Mechanism Model and Image Fusion Driven Photovoltaic Panel Defect Detection Method and System

Abstract

The present invention proposes a photovoltaic panel defect detection method and system driven by the fusion of a mechanism model and an image. The method includes: collecting photovoltaic panel image data and monitoring data at the same timestamp; the monitoring data includes illumination intensity, ambient temperature, wind speed, current, and voltage information; calculating current-voltage experimental data according to the monitoring data; inputting the image data and the current-voltage experimental data into a trained hierarchical recognition model for explicit and implicit defects to obtain the recognition results of explicit defects and implicit defects of the photovoltaic panel. The hierarchical recognition model for explicit and implicit defects is designed in series, enabling the model to be more targeted when performing defect recognition. Only when the photovoltaic panel is initially judged to have defects will the explicit defect recognition network and the implicit defect recognition network be further called for processing, thus avoiding the indiscriminate processing of all photovoltaic panel-related data and saving computing resources and time.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent defect detection, and particularly to a photovoltaic panel defect detection method and system driven by the fusion of a mechanism model and an image. Background Art

[0002] With the rapid development of the photovoltaic industry, the usage of photovoltaic panels is increasing day by day. However, during long-term operation, photovoltaic panels are prone to be affected by environmental factors, manufacturing defects, aging, etc., resulting in a decline in their performance and the occurrence of failures. Therefore, timely and accurately detecting and diagnosing the defects of photovoltaic panels to ensure their efficient and stable operation has become an important task in the maintenance of photovoltaic power generation systems. Traditional photovoltaic panel detection methods mainly rely on manual inspection and simple instrument detection, which are not only inefficient but also difficult to accurately identify the types of subtle defects of photovoltaic panels. Therefore, researching intelligent and efficient photovoltaic panel defect detection methods has become a key requirement for the maintenance of photovoltaic power generation systems.

[0003] Most existing intelligent detection methods adopt defect detection methods driven by a single data type. When only image data is used, internal defects of photovoltaic panels cannot be detected; if only electrical data is relied on for defect diagnosis, the changes in electrical data caused by obstacles (such as dust, leaves, shadows, etc.) on the surface or near the photovoltaic panel may interfere with or mislead the accurate diagnosis of internal defects of the photovoltaic panel. Therefore, there is an urgent need for a method that can accurately detect the defects of photovoltaic panels. Summary of the Invention

[0004] To solve the above problems, the present invention proposes a photovoltaic panel defect detection method and system driven by the fusion of a mechanism model and an image. The method constructs a defect recognition model composed of a recessive defect recognition network and a dominant defect recognition network based on explicit and recessive hierarchical decision-making. By using the current-voltage data output by the photovoltaic panel mechanism model, the actual current-voltage data, and the photovoltaic panel image data as the input data for training and detection, and introducing the constraints of the mechanism model to guide different networks to perform training and detection of defect recognition, various defects of the photovoltaic panel can be accurately recognized.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] In the first aspect, the present invention provides a photovoltaic panel defect detection method driven by the fusion of a mechanism model and an image, including:

[0007] Collecting photovoltaic panel image data and monitoring data at the same time stamp; the monitoring data includes light intensity, ambient temperature, wind speed, current, and voltage information;

[0008] Calculating current-voltage experimental data based on the monitoring data;

[0009] Input the image data and current-voltage experimental data into a trained layered recognition model for explicit and implicit defects to obtain recognition results of explicit and implicit defects of photovoltaic panels;

[0010] Among them, the hierarchical identification model of explicit and implicit defects includes a defect classification network, an explicit defect recognition network and a implicit defect recognition network connected in series in sequence; the defect classification network is used to identify whether there are defects in the photovoltaic panel based on current-voltage experimental data; if there are defects, the image data corresponding to the current-voltage experimental data is input into the explicit defect recognition network to identify the explicit defects; the current-voltage experimental data corresponding to the image data without explicit defects is input into the implicit defect recognition network to identify the implicit defects.

[0011] Preferably, the calculating of current-voltage experimental data according to the monitoring data specifically includes:

[0012] According to the photovoltaic panel mechanism, the theoretical current-voltage data is calculated based on the light intensity, ambient temperature and wind speed information;

[0013] Based on the current and voltage information, calculate the actual current-voltage data;

[0014] Subtract the theoretical current-voltage data from the actual current-voltage data to obtain the deviation current-voltage data;

[0015] The theoretical current-voltage data and the deviation current-voltage data are added together to obtain the current-voltage experimental data.

[0016] Preferably, the method of calculating the current-voltage experimental data based on the monitoring data and inputting the image data and the current-voltage experimental data into the trained layered recognition model for explicit and implicit defects also includes: performing preprocessing operations of image enhancement, position correction, and unifying the image size on the image data; and performing preprocessing of data cleaning and data denoising on the monitoring data.

[0017] Preferably, the dominant defects include cracks, breakages and shielding; the recessive defects include hidden cracks, current leakage and PID effect.

[0018] Preferably, the defect classification network is a binary classification network constructed based on a residual network; the training process of the defect classification network includes:

[0019] Inputting the current-voltage experimental data samples into the defect classification network, and training based on the non-defect label and the defect label; the defect label includes an explicit defect label and an implicit defect label; the explicit defect label includes a crack label, a damage label and an occlusion label, and the implicit defect label includes a hidden crack label, a current leakage label and a PID effect label;

[0020] Calculate the first loss function. When the first loss function is minimized, the training of the defect classification network is completed.

[0021] Preferably, the training process of the dominant defect recognition network includes:

[0022] Obtain the image data samples corresponding to the current-voltage experimental data samples recognized as having defects by the defect classification network;

[0023] Input the image data samples into the dominant defect recognition network, take the overall recessive defect label as the dominant non-defect label, and perform training based on the dominant defect label and the dominant non-defect label;

[0024] Calculate the second loss function. When the second loss function is minimized, the training of the dominant defect recognition network is completed.

[0025] Preferably, the training process of the recessive defect recognition network includes:

[0026] Obtain the current-voltage experimental data samples corresponding to the image data samples recognized as dominant non-defects by the dominant defect recognition network;

[0027] Input the current-voltage experimental data samples into the recessive defect recognition network and perform training based on the recessive defect label;

[0028] Calculate the third loss function. When the third loss function is minimized, the training of the recessive defect recognition network is completed.

[0029] In a second aspect, the present invention provides a photovoltaic panel defect detection system driven by the fusion of a mechanism model and an image, including:

[0030] A data acquisition module for collecting photovoltaic panel image data and monitoring data at the same time stamp; the monitoring data includes light intensity, environmental temperature, wind speed, current, and voltage information;

[0031] A photovoltaic panel mechanism module for calculating current-voltage experimental data based on the monitoring data;

[0032] A defect recognition module for inputting the image data and the current-voltage experimental data into the trained dominant and recessive defect hierarchical recognition model to obtain the recognition results of the dominant and recessive defects of the photovoltaic panel;

[0033] Among them, the dominant and recessive defect hierarchical recognition model includes a defect classification network, a dominant defect recognition network, and a recessive defect recognition network connected in series in sequence; the defect classification network is used to identify whether there are defects in the photovoltaic panel based on the current-voltage experimental data; if there are defects, the image data corresponding to the current-voltage experimental data is input into the dominant defect recognition network to identify dominant defects; the current-voltage experimental data corresponding to the image data without dominant defects is input into the recessive defect recognition network to identify recessive defects.

[0034] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the steps in the method for detecting photovoltaic panel defects driven by a mechanism model and image fusion described in the first aspect.

[0035] In a fourth aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the method for detecting photovoltaic panel defects driven by a mechanism model and image fusion described in the first aspect.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] 1. The present invention constructs a photovoltaic panel mechanism model through environmental data, fuses electrical data and image data, and introduces the constraints of the photovoltaic panel mechanism model in the process of deep neural network training, enhancing the model's understanding of the physical process and realizing the detection of photovoltaic panel defects driven by the model and image fusion;

[0038] 2. The present invention uses the dominant and recessive hierarchical decision-making method to first detect the visible dominant defect types of the photovoltaic panel, and then detect the internal recessive defect types of the photovoltaic panel, which well solves the situation where the electrical data of the dominant defects is misjudged as the recessive defect type, and the image data of the recessive defects is misjudged as no defects, greatly improving the model detection accuracy;

[0039] 3. The present invention realizes an automated photovoltaic panel defect detection process, greatly improving the detection efficiency, significantly reducing the dependence on manual detection, and improving the operation and management efficiency of large-scale photovoltaic power stations.

[0040] The advantages of the additional aspects of the present invention will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings forming a part of this specification are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute a limitation to the present invention.

[0042] Figure 1 The main flowchart of a photovoltaic panel defect detection method driven by the fusion of a mechanism model and an image provided by an embodiment of the present invention;

[0043] Figure 2 The structure diagram of the dominant and recessive defect hierarchical recognition model provided by an embodiment of the present invention;

[0044] Figure 3 The schematic diagram of a photovoltaic panel defect detection system driven by the fusion of a mechanism model and an image provided by an embodiment of the present invention. Specific implementation manners

[0045] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0046] Embodiment 1

[0047] As Figure 1 shown, this embodiment discloses a photovoltaic panel defect detection method driven by the fusion of a mechanism model and an image, including the following steps:

[0048] S1: Collect the photovoltaic panel image data and monitoring data at the same timestamp; the monitoring data includes light intensity, ambient temperature, wind speed, current, and voltage information;

[0049] S2: Collect the photovoltaic panel image data and monitoring data at the same timestamp; the monitoring data includes light intensity, ambient temperature, wind speed, current, and voltage information;

[0050] S3: Input the image data and current-voltage experimental data into the trained dominant and recessive defect hierarchical recognition model to obtain the recognition results of the dominant and recessive defects of the photovoltaic panel;

[0051] Among them, the dominant and recessive defect hierarchical recognition model includes a defect classification network, a dominant defect recognition network, and a recessive defect recognition network connected in series in sequence; the defect classification network is used to identify whether there are defects in the photovoltaic panel based on the current-voltage experimental data; if there are defects, input the image data corresponding to the current-voltage experimental data into the dominant defect recognition network to identify the dominant defects; input the current-voltage experimental data corresponding to the image data without dominant defects into the recessive defect recognition network to identify the recessive defects.

[0052] Next, in combination with Figure 1 , a photovoltaic panel defect detection method driven by the fusion of a mechanism model and an image disclosed in this embodiment will be described in detail.

[0053] (1) Data collection

[0054] The data acquisition module is used to collect the real-time operation data of photovoltaic panels, including image data and monitoring data, where the monitoring data includes data such as light intensity, ambient temperature, wind speed, etc. For the collected image data and monitoring data, it is necessary to ensure that the data is aligned to a unified timestamp.

[0055] 1. Image acquisition: Use a drone equipped with a high-resolution camera to collect the image data of the on-site photovoltaic panels. Perform preprocessing operations on the collected images, such as image enhancement, position correction, and unifying the image size, and then summarize the image X image into the photovoltaic panel defect recognition database.

[0056] 2. Monitoring data acquisition: Mainly responsible for collecting the real-time data of the operating environment of photovoltaic panels, including parameters such as light intensity, ambient temperature, wind speed, current, and voltage, and performing preprocessing such as data cleaning and data denoising on the collected monitoring data.

[0057] Among them, the light intensity data G uses a light meter sensor to measure the solar radiation intensity, which is installed at a position parallel to the photovoltaic panel to ensure that the measured light intensity is consistent with the light intensity received by the photovoltaic panel.

[0058] The ambient temperature data T uses a temperature sensor to measure the ambient temperature around the photovoltaic panel, and the selection of its installation position should avoid situations such as direct sunlight and human interference.

[0059] The wind speed data W uses an anemometer to measure the wind speed around the photovoltaic panel, and its installation position should be in the development area to avoid being affected by photovoltaic panels, buildings, and other obstacles.

[0060] The current I and voltage V data use current sensors and voltage sensors to measure the output current and voltage of the photovoltaic panel. The sensors are directly installed at the output end of the photovoltaic panel to record the current and voltage data in real time.

[0061] (2) Labeling

[0062] Before performing the labeling task, set the labeling strategy according to the types and characteristics of photovoltaic panel defects, and classify the defects into obvious defects and hidden defects.

[0063] The obvious defects mainly include three categories: cracks, breakages, and occlusions, which are composed of the defect types visualized in the photovoltaic panel images. Label the existing defect types on the collected visible light image samples of the photovoltaic panels to form a training dataset. The labelers use the labeling tool to interactively label the obvious defect type labels existing in the photovoltaic panels.

[0064] The latent defects mainly include three categories: hidden cracks, current leakage, and PID effect. Their visible light images are normal without defects, and the internal defects cannot be identified. Therefore, the current and voltage data in the monitoring data can be used to draw the actual current-voltage curve, and the types of internal defects existing in the photovoltaic panel are marked as the latent defect type labels.

[0065] For the same image, it is labeled multiple times, and the latent labels are no longer labeled for those already labeled with dominant labels. The labeled images are saved in the same format and stored in the photovoltaic panel defect recognition annotation database together with the labels. Since the monitoring data and the image data are already aligned in time, the defect labels for the images are used as the labels for the monitoring data.

[0066] The data with category labels stored in the annotation database is used as training data and input into the defect recognition module of the deep neural network to train the defect recognition model.

[0067] (3) Photovoltaic panel mechanism model

[0068] The photovoltaic panel mechanism model module constructs a mechanism model based on the physical characteristics and working principles of the photovoltaic panel. The mechanism model of the photovoltaic panel is mainly constructed through information such as light intensity (G), ambient temperature (T), and wind speed (W), and the current-voltage characteristic curve under normal working conditions of the photovoltaic panel is calculated.

[0069] For the photovoltaic mechanism model, a single-diode model is used for construction, which is mainly composed of photocurrent I ph , diode saturation current I0, series resistance R s , parallel resistance R sh and diode ideality factor n to describe the output characteristics of the photovoltaic panel. The output current I m and output voltage V m of the photovoltaic panel satisfy the following equation:

[0070]

[0071] Among them, is the thermal voltage, k is the Boltzmann constant, and q is the electron charge.

[0072] The calculation processes of the five parameters are as follows:

[0073] 1. First, calculate the working temperature T c of the photovoltaic panel, which is affected by light intensity, ambient temperature, and wind speed, and the calculation is as follows:

[0074]

[0075] Among them, G ref is the light intensity under standard test conditions (STC), usually taking a value of 1000 W / m 2; T NOCT is the nominal operating temperature, measured under the conditions of an ambient temperature of 22 °C, a light intensity of 800 W / m 2 , and no wind; τ is the wind speed influence coefficient, indicating the influence of wind speed on the temperature of the photovoltaic panel.

[0076] 2. Calculate the five parameters I ph , I0, R s , R sh and n respectively under STC conditions. Among them, the photocurrent I ph and the series resistance R s are related to the light intensity G and the operating temperature T c ; the diode saturation current I0 and the diode ideality factor n are related to the operating temperature T c ; the shunt resistance R sh is related to the light intensity G.

[0077] 3. Substitute the five parameters into the current-voltage equation of the photovoltaic panel mechanism model to calculate the theoretical current-voltage curve of the photovoltaic panel under the input of light intensity, ambient temperature and wind speed data at a certain moment.

[0078] (IV) Dominant and recessive defect hierarchical recognition model

[0079] This embodiment constructs a defect recognition model composed of a recessive defect recognition network and a dominant defect recognition network based on the dominant and recessive hierarchical decision-making method. By using the current-voltage data, actual current-voltage data and photovoltaic panel image data output by the photovoltaic panel mechanism model as the input data for training and detection, adopting the dominant and recessive hierarchical decision-making method, and introducing the constraints of the mechanism model, different networks are guided to perform training and detection of defect recognition, and then various defects of the photovoltaic panel are accurately recognized. The overall process is as Figure 2 shown:

[0080] 1. Defect classification network and training

[0081] The defect classification network adopts an architecture of a mechanism-constrained residual network (MCResNet), that is, the mechanism rules of the mechanism model are embedded as constraints into the training process of ResNet.

[0082] The defect classification network includes a defect recognition binary classification unit for distinguishing between a non-defective photovoltaic panel θ normal and a defective photovoltaic panel θ abnormal . Among them, the label corresponding to the non-defective photovoltaic panel θ normal is represented by y = 0, and the label corresponding to the defective photovoltaic panel θ abnormal is represented by y = 1. The category of the defective photovoltaic panel θ abnormal includes three types of dominant defects {θ d1 , θ d2 , θd3}, three types of hidden defects {θ r1 , θ r2 , θ r3}, and one type of dominant non - defect and hidden defect (i.e., the case classified as non - defect in the dominant defect recognition network, which includes three types of hidden defect types); where, θ d1 , θ d2 , θ d3 are crack, breakage, and occlusion respectively, and the corresponding labels are represented by y = 2, 3, 4; is dominant non - defect and hidden defect, and the corresponding label is represented by y = 5; θ r1 , θ r2 , θ r3 are hidden crack, current leakage, and PID effect respectively, and the corresponding labels are represented by y = 6, 7, 8.

[0083] The specific training steps are as follows:

[0084] S321: Deviation data processing. According to the output current I m and output voltage V m of the mechanism model, draw the current - voltage curve l m , fuse the theoretical current - voltage curve l m and the actually measured current - voltage curve l0. First, calculate the deviation current - voltage data l res = l m - l0, then add the deviation current - voltage data l res to the output current - voltage data l m of the mechanism model to obtain the current - voltage experimental data l = l res + l m , and divide l into a training set l train and a test set l test .

[0085] It should be understood that the calculations related to the curve can be achieved by those skilled in the art.

[0086] S322: Construct a residual network model, initialize the network weights, and input the training current - voltage data l train into the defect recognition binary classification unit for training, and output the classification result.

[0087] S323: Calculate the loss function to update the ResNet model parameters, and at the same time embed the photovoltaic panel mechanism model equation as a constraint into the network model training to make the network model satisfy the physical laws. The reconstructed loss function L IV includes the network model loss term and the mechanism model loss term. The calculation method of the first loss function is as follows:

[0088]

[0089] Among them, λ1 and λ2 are the weights of the loss term of the network model and the loss term of the mechanism model; N is the number of samples; C is the number of categories, a total of two categories, namely no defect of the photovoltaic panel and defect of the photovoltaic panel; is the classification probability; y ij is the true fault type label of the i-th sample. If sample i belongs to the j-th category, then y ij = 1, otherwise y ij = 0; 1(y i = j) is the indicator function, and y i is the true label of the i-th sample. When y i = c, the value is 1, otherwise it is 2; l res_i is the deviation current-voltage curve of the i-th sample. The deviation data is used to make a difference, that is, only the relative change amount is considered to prevent the value obtained by the loss term of the mechanism model from being too large; l target_c is the target deviation current-voltage curve corresponding to the category c; it is defined according to prior knowledge or physical laws, and l target_c can take l target_0 and l target_1 : When l res_i is a sample without defects, l target_0 is selected for calculation, and l target_0 is the target deviation current-voltage curve of the no-defect category of the photovoltaic panel; when l res_i is a defective sample, l target_1 is selected for calculation, and l target_1 is the target deviation current-voltage curve of the defect category of the photovoltaic panel.

[0090] In this embodiment, detailed consideration is given to the selection of current-voltage data. For the samples in network training, current-voltage experimental data is required. Because when the photovoltaic panel has no defects, the actually measured current-voltage curve and the theoretical current-voltage curve generated by the mechanism model may be the same, and the difference after subtracting the deviation may be 2. The data of 2 cannot help the network training purpose of the training network. Therefore, the experimental data is the ideal data generated by the mechanism model plus the deviation data. For the calculation of the loss function, the purpose is to make a difference with a target deviation data to form the loss term of the mechanism constraint. Therefore, there is no need to use experimental data, and the deviation data can be used to complete the calculation of the mechanism constraint loss; at the same time, the deviation data has a small value, which can effectively prevent the loss of the mechanism constraint term from being too large, thus affecting the adjustment and update of the model parameters by the loss function.

[0091] In this embodiment, by using the deviation data curve between the output of the mechanism model and the actually measured value, the loss term of the mechanism model is constructed, the mechanism constraint is introduced, and it is better adapted to the classification task and the training network model.

[0092] S324: Use test set l test Evaluate the trained model to determine whether the defect classification network can correctly classify defective items based on six defect categories, and obtain the defect detection classification results.

[0093] In this embodiment, the current-voltage curve output by the mechanism model is the ideal value generated under the current environment. By performing deviation processing on the actually measured current-voltage data, it is equivalent to realizing the fusion of the two types of data, and obtaining current-voltage curve data with rich characteristic information.

[0094] Furthermore, by generating an ideal value using the mechanism model under the current environment, this value can be used as a reference value under the current environment and space-time. By performing deviation processing on the measured value, a deviation value is obtained, and then the deviation value is added to the ideal value and input into the classification detection model, making the input characteristic information more abundant and contributing to defect detection.

[0095] In this embodiment, the photovoltaic panel mechanism model equations are constructed based on physical principles and the working principle of the photovoltaic panel, and can accurately describe the behavior of the photovoltaic panel under different conditions. Embedding these equations as constraints into the network model can ensure that the network model follows physical laws while learning. Since the network model is constrained by physical laws during the training process, it can reduce prediction errors caused by data noise, model overfitting, etc., making the prediction results of the model more reliable and improving the physical accuracy of the model output.

[0096] 2. Explicit defect recognition network and training

[0097] When the defect classification network classifies a defective type, activate the explicit defect recognition network to perform defect detection on the photovoltaic panel image data. For the explicit defect recognition network, use the Vision Transformer (ViT) model architecture, and use the labeled image data to train the network model. The specific training steps are as follows:

[0098] S311: Preprocess the image data, including operations such as removing the background, adjusting the image size, and normalizing, and divide the processed images into a training set and a test set. Use the three true labels {θ r1 , θ r2 , θ r3} corresponding to the latent defects as a new label, that is This label corresponds to the explicit non-defective classification result of the training images.

[0099] S312: Construct a ViT network model, initialize the network parameters, input the image training set into the network for training, and obtain the classification results.

[0100] S313: Calculate the model loss function and update the model parameters. Using the cross - entropy loss function as the optimization objective, the second loss function is calculated as follows:

[0101]

[0102] where G is the number of categories, and there are four categories in total M is the number of samples; is the classification probability; y pq is the true label.

[0103] S314: Use the image test set to evaluate the trained model and obtain the defect detection classification results. When the dominant defect recognition network classifies an image as non - defective, activate the recessive defect recognition multi - classification network, and input the current - voltage experimental data corresponding to this image into the recessive defect recognition multi - classification network to detect the recessive defects of the photovoltaic panel.

[0104] 3. Recessive Defect Recognition Network and Training

[0105] The structure of the recessive defect recognition network is similar to that of the defect classification network, except that the defect recognition binary classification unit is replaced by a recessive defect recognition multi - classification unit.

[0106] The recessive defect recognition multi - classification unit performs multi - classification tasks corresponding to three recessive defect categories {θ r1 , θ r2 , θ r3}. The corresponding third loss function is as follows:

[0107]

[0108] where K is the number of samples; C is the number of categories, which are three in total, namely hidden crack, current leakage, and PID effect; l target_c is the target deviation current - voltage curve for the corresponding category c, defined based on prior knowledge or physical laws, l target_6 is the target deviation current - voltage curve for the hidden crack category, l target_7 is the target deviation current - voltage curve for the current leakage category, l target_8 is the target deviation current - voltage curve for the PID effect category.

[0109] The specific implementation process will not be elaborated. Similarly, use the test set l test to evaluate the trained model and evaluate whether the recessive defect recognition network can accurately classify the three recessive defect categories to obtain the defect detection classification results.

[0110] The defect recognition binary classification unit of the defect classification network and the hidden defect recognition multi-classification network in this embodiment are used to solve the situation where the current-voltage curves of obvious defects and hidden defects are similar, and to avoid misidentifying defects caused by similar curves. Since obvious defects and hidden defects may exhibit similar characteristics in the current-voltage curve, traditional single classification methods are prone to misjudgment. By introducing a multi-classification hidden defect recognition network in this embodiment, different types of hidden defects can be more carefully distinguished on the basis of determining that there are no obvious defects, reducing the recognition errors caused by similar curves and improving the accuracy of defect recognition.

[0111] As a specific implementation method, the explicit and implicit decision fusion method mainly uses the explicit and implicit defect recognition network to perform hierarchical decision-making to obtain classification results. The photovoltaic panel defect detection task has obvious defects and hidden defects. First, overall defect detection is performed to detect whether there are defects in the photovoltaic panel. If there are defects, the defect data is subjected to explicit detection, and the detection stops when the detection result is an obvious defect. If no obvious defect is detected, the hidden defect is further detected. The main steps of this method are as follows:

[0112] (1) Overall defect detection: First, use the defect classification network to detect defects in the X IV data. The classification categories are no defect θ normal and defective θ abnormal ={θ d1 ,θ d2 ,θ d3 ,θ r1 ,θ r2 ,θ r3}. If the detection result is no defect, the decision result y binary is output. If there are defects, the obvious defect recognition network is activated;

[0113] (2) Obvious defect detection: Input the image data X IV_abnormal corresponding to the defective X image data into the obvious defect recognition network for defect detection. If the detection result is defective θ' normal ={θ d1 ,θ d2 ,θ d3}, it is one of the three types of obvious defects {θ d1 ,θ d2 ,θ d3}, and the decision result y explicit is output. If the detection result is no defect θ a ' bnormal ={θ r1 ,θ r2 ,θ r3}, the hidden defect recognition network is activated;

[0114] (3) Hidden defect detection: Input the image data X without defects image_normal corresponding to the data into the hidden defect recognition network for defect detection. The detection result is one of the three types of hidden defects {θ r1 , θ r2 , θ r3}, and output the decision result y multi ;

[0115] (4) Decision making completion: Aggregate the decision results of the three networks and output the final decision result y.

[0116] y = {y binary || y explicit || y multi} (6)

[0117] The ontology embodiment adopts a trained deep neural network defect recognition module to perform real-time diagnosis on the on-site photovoltaic panel defect situation. By deploying a data acquisition module on-site, image information, environmental information, and current-voltage information of the photovoltaic panel are collected, and the three types of information are preprocessed. The preprocessed environmental information is input into the photovoltaic panel mechanism model to obtain the corresponding theoretical current-voltage data, and the deviation data processing is performed on the theoretical current-voltage data and the measured current-voltage data to obtain the current-voltage experimental data. The current-voltage experimental data and the image data are input into the trained hierarchical recognition model of explicit and hidden defects. First, the current-voltage experimental data is used for overall defect diagnosis. If a defect is diagnosed, the input of the image data is introduced and the explicit defect diagnosis is activated. In the explicit defect diagnosis, if no defect is detected, the corresponding current-voltage experimental data is introduced again, and the hidden defect diagnosis is activated for diagnosis operations. The method provided by this embodiment significantly improves the accuracy and efficiency of fault detection and effectively avoids misjudgment caused by a single information source.

[0118] When explicit and hidden faults are diagnosed, a fault alarm can be issued, and the operation and maintenance personnel can perform maintenance on the photovoltaic panel in a timely manner. At the same time, the diagnosis results and maintenance records are uploaded to the photovoltaic panel diagnosis database to facilitate the management and analysis of fault types and maintenance methods in the future.

[0119] Embodiment 2

[0120] As Figure 3 shown, this embodiment provides a photovoltaic panel defect detection system driven by the fusion of a mechanism model and an image, including:

[0121] A data acquisition module for collecting photovoltaic panel image data and monitoring data at the same time stamp; the monitoring data includes light intensity, environmental temperature, wind speed, current, and voltage information;

[0122] The photovoltaic panel mechanism module is used to calculate current-voltage experimental data based on the monitoring data;

[0123] The defect identification module is used to input the image data and the current-voltage experimental data into the trained explicit and implicit defect hierarchical identification model to obtain the identification results of the explicit defects and implicit defects of the photovoltaic panel;

[0124] Among them, the explicit and implicit defect hierarchical identification model includes a defect classification network, an explicit defect identification network, and an implicit defect identification network connected in series in sequence; the defect classification network is used to identify whether there are defects in the photovoltaic panel based on the current-voltage experimental data; if there are defects, the image data corresponding to the current-voltage experimental data is input into the explicit defect identification network to identify the explicit defects; the current-voltage experimental data corresponding to the image data without explicit defects is input into the implicit defect identification network to identify the implicit defects.

[0125] The explicit and implicit defect hierarchical identification model provided by this specific embodiment realizes the refined classification of the photovoltaic panel defects through the defect classification network, the explicit defect identification network, and the implicit defect identification network connected in series in sequence.

[0126] First, the defect classification network preliminarily judges whether there are defects in the photovoltaic panel based on the current-voltage experimental data, which helps to quickly screen out the photovoltaic panels that need further detection and reduce the ineffective detection; subsequently, for the photovoltaic panels with defects, the model will input the image data into the explicit defect identification network or the implicit defect identification network according to the current-voltage experimental data for further analysis. This targeted detection method improves the accuracy and efficiency of the detection. The design of the series network enables the model to be more targeted when processing the image data. Only when the photovoltaic panel is preliminarily judged to have defects, will the explicit defect identification network or the implicit defect identification network be further called for processing, thus avoiding the indiscriminate processing of all photovoltaic panel image data and saving computing resources and time.

[0127] Embodiment Three

[0128] This embodiment provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it realizes the steps in a photovoltaic panel defect detection method driven by a mechanism model and image fusion as described in Embodiment One above.

[0129] Embodiment Four

[0130] This embodiment provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it realizes the steps in a photovoltaic panel defect detection method driven by a mechanism model and image fusion as described in Embodiment One above.

[0131] The steps or modules involved in the second to fourth embodiments above correspond to those in the first embodiment. For the specific implementation manners, reference may be made to the relevant description part of the first embodiment. The term "computer-readable storage medium" should be understood to include a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0132] The foregoing are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A photovoltaic panel defect detection method driven by mechanism model and image fusion, characterized in that: include: Collect photovoltaic panel image data and monitoring data with the same time stamp; The monitoring data includes light intensity, ambient temperature, wind speed, current and voltage information; Calculate the current-voltage experimental data according to the monitoring data, specifically including: calculate the theoretical current-voltage data according to the photovoltaic panel mechanism, based on the light intensity, ambient temperature and wind speed information; calculate the actual current-voltage data based on the current and voltage information; make a difference between the theoretical current-voltage data and the actual current-voltage data to obtain the deviation current-voltage data; add the theoretical current-voltage data and the deviation current-voltage data to obtain the current-voltage experimental data; Input the image data and current-voltage experimental data into a trained layered recognition model for explicit and implicit defects to obtain recognition results of explicit and implicit defects of photovoltaic panels; Among them, the hierarchical recognition model for explicit and implicit defects includes a defect classification network, an explicit defect recognition network and a implicit defect recognition network connected in series in sequence; the defect classification network is used to identify whether there are defects in the photovoltaic panel based on current-voltage experimental data; if there are defects, the image data corresponding to the current-voltage experimental data is input into the explicit defect recognition network to identify the explicit defects; the current-voltage experimental data corresponding to the image data without explicit defects is input into the implicit defect recognition network to identify the implicit defects; the training process of the implicit defect recognition network includes: obtaining current-voltage experimental data samples corresponding to image data samples identified as explicitly defect-free by the explicit defect recognition network; inputting the current-voltage experimental data samples into the implicit defect recognition network, and training based on the implicit defect labels; calculating the third loss function, and when the third loss function is minimized, the training of the implicit defect recognition network is completed.

2. The photovoltaic panel defect detection method driven by mechanism model and image fusion as claimed in claim 1, characterized in that: The method further includes: performing preprocessing operations of image enhancement, position correction, and uniform image size on the image data; and performing preprocessing of data cleaning and data denoising on the monitoring data.

3. The photovoltaic panel defect detection method driven by mechanism model and image fusion as claimed in claim 1, characterized in that: The explicit defects include cracks, breakages and obstructions; the implicit defects include hidden cracks, current leakage and PID effects.

4. The photovoltaic panel defect detection method driven by mechanism model and image fusion as claimed in claim 1, characterized in that: The defect classification network is a binary classification network constructed based on a residual network; The training process of the defect classification network includes: Inputting the current-voltage experimental data samples into the defect classification network, and training based on the non-defect label and the defect label; the defect label includes an explicit defect label and an implicit defect label; the explicit defect label includes a crack label, a damage label and an occlusion label, and the implicit defect label includes a hidden crack label, a current leakage label and a PID effect label; The first loss function is calculated. When the first loss function is minimized, the defect classification network training is completed.

5. The photovoltaic panel defect detection method driven by mechanism model and image fusion as claimed in claim 1, characterized in that: The training process of the dominant defect recognition network includes: Obtaining image data samples corresponding to current-voltage experimental data samples identified as having defects by the defect classification network; Inputting the image data sample into an explicit defect recognition network, taking the implicit defect label as a whole as an explicit non-defect label, and performing training based on the explicit defect label and the explicit non-defect label; The second loss function is calculated. When the second loss function is minimized, the training of the dominant defect recognition network is completed.

6. A photovoltaic panel defect detection system driven by mechanism model and image fusion, characterized in that: include: A data acquisition module is used to collect photovoltaic panel image data and monitoring data with the same time stamp; the monitoring data includes light intensity, ambient temperature, wind speed, current and voltage information; The photovoltaic panel mechanism module is used to calculate the current-voltage experimental data according to the monitoring data, specifically including: calculating the theoretical current-voltage data according to the photovoltaic panel mechanism, based on the light intensity, ambient temperature and wind speed information; calculating the actual current-voltage data based on the current and voltage information; subtracting the theoretical current-voltage data from the actual current-voltage data to obtain the deviation current-voltage data; adding the theoretical current-voltage data and the deviation current-voltage data to obtain the current-voltage experimental data; A defect recognition module, used for inputting the image data and current-voltage experimental data into a trained layered recognition model for explicit and implicit defects to obtain recognition results of explicit and implicit defects of photovoltaic panels; Among them, the hierarchical recognition model for explicit and implicit defects includes a defect classification network, an explicit defect recognition network and a implicit defect recognition network connected in series in sequence; the defect classification network is used to identify whether there are defects in the photovoltaic panel based on current-voltage experimental data; if there are defects, the image data corresponding to the current-voltage experimental data is input into the explicit defect recognition network to identify the explicit defects; the current-voltage experimental data corresponding to the image data without explicit defects is input into the implicit defect recognition network to identify the implicit defects; the training process of the implicit defect recognition network includes: obtaining current-voltage experimental data samples corresponding to image data samples identified as explicitly defect-free by the explicit defect recognition network; inputting the current-voltage experimental data samples into the implicit defect recognition network, and training based on the implicit defect labels; calculating the third loss function, and when the third loss function is minimized, the training of the implicit defect recognition network is completed.

7. 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 a photovoltaic panel defect detection method driven by a mechanism model and image fusion as described in any one of claims 1 to 5 are implemented.

8. 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 in the photovoltaic panel defect detection method driven by mechanism model and image fusion as described in any one of claims 1-5 are implemented.

Images