and obtaining the corresponding label according to the ratio of the majority class to the minority class in the sample, thereby obtaining an unbalanced feature mixed sample; Step S6: Use the unbalanced feature mixed sample obtained in step S5 to train the classification network model to obtain the optimal classification model; Step S7: normalize the sample image to be classified, adjust the resolution, and finally input it into the optimal classification model for fault classification; Step S8: Output the prediction matrix of the classification model, find the index value of the maximum value in each row, map the index value to the actual fault type, and obtain the fault classification result corresponding to the sample image; Wherein, step S5 includes the following steps: Step S5.1: traverse the structural information of the classification network model and save the structural information in a variable to obtain the hidden layer set of the network; Step S5.2: Sampling from the random sample set and the class-balanced sample set obtained in step S3, select samples (x b ,y b ), and select samples from the class-balanced sample set that is biased towards the minority class (x p ,y p ), and input the classification network model in parallel for feature extraction; then, randomly select the middle hidden layer k from the hidden layer set; Step S5.3: Draw samples (x b ,y b ) The feature representation of the kth hidden layer (g k (x b ),y b ), and the sample (x p ,y p ) The feature representation of the kth hidden layer (g k (x p ),y p );to (g k (x b ),y b ) and (g k (x p ),y p ) performs block-level feature mixing; Step S5.4: Create a binary mask to determine the features involved in the blending in the feature map.

2. A solar photovoltaic module fault classification method based on a category balancing strategy according to claim 1, characterized in that: Step S1 specifically includes: S1.1 uses a drone equipped with infrared thermal imaging and GPS positioning to photograph photovoltaic modules according to a fixed inspection route; S1.2 classify and mark the fault sample images according to hot spots, dirt, offline modules, shadows, and cracks; S1.3 normalizes the fault sample image; S1.4 uniformly adjusts the resolution of the normalized sample images; S1.5 Divide the dataset into training set and validation set in proportion.

3. The solar photovoltaic module fault classification method based on the category balancing strategy according to claim 1, characterized in that: Step S2 specifically includes: All sample images in the training set are expanded in parallel by horizontal flipping, vertical flipping, and rotation. After enhancement, the number of sample images in the training set increases three times.

4. The solar photovoltaic module fault classification method based on the category balancing strategy according to claim 1, characterized in that: Step S3 includes the following steps: Step S3.1: Construct a randomly distributed sample set; An unbiased data selection method is used to randomly select sample images from the enhanced training set. Step S3.2: Construct a class-balanced sample set; Define the sampling probability of the i-th class, and sample from the enhanced training set according to the sampling probability of each class.

5. The solar photovoltaic module fault classification method based on the category balancing strategy according to claim 1, characterized in that: Step S5.4 is specifically: First, create a tensor Q with the same size as the feature map of the middle hidden layer of the kth layer, and add the tensor Q to the adjustment factor γ adj Multiply them together to get the probability tensor P; Subsequently, Bernoulli sampling is performed on each feature in the probability tensor P; if the feature f ij If the sampling result is 0, then set the corresponding mask M ij =1, if feature f ij The sampling result is 1, then select the corresponding M in the mask ij is the center and the side length is b s The entire square area, let the mask M ij =0, indicating that the corresponding feature block will be modified; After obtaining the mask, the unblended part of the feature map remains unchanged. k (x b ), and the mixed feature block Patchmix and its label Y Patchmix Expressed as: unchangedg k (x b )=(1-Mask)⊙g k (x b ) Patchmix=Mix α (Patch b ,Patch p ) AND Patchmix =Mix α (and b ,and p ) Where ⊙ is element-wise multiplication, Patch b 、Patch p Respectively represent the mask from the feature map g k (x b ), g k (x p ) The extracted feature blocks; The mixing ratio β between the two feature blocks is sampled from the Beta distribution Beta(2,2), and the mixing operation is defined as: Mix β (a,b)=β×a+(1-β)×b Where a and b represent any two vectors; Finally, we get an unbalanced feature mixed sample Its characteristic diagram And the label Y is represented as: Y=Mix pc (AND Patchmix ,and b ) In the formula, pc represents the proportion of the features involved in the mixture to the total features in the feature map.

6. The solar photovoltaic module fault classification method based on the category balancing strategy according to claim 1, characterized in that: Step S6 specifically includes: Mixing feature samples As the input of the k+1th intermediate hidden layer of the network model, it continues to be passed through the network until the last layer; then, the loss function is minimized.

7. A solar photovoltaic module fault classification device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the solar photovoltaic module fault classification method based on the category balancing strategy described in any one of claims 1 to 6 is implemented.

8. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and the computer program is used to execute the solar photovoltaic module fault classification method based on the category balancing strategy according to any one of claims 1 to 6.

Images