Photovoltaic module damage detection method, system, storage medium and device

Abstract

The present invention relates to the field of photovoltaic intelligent detection technology, and provides a photovoltaic component damage detection method, system, storage medium and device. The method includes, based on a number of groups of feature combinations, using a data feature dynamic fusion module to obtain fusion features; the data feature dynamic fusion module includes four feature data selection blocks and a global feedback network, each feature data selection block is used to select feature combinations, and the selected feature combinations are input into the global feedback network to determine whether the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block have repeated combination selection, repeated data selection or all three types of data have been selected. If not, the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block are weightedly fused to obtain fusion features; a damage detection model is used to judge the damage of the photovoltaic components on site. Improve the detection efficiency.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic intelligent detection technology, and in particular to a photovoltaic component damage detection method, system, storage medium and equipment. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] At present, photovoltaic power generation systems have been widely installed and applied. As the core components of the system, the performance and reliability of photovoltaic modules directly affect the power generation efficiency and service life of the system. However, as time goes by, the core components of photovoltaic modules may produce cracks, hot spots, aging and shading, which will lead to a decrease in the power generation efficiency of the system.

[0004] Existing intelligent detection methods are mostly single-type data-driven damage detection methods, which cannot comprehensively and accurately detect various defects of photovoltaic modules and have problems such as poor detection accuracy and limited coverage. Summary of the invention

[0005] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a photovoltaic module damage detection method, system, storage medium and equipment. The present invention adopts an automated intelligent detection method to replace the traditional manual inspection, which greatly improves the detection efficiency, reduces the labor cost and time cost, and makes the maintenance of photovoltaic modules more efficient.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A first aspect of the present invention provides a method for detecting damage to a photovoltaic module.

[0008] A photovoltaic module damage detection method, comprising:

[0009] Acquire electrical characteristic data, image data and text data of the photovoltaic module, perform preprocessing and encoding, and obtain electrical characteristic feature vectors, text feature vectors and image feature vectors;

[0010] The electrical characteristic feature vector, the text feature vector and the image feature vector aligned in the time dimension are converted into a multi-channel input tensor, the multi-channel input tensor is input into a convolutional neural network, and the first feature, the second feature and the third feature of the output are combined to obtain several groups of feature combinations;

[0011] Based on several groups of feature combinations, a data feature dynamic fusion module is used to obtain fusion features; wherein the data feature dynamic fusion module includes four feature data selection blocks and a global feedback network, each feature data selection block is used to select a feature combination, and the selected feature combination is input into the global feedback network to determine whether there is a combination selection duplication, data selection duplication or all three types of data have been selected between the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block; if not, the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block are weightedly fused to obtain a fusion feature; if so, the fusion feature is output;

[0012] Based on the fusion features, a damage detection model is used to judge the damage of on-site photovoltaic modules.

[0013] Furthermore, the output first feature, second feature and third feature are combined to obtain several groups of feature combinations, and the method includes: mapping the first feature, the second feature and the third feature to obtain three groups of mapped feature combinations, summing the first feature, the second feature and the third feature in pairs and summing them together to obtain four groups of summed feature combinations; setting weight coefficients of the first feature, the second feature and the third feature, performing weighted summation, and obtaining a group of weighted feature combinations.

[0014] Furthermore, the feature data selection block adopts a Lasso regression feature selection algorithm, introduces random disturbance, calculates the coefficient of the Lasso regression, and selects a feature combination corresponding to the maximum value of the coefficient of the Lasso regression.

[0015] Furthermore, the data feature dynamic fusion module also includes four attention blocks, one of which is after each feature data selection block. The feature combination selected by the previous feature data selection block is input into the previous attention block, the feature combination selected by the current feature data selection block is input into the current attention block, and the fused features output by the previous feature data selection block are input into the current attention block and weightedly fused with the feature combination selected by the current feature data selection block.

[0016] Furthermore, electrical characteristic data, image data and text data of the photovoltaic module are obtained, and the method includes: collecting electrical characteristic data through a voltage sensor deployed on the photovoltaic module, capturing image data of the photovoltaic module through a camera, and collecting text data through daily maintenance records.

[0017] Further, preprocessing and encoding are performed, and the method includes:

[0018] Eliminate outliers in the electrical characteristic data, fill in missing data, and normalize the electrical characteristic data; use the LSTM algorithm to encode the preprocessed electrical characteristic data and extract the electrical characteristic feature vector;

[0019] The word embedding method is used to map the words in the text data into a high-dimensional vector space, and the word embedding matrix is ​​calculated; the word embedding matrix is ​​input into the bidirectional LSTM encoder to obtain the text feature vector;

[0020] Enhance the image data, remove the background, unify the size of the image data, remove noise and normalize it; perform CNN encoding on the preprocessed image data and extract the image feature vector.

[0021] Furthermore, the fusion features are used to train the damage detection model, and the loss function constructed by the cross entropy loss function and the mean square error loss function is designed to train the damage detection model.

[0022] A second aspect of the present invention provides a photovoltaic module damage detection system.

[0023] A photovoltaic module damage detection system, comprising:

[0024] A data acquisition and processing module is configured to: acquire electrical characteristic data, image data and text data of the photovoltaic module, and perform preprocessing and encoding to obtain an electrical characteristic feature vector, a text feature vector and an image feature vector;

[0025] A feature combination module is configured to: convert the electrical characteristic feature vector, the text feature vector and the image feature vector aligned in the time dimension into a multi-channel input tensor, input the multi-channel input tensor into a convolutional neural network, and perform feature combination on the output first feature, the second feature and the third feature to obtain a plurality of groups of feature combinations;

[0026] The data feature dynamic fusion module is configured as follows: based on a number of feature combinations, the data feature dynamic fusion module is used to obtain fusion features; wherein the data feature dynamic fusion module includes four feature data selection blocks and a global feedback network, each feature data selection block is used to select a feature combination, and the selected feature combination is input into the global feedback network to determine whether the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block have repeated combination selection, repeated data selection, or all three types of data have been selected; if not, the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block are weightedly fused to obtain fusion features; if so, the fusion features are output;

[0027] The damage detection module is configured to: judge the damage of the on-site photovoltaic components based on the fusion features and using the damage detection model.

[0028] A third aspect of the present invention provides a computer-readable storage medium.

[0029] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the photovoltaic module damage detection method as described in the first aspect above.

[0030] A fourth aspect of the present invention provides a computer device.

[0031] A computer device comprises 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 steps in the photovoltaic module damage detection method as described in the first aspect above are implemented.

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

[0033] The present invention ensures the richness and effectiveness of feature information by fusing the electrical characteristics, images and maintenance log text data of photovoltaic modules, improves the accuracy and robustness of the detection model, and more comprehensively detects various damages of photovoltaic modules.

[0034] The present invention realizes the combination of different data by utilizing the selection network, and then utilizes the data selection rules to select the most important feature combination from multiple combinations, and dynamically monitors the data selection and fusion through the global feedback network to ensure that the fused data with sufficient information is obtained.

[0035] The present invention adopts an automated intelligent detection method to replace traditional manual inspections, greatly improving detection efficiency, reducing labor costs and time costs, and making the maintenance of photovoltaic components more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] Figure 1 is a flow chart of a photovoltaic module damage detection method shown in the present invention;

[0038] Figure 2 is a structural diagram of a data preprocessing module shown in the present invention;

[0039] Figure 3 It is a framework diagram of the rule-based data selection method shown in the present invention;

[0040] Figure 4 It is a structural diagram of the data fusion dynamic module shown in the present invention;

[0041] Figure 5 is a structural diagram of a damage detection network training module shown in the present invention;

[0042] Figure 6 It is a structural diagram of the photovoltaic module damage detection system shown in the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0044] Embodiment 1

[0045] like Figure 1 As shown, this embodiment provides a photovoltaic module damage detection method, comprising the following steps:

[0046] Acquire electrical characteristic data, image data and text data of the photovoltaic module, perform preprocessing and encoding, and obtain electrical characteristic feature vectors, text feature vectors and image feature vectors;

[0047] The electrical characteristic feature vector, the text feature vector and the image feature vector aligned in the time dimension are converted into a multi-channel input tensor, the multi-channel input tensor is input into a convolutional neural network, and the first feature, the second feature and the third feature of the output are combined to obtain several groups of feature combinations;

[0048] Based on several groups of feature combinations, a data feature dynamic fusion module is used to obtain fusion features; wherein the data feature dynamic fusion module includes four feature data selection blocks and a global feedback network, each feature data selection block is used to select a feature combination, and the selected feature combination is input into the global feedback network to determine whether there is a combination selection duplication, data selection duplication or all three types of data have been selected between the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block; if not, the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block are weightedly fused to obtain a fusion feature; if so, the fusion feature is output;

[0049] Based on the fusion features, a damage detection model is used to judge the damage of on-site photovoltaic modules.

[0050] This embodiment includes: a data acquisition and processing module, a data feature dynamic fusion module, a damage detection network training module and a real-time detection module, such as Figure 6 shown.

[0051] Module 1: Data acquisition and processing module, such as Figure 2 shown.

[0052] 1. Data collection part

[0053] The data collection part is used to collect the electrical characteristic data, image data, and text data of photovoltaic modules. The electrical characteristic data is collected by deploying voltage sensors on photovoltaic modules. The collected data samples are timestamped and labeled with damage type labels to build a time-series electrical sample database. The image data is captured by high-resolution cameras deployed on site. The captured images must be matched with the electrical data in time, and the images are also labeled with damage type labels. The image database is then built using the collected images. The text data is collected from daily maintenance records. The maintenance records are stored together with the labels to build a text database.

[0054] 2. Data processing part

[0055] The data processing part is to perform preprocessing operations on the three types of collected data, encode different categories of data, and align the encoded three types of data in the time dimension.

[0056] (1) Data preprocessing: Remove outliers from the electrical characteristics data, fill in missing data, and normalize the data to obtain electrical characteristics time series data .

[0057] Maintenance log text data ,in, Representative The damage description and maintenance record information of each photovoltaic panel is used to assist in diagnosing the damage of the photovoltaic module.

[0058] Image data To collect infrared thermal imaging images of photovoltaic modules, first Data enhancement removes the background, unifies the image size, removes noise, and normalizes the image to obtain pre-processed image data .

[0059] (2) Data encoding: Encoding the three types of preprocessed data, extracting high-dimensional feature vectors from the three types of data, and setting high-dimensional vector spaces of the same dimension size.

[0060] The LSTM algorithm is used to process the preprocessed electrical characteristic timing data: Encode and extract its high-dimensional feature vector .

[0061] Using the Embedding word embedding method to maintain text log data Map the words in to a high-dimensional vector space and calculate the word embedding matrix , and then these word embeddings are fed into a bidirectional LSTM encoder to obtain the utterance-level language representation, i.e., the text feature vector .

[0062] The preprocessed image data is Use CNN for encoding operation to extract its high-dimensional feature vector .

[0063] (3) Data time alignment: Align the three types of data collected on the same photovoltaic panel Align in the time dimension to ensure the relevance between data. According to the time alignment principle, the timestamp is set to ,in , is the start time of data collection, is the end time of data collection. The data is at this timestamp. , the alignment operation is:

[0064] (1)

[0065] Module 2: Data feature dynamic fusion module

[0066] This module calls the data feature dynamic fusion method to dynamically fuse the three types of collected data. This method internally calls the rule-based data selection method four times, and controls the number of times the selected data combination is fused through the global feedback network, thereby obtaining fusion features with sufficient feature content.

[0067] 1. Rule-based data selection method

[0068] like Figure 3 As shown in the figure, the rule-based data selection method mainly selects and combines three types of data, using the selection network Combining three kinds of data, a total of eight combinations are generated, and a data selection rule is set to select a combination with a large number of feature contents from the combination results as the selection result. The method includes the following specific steps:

[0069] (1) Select a network Construction: The main function of the selection network is to make a preliminary combination of data features in the same feature space, and combine the three types of data through a lightweight convolutional neural network. First, merge the three types of data The feature vector of is a multi-channel input tensor , input it into the convolutional neural network structure for processing, and then the three types of data features output by the fully connected layer The main combination operations are simple mappings, which are 3 groups in total. , , ; There are 4 groups of summation combinations, namely , , , ; and weighted combinations totaling 1 group, ,in , is the weight coefficient, m=1, 2, 3.

[0070] (2) Setting of data feature selection rules: Based on the Lasso regression feature selection algorithm, the most important feature combination is selected from the eight combination results. The eight combinations are used as input, and their regularization characteristics are used to select the feature combination that contributes most to the damage detection target variable. In subsequent processing, in order to prevent the same selection results from being obtained by multiple feature selection operations on the same set of data, random perturbations are introduced to make the regularization strength of each regression operation slightly different, and possible different feature data combinations are selected, which better reflects the diversity of data feature selection:

[0071] (2)

[0072] in, It is The actual value of the samples, It is The sample The modal combination eigenvalues, It is The regression coefficients of the modal combinations are is the regularization parameter, which controls the strength of the regularization term. is the number of samples contained in the dataset, It follows a normal distribution of random disturbances.

[0073] According to the coefficient of Lasso regression, select the maximum non-zero coefficient The corresponding feature index value , and then select the feature combination that contributes most to the target variable .

[0074] 2. Dynamic fusion method of data features

[0075] like Figure 4 As shown in the figure, the dynamic fusion of data features mainly uses the global feedback network to determine whether there is a combination selection duplication, data selection duplication, or all three types of data have been selected in the feature data selection block (i.e., the rule-based data selection method). If such a situation occurs, the subsequent combination data selection and fusion are skipped, and the fused features are directly input into the training module for model training. This module contains the following specific contents:

[0076] (1) A global feedback network is used to monitor the three types of repeated selections that may occur during the dynamic fusion process. If any of the above situations occurs, the subsequent data selection and fusion process is skipped. The main steps of the network monitoring process are:

[0077] ① Initialize the selected recorder : Selection results of each feature data selection block Recorded in middle;

[0078] ②Record selection results: Assume that the first set of feature data selection blocks is selected , i.e. the recorder , the second set of feature data selection blocks selects , then the recorder , enter the selection result detection phase;

[0079] ③Result detection stage: detection Check whether there are repeated combination selections, repeated data selections, or all three types of data have been selected:

[0080] (3)

[0081] The test content must include testing Whether Same, also need to detect the selection results Data in Whether Included, but also need to detect and All data in If duplication is detected, go to step ④; if no duplication is detected, go back to step ② to continue recording the selection results of the third set of feature data selection blocks, and then go to step ③ for duplication detection.

[0082] ④ Stop operation: If repeated selection is detected, stop subsequent data selection and fusion, and input the fused features into model training.

[0083] The three types of repeated selection are shown in the following three examples. The feature data selection block and the attention block are taken as one unit, and a total of four units are set:

[0084] Case 1: Units 1 and 2 are both selected , in this case, the combination selects repetition and will not enter the attention block of unit 2, unit 3 and 4; in the second case, unit 1 selects , Unit 2 selected , while unit 3 selected , at this time, the data selection is repeated and will not enter the attention block of unit 3 and unit 4; the third case: unit 1 selects , Unit 2 selected , while unit 3 selected , at this time, all three types of data have been selected and will not enter unit 4. In most cases, data fusion is completed in unit 3. In order to prevent the situation where only one type of data is selected in each unit, unit 4 is set up for auxiliary selection.

[0085] (2) The data feature fusion process uses the self-attention algorithm in the attention block to perform weighted fusion on the selection results of the feature data selection block. The selection results of the feature data selection block of unit 1 are directly input into its attention block for feature fusion. The other three units need to input the features selected by this unit into the attention block, and at the same time input the fused features output by the previous unit into the attention block of this unit, and fuse the two parts of features. In the last three units, the fused data of the previous unit is input into the feature data selection block of the current unit, but this fused data only participates in the feature fusion process of the attention block and does not participate in the data selection process.

[0086] Module 3: Damage detection network training module, such as Figure 5 shown.

[0087] This module implements the training of the photovoltaic module damage detection network, inputs the multimodal fusion data into the Transformer network for training, and obtains a damage detection network module with good detection effect by updating the network model weights multiple times. The specific implementation steps of this module are as follows:

[0088] (1) Dataset division: The fused data is divided into a training set and a test set. The training set is used for model training, and the test set is used for model verification and evaluation.

[0089] (2) Transformer network model construction and training: Use fused data to train the Transformer network model, calculate the loss during the training process, update the model's weight parameters, and build a better Transformer network model. Calculate model damage By using a combination of loss functions, including cross entropy loss and mean square error loss :

[0090] (4)

[0091] in, is the true label, is the prediction result, and Weight coefficients of the two loss functions.

[0092] (3) Model testing and evaluation: Use the test set to test the model and and the true label Analyze and calculate the loss , and then update the model weights in the training process until the model loss drops to the ideal value, and then obtain a photovoltaic module damage detection network model with higher detection accuracy.

[0093] Module 4: Real-time detection module

[0094] The real-time detection module uses a trained damage detection model to judge the damage of photovoltaic modules on site. Data preprocessing is performed by collecting three types of data of photovoltaic modules on site or retrieving data collected daily and stored in three types of databases, and the processed data is input into the real-time diagnosis module of photovoltaic modules for real-time damage detection. The damage detection results are stored as text data and uploaded to the database. If damage is detected, an alarm is issued, and the operation and maintenance personnel perform maintenance and upload the maintenance results to the database; if there is no damage, it is uploaded to the database as the daily inspection and maintenance text data of the tested photovoltaic module, and reserved for the next inspection.

[0095] Embodiment 2

[0096] This embodiment provides a photovoltaic module damage detection system.

[0097] A photovoltaic module damage detection system, comprising:

[0098] A data acquisition and processing module is configured to: acquire electrical characteristic data, image data and text data of the photovoltaic module, and perform preprocessing and encoding to obtain an electrical characteristic feature vector, a text feature vector and an image feature vector;

[0099] A feature combination module is configured to: convert the electrical characteristic feature vector, the text feature vector and the image feature vector aligned in the time dimension into a multi-channel input tensor, input the multi-channel input tensor into a convolutional neural network, and perform feature combination on the output first feature, the second feature and the third feature to obtain a plurality of groups of feature combinations;

[0100] The data feature dynamic fusion module is configured as follows: based on a number of feature combinations, the data feature dynamic fusion module is used to obtain fusion features; wherein the data feature dynamic fusion module includes four feature data selection blocks and a global feedback network, each feature data selection block is used to select a feature combination, and the selected feature combination is input into the global feedback network to determine whether the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block have repeated combination selection, repeated data selection, or all three types of data have been selected; if not, the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block are weightedly fused to obtain fusion features; if so, the fusion features are output;

[0101] The damage detection module is configured to: judge the damage of the on-site photovoltaic components based on the fusion features and using the damage detection model.

[0102] In some embodiments, the feature combination module is further configured to: map the first feature, the second feature, and the third feature to obtain three groups of mapped feature combinations, sum the first feature, the second feature, and the third feature in pairs and sum them together to obtain four groups of summed feature combinations; set weight coefficients for the first feature, the second feature, and the third feature, perform weighted summation, and obtain a group of weighted feature combinations.

[0103] In some embodiments, the data feature dynamic fusion module is further configured as follows: the feature data selection block adopts the Lasso regression feature selection algorithm, introduces random disturbance, calculates the coefficient of Lasso regression, and selects the feature combination corresponding to the maximum value of the Lasso regression coefficient.

[0104] In some embodiments, the data feature dynamic fusion module is further configured as follows: the data feature dynamic fusion module also includes four attention blocks, with one attention block after each feature data selection block, inputting the feature combination selected by the previous feature data selection block into the previous attention block, inputting the feature combination selected by the current feature data selection block into the current attention block, inputting the fused features output by the previous feature data selection block into the current attention block, and performing weighted fusion with the feature combination selected by the current feature data selection block.

[0105] In some embodiments, the data acquisition and processing module is further configured to: obtain electrical characteristic data, image data and text data of the photovoltaic component, and the method includes: collecting electrical characteristic data through a voltage sensor deployed on the photovoltaic component, capturing image data of the photovoltaic component through a camera, and collecting text data through daily maintenance records.

[0106] In some embodiments, the data acquisition and processing module is further configured to: remove abnormal values ​​of the electrical characteristic data, fill in missing data, and normalize the electrical characteristic data; use the LSTM algorithm to encode the preprocessed electrical characteristic data and extract the electrical characteristic feature vector;

[0107] The word embedding method is used to map the words in the text data into a high-dimensional vector space, and the word embedding matrix is ​​calculated; the word embedding matrix is ​​input into the bidirectional LSTM encoder to obtain the text feature vector;

[0108] Enhance the image data, remove the background, unify the size of the image data, remove noise and normalize it; perform CNN encoding on the preprocessed image data and extract the image feature vector.

[0109] In some embodiments, the system also includes a damage detection network training module, which is configured to: use fusion features to train a damage detection model, and design a loss function constructed by a cross entropy loss function and a mean square error loss function to train the damage detection model.

[0110] Embodiment 3

[0111] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the photovoltaic module damage detection method described in the above-mentioned embodiment 1 are implemented.

[0112] Embodiment 4

[0113] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the photovoltaic module damage detection method described in the first embodiment are implemented.

[0114] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A photovoltaic module damage detection method, characterized in that: include: Acquire electrical characteristic data, image data and text data of the photovoltaic module, perform preprocessing and encoding, and obtain electrical characteristic feature vectors, text feature vectors and image feature vectors; The electrical characteristic feature vector, the text feature vector and the image feature vector aligned in the time dimension are converted into a multi-channel input tensor, the multi-channel input tensor is input into a convolutional neural network, and the first feature, the second feature and the third feature of the output are combined to obtain several groups of feature combinations; Based on several groups of feature combinations, a data feature dynamic fusion module is used to obtain fusion features; wherein the data feature dynamic fusion module includes four feature data selection blocks and a global feedback network, each feature data selection block is used to select a feature combination, and the selected feature combination is input into the global feedback network to determine whether there is a combination selection duplication, data selection duplication or all three types of data have been selected between the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block; if not, the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block are weightedly fused to obtain a fusion feature; if so, the fusion feature is output; Based on the fusion features, a damage detection model is used to judge the damage of on-site photovoltaic modules.

2. The photovoltaic module damage detection method according to claim 1, characterized in that: The method of combining the output first feature, second feature and third feature to obtain several groups of feature combinations includes: mapping the first feature, the second feature and the third feature to obtain three groups of mapping feature combinations, summing the first feature, the second feature and the third feature in pairs and summing the three features together to obtain four groups of summed feature combinations; setting weight coefficients of the first feature, the second feature and the third feature, performing weighted summation, and obtaining a group of weighted feature combinations.

3. The photovoltaic module damage detection method according to claim 1, characterized in that: The feature data selection block adopts the Lasso regression feature selection algorithm, introduces random disturbance, calculates the coefficient of the Lasso regression, and selects the feature combination corresponding to the maximum value of the coefficient of the Lasso regression.

4. The photovoltaic module damage detection method according to claim 1, characterized in that: The data feature dynamic fusion module also includes four attention blocks, one of which is after each feature data selection block. The feature combination selected by the previous feature data selection block is input into the previous attention block, the feature combination selected by the current feature data selection block is input into the current attention block, and the fused features output by the previous feature data selection block are input into the current attention block and weightedly fused with the feature combination selected by the current feature data selection block.

5. The photovoltaic module damage detection method according to claim 1, characterized in that: The electrical characteristic data, image data and text data of a photovoltaic module are obtained, and the method includes: collecting electrical characteristic data through a voltage sensor deployed on the photovoltaic module, taking image data of the photovoltaic module through a camera, and collecting text data through daily maintenance records.

6. The photovoltaic module damage detection method according to claim 1, characterized in that: Preprocessing and encoding are performed, including: Eliminate outliers in the electrical characteristic data, fill in missing data, and normalize the electrical characteristic data; use the LSTM algorithm to encode the preprocessed electrical characteristic data and extract the electrical characteristic feature vector; The word embedding method is used to map the words in the text data into a high-dimensional vector space, and the word embedding matrix is ​​calculated; the word embedding matrix is ​​input into the bidirectional LSTM encoder to obtain the text feature vector; Enhance the image data, remove the background, unify the size of the image data, remove noise and normalize it; perform CNN encoding on the preprocessed image data and extract the image feature vector.

7. The photovoltaic module damage detection method according to claim 1, characterized in that: The damage detection model is trained by using fusion features, and a loss function constructed by the cross entropy loss function and the mean square error loss function is designed to train the damage detection model.

8. A photovoltaic module damage detection system, characterized in that: include: A data acquisition and processing module is configured to: acquire electrical characteristic data, image data and text data of the photovoltaic module, and perform preprocessing and encoding to obtain an electrical characteristic feature vector, a text feature vector and an image feature vector; A feature combination module is configured to: convert the electrical characteristic feature vector, the text feature vector and the image feature vector aligned in the time dimension into a multi-channel input tensor, input the multi-channel input tensor into a convolutional neural network, and perform feature combination on the output first feature, the second feature and the third feature to obtain a plurality of groups of feature combinations; The data feature dynamic fusion module is configured as follows: based on a number of feature combinations, the data feature dynamic fusion module is used to obtain fusion features; wherein the data feature dynamic fusion module includes four feature data selection blocks and a global feedback network, each feature data selection block is used to select a feature combination, and the selected feature combination is input into the global feedback network to determine whether the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block have repeated combination selection, repeated data selection, or all three types of data have been selected; if not, the feature combination selected by the previous feature data selection block and the feature combination selected by the current feature data selection block are weightedly fused to obtain fusion features; if so, the fusion features are output; The damage detection module is configured to: judge the damage of the on-site photovoltaic components based on the fusion features and using the damage detection model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the photovoltaic component damage detection method as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the photovoltaic component damage detection method according to any one of claims 1 to 7 are implemented.

Images