Photovoltaic module online defect monitoring system and method

Abstract

The present invention discloses a method for online defect monitoring of photovoltaic modules, which relates to the field of intelligent manufacturing of photovoltaic modules. The method comprises: using a visible light camera device, an infrared thermal imaging device and a spectrum analysis device to collect and acquire three different sets of data respectively, and preprocessing the three sets of data respectively; performing feature enhancement and extraction processing on the three sets of preprocessed data respectively to obtain feature information corresponding to each of the extracted data; performing graphical processing on the feature information respectively; finally, using a trained convolutional long short-term memory network to automatically identify and locate faults on the feature information after graphical processing respectively, and obtain corresponding detection results respectively; if any detection result is abnormal, the detected faulty component is automatically diverted to a scrap channel or a recycling channel. The present invention solves the problems of fault identification limitations caused by a single detection method of traditional photovoltaic modules and insufficient traditional data processing capabilities.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing of photovoltaic modules, and in particular to an online defect monitoring system and method for photovoltaic modules. Background Art

[0002] Current photovoltaic module defect detection technologies mostly rely on a single detection method, including infrared thermal imaging analysis or visible light image analysis. However, using only a single detection method for photovoltaic module defect detection will lead to significant limitations in the detection system when identifying multiple defects. Visible light images can identify surface cracks but cannot detect internal hot spots, and are easily interfered by light reflections. Infrared thermal imaging has low sensitivity to defects without temperature rise (early PID attenuation of photovoltaics), and ambient temperature fluctuations can easily lead to misjudgments. Although spectral analysis can identify material aging, its spatial resolution is insufficient, making it difficult to accurately locate the defect.

[0003] Existing technologies have also attempted to use a pairwise combination solution for defect detection, but there are still unresolved technical bottlenecks:

[0004] 1. Combining visible light images with infrared thermal imaging: This will not accurately identify changes in material chemical properties, such as yellowing of EVA, and will also lead to registration errors due to resolution differences, with a typical error of ≥2mm;

[0005] 2. Combining infrared thermal imaging with spectral analysis: This lacks support for the morphological characteristics of photovoltaic modules, making it difficult to distinguish between hot spots and solder joints, with a misjudgment rate of >25%;

[0006] 3. Combining visible light images with spectral analysis: This approach cannot detect micro-area temperature rises caused by internal cracks, and the spectral sampling rate (typically <10 Hz) is difficult to synchronize with the visible light video stream (≥30 fps).

[0007] In addition, traditional technologies often use a serial computing architecture when processing multimodal data, which can result in processing delays of up to 2 seconds or more, and cannot meet the real-time sorting requirements of the production line (required to be ≤1 second).

[0008] The feature fusion of existing technologies mostly relies on the superposition of fixed weights and fails to dynamically adjust the contribution of each modality according to the environment. Under complex working conditions (such as strong reflections and high ambient temperature fluctuations), the detection accuracy drops by more than 30%. Therefore, due to the isolated detection methods, extensive fusion mechanisms and lack of real-time performance, existing technologies are difficult to achieve high-precision and high-efficiency defect detection in complex industrial scenarios. Summary of the Invention

[0009] In response to the significant limitations of traditional photovoltaic module defect detection technology due to its reliance on a single detection method, the present invention proposes a photovoltaic module online defect monitoring system and method. This system and method utilize advanced convolutional neural networks and long-short-term memory networks to achieve efficient data processing and intelligent analysis, ensuring comprehensive online monitoring of photovoltaic modules and significantly improving the accuracy and real-time performance of fault detection. This system addresses the existing problems of fault identification limitations and insufficient traditional data processing capabilities caused by a single detection method. To achieve the above objectives, the present invention utilizes the following technical solutions:

[0010] A photovoltaic module online defect monitoring method comprises the following steps:

[0011] Step S1: using a visible light camera device, an infrared thermal imaging device, and a spectrum analysis device to respectively acquire visible light image data, infrared thermal image data, and spectrum data of the photovoltaic module, and performing corresponding preprocessing on the visible light image data, infrared thermal image data, and spectrum data;

[0012] Step S2: performing feature enhancement and extraction on the pre-processed visible light image data, infrared thermal image data, and spectral data, respectively, to obtain extracted visible light image feature information, infrared thermal image feature information, and spectral image feature information;

[0013] Step S3, graphically processing the extracted visible light image feature information, infrared thermal image feature information, and spectral image feature information respectively;

[0014] Step S4: using a convolutional long short-term memory network to automatically identify and locate faults on the graphically processed visible light image feature information, infrared thermal image feature information, and spectral image feature information, respectively, to obtain detection results corresponding to the visible light image feature information, infrared thermal image feature information, and spectral image feature information;

[0015] If any test result is abnormal, the detected faulty component will be automatically diverted to the scrap channel or recycling channel;

[0016] The steps of automatically identifying and locating the fault in step S4 include:

[0017] Step S41: Classify the received feature information and use the fault probability formula output by the convolutional long short-term memory network to perform threshold judgment to determine the fault type;

[0018] The failure probability formula is as follows:

[0019] Formula 1

[0020] Where P(y=1|X) represents the probability of a fault occurring given the input feature X; σ is the activation function; h i is the hidden state at the previous moment; x i is the input feature at the current moment; W h and W x are the weight matrices of hidden state and input features respectively; b is the bias term;

[0021] Step S42: Based on the historical detection data of the photovoltaic module and the preset empirical threshold θ, the probability of a fault occurring is compared with the preset empirical threshold θ using the following formula 2 to determine whether a fault exists;

[0022] Formula 2

[0023] Where P(y=1|X) represents the probability of a fault occurring given the input feature X. When the inequality holds, it is determined that the PV module is faulty.

[0024] Step S43: Classify similar faults using a clustering algorithm.

[0025] Preferably, the specific steps of preprocessing the visible light image data in step S1 include:

[0026] Step S111: using Gaussian filtering technology to perform denoising on the visible light image data;

[0027] Step S112: performing image contrast enhancement processing on the visible light image data using a histogram equalization technique;

[0028] Step S113: normalize the visible light image data and standardize the pixel values ​​to a preset uniform range.

[0029] Preferably, the specific steps of preprocessing the infrared thermal image data in step S1 include:

[0030] Step S121: Smoothing the infrared thermal image data and using a median filter technique to remove isolated noise points in the infrared thermal image data;

[0031] Step S122, converting the infrared thermal image data processed in step S121 into a grayscale image;

[0032] Step S123: Convert the grayscale image into a binary image using Formula 3:

[0033] Formula 3

[0034] Among them, T(x,y) represents the binary image value at the position (x,y), I(x,y) is the grayscale value of the infrared thermal image data at the position (x,y), T a (x,y) is the adaptive threshold, calculated by formula 4:

[0035]

[0036] Among them, N(x,y) is the neighborhood pixel set centered at position (x,y), and N is the total number of pixels in the neighborhood.

[0037] Preferably, the step of preprocessing the spectral data in step S1 includes:

[0038] Step S131: performing background correction processing on the spectral data and subtracting the background signal to eliminate environmental interference;

[0039] Step S132: normalizing the spectral data by normalization processing;

[0040] Step S133: performing characteristic peak identification on the standardized spectral data.

[0041] Preferably, the step of feature enhancement in step S2 is: using morphological processing technology to perform expansion and corrosion operations on the features of visible light image data, infrared thermal image data and spectral data, and at the same time using convolutional neural networks to perform deep extraction on the visible light image data, infrared thermal image data and spectral data respectively to obtain the depth feature information corresponding to each of the visible light image data, infrared thermal image data and spectral data.

[0042] Preferably, the step S3 further includes transmitting the extracted feature information back to the remote monitoring station via 5G communication and performing graphical processing using a host computer; the specific operation steps of the transmission include:

[0043] The extracted feature information is compressed, and the compressed feature information is transmitted using the MQTT protocol through the 5G NR module and sent back to the remote monitoring station.

[0044] Preferably, the data transmission process further includes setting security measures, and the step of setting security measures includes:

[0045] Step S31: Before data transmission, the compressed feature information is encrypted using AES;

[0046] Step S32: Implement an authentication mechanism for the device connected to the 5G NR module for data transmission;

[0047] Step S33: During data transmission, use digital certificates for security authentication;

[0048] Step S34: regularly update the AES encryption key in step S31 and the identity authentication mechanism in step S32.

[0049] Preferably, the graphical processing in step S3 is performed by using a data visualization tool to display image feature information in real time.

[0050] Preferably, the data visualization tools include real-time charts, heat maps and trend lines.

[0051] Preferably, the classification method in step S43 includes: performing cluster analysis on the feature information using a K-means algorithm, and classifying similar fault modes into the same category by calculating the Euclidean distance between the feature information.

[0052] Preferably, before step S4, the photovoltaic module online defect monitoring method further includes:

[0053] Step S3': train the convolutional long short-term memory network.

[0054] Preferably, the training steps specifically include:

[0055] The photovoltaic module failure dataset is used for supervised learning. The convolutional long short-term memory network is optimized using the cross entropy loss function according to the preset learning rate and batch size.

[0056] Preferably, the step S4 further includes:

[0057] The conveyor belt is controlled based on the test results to allow normal photovoltaic modules to continue to pass through the production line; the detected faulty modules are automatically diverted to the scrap channel or recycling channel.

[0058] Preferably, the method of controlling the conveyor belt includes:

[0059] Use PLC to receive fault detection results and control the switch of the conveyor belt according to the fault type;

[0060] For normal components, the conveyor belt remains operational;

[0061] For faulty components, the conveyor belt automatically stops and diverts the faulty components to designated scrap channels or recycling channels.

[0062] A photovoltaic module online defect monitoring system is used to run a photovoltaic module online defect monitoring method, the photovoltaic module online defect monitoring system includes a high-resolution visible light camera, an infrared thermal imager, a spectrometer, a PLC controller, a servo motor, a contactor, a frequency converter, a transmission belt, a 5G NR module, a human-machine interface HMI and an alarm module; the high-resolution visible light camera is divided into a first high-resolution visible light camera and a second high-resolution visible light camera; the PLC controller includes a conventional transmission PLC controller and a fault transmission PLC controller; the servo motor includes a first servo motor and a second servo motor; the contactor includes a first contactor and a second contactor; the transmission belt includes a conventional transmission belt and a fault transmission belt; the 5G NR module uses a Qualcomm 5G module.

[0063] Preferably, the first high-resolution visible light camera and the second high-resolution visible light camera are installed in parallel on the side of the conventional conveyor belt for real-time monitoring of the appearance of the photovoltaic module; the infrared thermal imager is installed on the other side of the conventional conveyor belt for monitoring the temperature change of the photovoltaic module; the spectrometer is set in the middle position of the conventional conveyor belt and is located between the conventional conveyor belt and the fault conveyor belt for photoluminescence detection; the conventional transmission PLC controller is responsible for the transmission control of the photovoltaic module under normal conditions to ensure the smooth transmission of the photovoltaic module; the fault transmission PLC controller is dedicated to handling faulty photovoltaic modules to ensure that the faulty photovoltaic modules are effectively guided to the fault conveyor belt; the first servo motor is used to drive the conventional conveyor belt to operate, and the second servo motor is used to drive the fault conveyor belt to operate; the first servo motor is controlled by the conventional transmission PLC controller, and the second servo motor is controlled by the fault transmission PLC controller; the first contactor is used to control the start and stop of the conventional conveyor belt, and the second contactor is used to control the start and stop of the fault conveyor belt; the inverter is installed in the control cabinet to adjust the speed of the servo motor; the conventional conveyor belt is used for the transmission of normal photovoltaic modules, and the fault conveyor belt is used for the transmission of faulty photovoltaic modules; the 5G The NR module is used to realize real-time data transmission and remote monitoring; the human-machine interface HMI is used to provide the operator with an interactive interface of the system and display the system status, fault alarm and parameter settings in real time; the alarm module is responsible for monitoring the system operation status and issuing alarms in time to notify the operator to intervene.

[0064] The present invention has the following advantages and beneficial effects compared to the prior art:

[0065] 1. This invention combines visible light, infrared, and spectral detection methods to achieve comprehensive online defect monitoring of photovoltaic modules, significantly improving the accuracy and real-time performance of fault detection. It also uses convolutional neural networks and long-short-term memory networks for intelligent data processing, enabling rapid analysis and identification of multiple types of defects, overcoming the limitations of traditional single detection methods.

[0066] 2. Traditional single detection or two-by-two detection will lead to detection blind spots such as infrared being unable to identify defects without temperature rise and visible light being easily interfered by reflections due to inherent defects. However, the present invention achieves defect feature complementarity and error correction by deeply integrating visible light, infrared and spectral detection. Visible light imaging is used to accurately locate surface cracks with a spatial resolution of up to 0.1mm. Infrared thermal imaging is used to correct misjudgments caused by visible light surface reflections with a temperature sensitivity of 0.1°C. Spectral analysis is used to identify material aging characteristics with a wavelength accuracy of 0.1nm. Its low spatial resolution defects are compensated by spatial registration through visible light imaging and infrared thermal imaging with an error of <0.5mm. Infrared thermal imaging is used to avoid missing defects without temperature rise, and a secondary review is triggered by a spectral feature peak offset greater than 5nm, realizing a dynamic weight fusion mechanism and breaking through the limitations of traditional simple superposition.

[0067] 3. The present invention solves the problem of multimodal data conflicts and the defects caused by the use of fixed weights or simple splicing and fusion in existing technologies through environmentally adaptive dynamic weight allocation. In addition, the present invention also realizes rapid data transmission and real-time analysis through 5G networks, reducing the need for manual inspections and lowering maintenance costs. It also utilizes the system's multimodal data fusion capabilities to improve the reliability of fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flowchart of the execution of the photovoltaic module online defect monitoring method of the present invention. DETAILED DESCRIPTION

[0069] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0070] In order to make the objectives, technical solutions and advantages of the present invention more clear and explicit, the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0071] Example 1:

[0072] A photovoltaic module online defect monitoring method comprises the following steps:

[0073] Step S1: using a visible light camera device, an infrared thermal imaging device, and a spectrum analysis device to respectively acquire visible light image data, infrared thermal image data, and spectrum data of the photovoltaic module, and performing corresponding preprocessing on the visible light image data, infrared thermal image data, and spectrum data;

[0074] Among them, preprocessing the visible light image data can significantly improve the image quality. The steps are as follows: first, Gaussian filtering technology is used to denoise the visible light image data to reduce random noise in the image and improve image quality; then histogram equalization technology is used to enhance the image contrast of the visible light image data to improve the visibility of defect areas in the image; finally, the visible light image data is normalized and the pixel values ​​are standardized to a preset uniform range to facilitate subsequent feature extraction and analysis.

[0075] By preprocessing the infrared thermal image data, the temperature abnormality area can be extracted; the steps are as follows:

[0076] Step S121: Smoothing the infrared thermal image data and using median filtering technology to remove isolated noise points in the infrared thermal image data to improve the overall image quality;

[0077] Step S122, converting the infrared thermal image data processed in step S121 into a grayscale image;

[0078] Step S123: Convert the grayscale image into a binary image using Formula 3:

[0079] Formula 3

[0080] Among them, T(x,y) represents the binary image value at the position (x,y), I(x,y) is the grayscale value of the infrared thermal image data at the position (x,y), T a (x,y) is the adaptive threshold, calculated by formula 4:

[0081]

[0082] Where N(x,y) is the neighborhood pixel set centered at position (x,y), and N is the total number of pixels in the neighborhood to ensure effective identification of temperature anomaly areas.

[0083] Finally, the spectral data is preprocessed to eliminate environmental interference and ensure the accuracy of subsequent analysis; the steps are as follows:

[0084] Step S131: performing background correction on the spectral data and subtracting the background signal to eliminate environmental interference to ensure the accuracy of the spectral data;

[0085] Step S132: normalizing the spectral data to facilitate comparison and analysis between different samples;

[0086] Step S133: performing characteristic peak identification on the standardized spectral data.

[0087] Step S2: Feature enhancement and extraction processing are performed on the pre-processed visible light image data, infrared thermal image data, and spectral data respectively to obtain visible light image feature information, infrared thermal image feature information, and spectral image feature information after extraction processing; the feature enhancement steps in this step are: using morphological processing technology to perform expansion and corrosion operations on the features of visible light image data, infrared thermal image data, and spectral data to enhance the connectivity of the target area; at the same time, using convolutional neural networks to perform deep extraction on the visible light image data, infrared thermal image data, and spectral data respectively to obtain the depth feature information corresponding to each of the visible light image data, infrared thermal image data, and spectral data, thereby improving the accuracy of fault detection.

[0088] Step S3: The extracted visible light image feature information, infrared thermal image feature information, and spectral image feature information are transmitted back to the remote monitoring station via 5G communication, and then graphically processed by the host computer. The specific operation steps of the transmission in this step include: compressing the extracted feature information, transmitting the compressed feature information using the MQTT protocol through the 5G NR module, and transmitting it back to the remote monitoring station.

[0089] The specific implementation method of transmitting the extracted feature information to the remote monitoring station via 5G communication in this step is as follows: compressing the feature information extracted from the visible light image data, infrared thermal image data, and spectral data to reduce the amount of data transmission; then transmitting the compressed feature information using the MQTT protocol through the 5G NR module to ensure the real-time and reliability of the data, ensuring that the monitoring station can receive the latest monitoring information in a timely manner and finally transmit it back to the remote monitoring station.

[0090] The data transmission process also requires setting security measures, and the steps for setting security measures include:

[0091] Step S31: Before data transmission, the compressed feature information is encrypted using AES;

[0092] Step S32: Implement an authentication mechanism for the device connected to the 5G NR module for data transmission;

[0093] Step S33: During data transmission, use digital certificates for security authentication;

[0094] Step S34: regularly update the AES encryption key in step S31 and the identity authentication mechanism in step S32. The regular period can be 7 days, 15 days, or 30 days, which can be set according to security requirements to ensure the security of data transmission.

[0095] Graphical processing can be achieved by using data visualization tools to display image feature information in real time, allowing operators to quickly understand system status and make appropriate decisions. Data visualization tools include real-time charts, heat maps, and trend lines, which are used to display monitoring data trends and fault detection history, facilitating comprehensive analysis and decision-making.

[0096] Step S4: Using a convolutional long short-term memory network, the visible light image feature information, infrared thermal image feature information, and spectral image feature information after graphical processing are automatically identified and fault located to obtain the detection results corresponding to the visible light image feature information, infrared thermal image feature information, and spectral image feature information; if any detection result is abnormal, the detected faulty component is automatically diverted to a scrap channel or a recycling channel; it should be noted that the convolutional long short-term memory network here is a trained convolutional long short-term memory network; the specific training process refers to the following training steps.

[0097] The steps for automatically identifying and locating faults in this step include:

[0098] Step S41: Classify the received feature information and use the fault probability formula output by the trained convolutional long short-term memory network to perform threshold judgment to determine the fault type; the fault probability formula is as follows:

[0099] Formula 1

[0100] Where P(y=1|X) represents the probability of a fault occurring given the input feature X; σ is the activation function, h i is the hidden state at the previous moment, x i is the input feature at the current moment; W h and W x are the weight matrices of hidden state and input features respectively, and b is the bias term;

[0101] Step S42: A threshold value θ is set based on historical test data and experience of the photovoltaic module. The threshold value θ ensures that the system can accurately identify potential faults. The probability of a fault occurring is compared with the threshold value θ using the following formula 2 to determine whether a fault exists.

[0102] Formula 2

[0103] Where P(y=1|X) represents the probability of a fault occurring given the input feature X. θ is preferably set to 0.92, but this value can be lowered based on production needs. When the inequality holds, the PV module is deemed to be faulty.

[0104] Step S43: Classify similar faults using a clustering algorithm to facilitate subsequent fault analysis and processing. This classification method uses a K-means algorithm to perform cluster analysis on feature information and calculate the Euclidean distance between feature information to group similar fault modes into the same category.

[0105] Step S4 of the present invention also includes a solution that controls the conveyor belt based on the detection results, allowing healthy photovoltaic modules to continue through the production line and automatically diverting detected faulty modules to scrap or recycling channels. The fault detection results can also be shared via a cloud platform, allowing different monitoring stations to exchange data and share experience, further improving the efficiency and accuracy of fault detection.

[0106] The methods for controlling the conveyor belt include:

[0107] Use PLC to receive fault detection results and control the switch of the conveyor belt according to the fault type;

[0108] For normal components, the conveyor belt remains operational;

[0109] For faulty components, the conveyor belt automatically stops and diverts the faulty components to designated scrap channels or recycling channels.

[0110] Specifically, the photovoltaic module online defect monitoring method of the present invention further includes, before step S4, step S3', training a convolutional long short-term memory network; and the training step includes: using a photovoltaic module failure dataset for supervised learning, optimizing the convolutional long short-term memory network using a cross-entropy loss function according to a preset learning rate and batch size; and during the training process, it is also necessary to regularly evaluate the model performance and adjust hyperparameters to improve the generalization ability of the model.

[0111] In order to better implement this plan, the training process of the convolutional long short-term memory network is described as follows:

[0112] The photovoltaic module failure dataset is a training dataset. Data is collected using a trimodal acquisition system. Defective photovoltaic samples were collected and labeled from the photovoltaic module production line for six consecutive months. This means that visible light image data, infrared thermal image data, and spectral data of the photovoltaic modules were collected continuously for six months on a normal photovoltaic production line using visible light cameras, infrared thermal imaging devices, and spectral analysis devices. The dataset consists of defective photovoltaic samples collected over a six-month period. The labeling rules include setting the first-level label as the defect category, including cracks, hot spots, and material aging; setting the second-level label as the defect severity level 0-5, which is quantified based on the international standard IEC 61215; and setting the third-level label as the position label, using pixel-level masking to annotate spatial defects such as visible light cracks and infrared hot spots.

[0113] Specifically, the construction of the PV module failure dataset includes the following steps:

[0114] (1) Collect trimodal sample data including visible light images, infrared thermal images and spectral data, of which normal samples shall be no less than 3,200 groups and defect samples shall cover 12 categories such as cracks, hot spots, material aging, etc., and the total number of samples collected shall be no less than 4,800 groups; (2) According to the international standard IEC 61215, at least three certified engineers shall independently label the defect type, location and severity, and only retain samples with labeling consistency ≥ 95%; (3) Apply random rotation and Gaussian noise to the visible light image, where the rotation angle is controlled within ±15° and the Gaussian noise is ≤0.05dB, and simulate the perturbation of the infrared thermal image based on the Fourier heat conduction equation, and add random baseline drift to the spectral data with an amplitude ≤5%.

[0115] The specific implementation of supervised learning includes the following steps:

[0116] (1) The network input is a three-channel tensor, where channel 1 is a 256×256 visible light image feature map, channel 2 is a 256×256 infrared thermal image pseudo-color coding map, and channel 3 is a 256×256 two-dimensional matrix of spectral feature vector expansion; (2) The AdamW optimizer is used, the initial learning rate is set to 0.001, and the cosine annealing strategy is used to update the learning rate. The formula is:

[0117] Formula 5

[0118] Where, is the learning rate of the tth training cycle, is the maximum value of the learning rate, , is the minimum value of the learning rate, is set to 0.0001, the period T is 50 epochs, and epoch represents a complete dataset training cycle; (3) The loss function uses weighted cross entropy:

[0119]

[0120] in, is the weighted cross entropy loss value, is the weight coefficient of the i-th type defect, is the true label of the i-th type of defect, that is, one-hot encoding, with a value of 0 or 1, is the probability of the i-th type defect predicted by the model, is the number of samples of the i-th type of defects in the training set.

[0121] The steps for regularly evaluating model performance include: (1) calculating the macro-average Macro-F1 on the validation set every 5 epochs, using the formula:

[0122]

[0123] in, is the accuracy, = , that is, the ratio of the number of samples correctly predicted to be class II to the number of samples predicted to be class II;

[0124] is the recall rate, , that is, the ratio of the number of samples correctly predicted to be class II to the total number of samples actually belonging to class II;

[0125] is a true positive, A false positive It is a false negative;

[0126] (2) When the improvement of Macro-F1 is less than 0.5% in three consecutive evaluations, the early stopping mechanism is triggered to terminate the training;

[0127] (3) Identify low-performance defect categories based on the confusion matrix, oversample their samples, and then retrain.

[0128] The steps of adjusting hyperparameters include: (1) using Bayesian optimization algorithm to automatically find the optimal solution in the search space, which is defined as:

[0129]

[0130] (2) The optimization objective function is the weighted sum of the validation set Macro-F1 and the training time:

[0131] Formula 8

[0132] Among them, α is the value of the hyperparameter optimization objective function, Indicates the time taken for a single training session in seconds.

[0133] The steps of model testing include:

[0134] (1) The test set contains time-separated samples and two new types of defects;

[0135] (2) Performance index requirements:

[0136]

[0137] (3) Comparative experiments show that on the same test set, the Macro-F1 of the ConvLSTM model of the present invention needs to be improved by ≥15% compared with the traditional CNN model.

[0138] Furthermore, in visible light image processing, edge detection algorithms, such as Canny edge detection, can be used to enhance edge information in images and help identify defects such as cracks and broken grids. In infrared thermal image processing, thermal threshold analysis can be combined to calibrate abnormal areas within different temperature ranges to facilitate rapid fault location. In spectral data processing, smoothing filters, such as Savitzky-Golay filtering, can be used to reduce fluctuations in spectral data and improve the accuracy of feature extraction. This invention combines multiple detection methods with intelligent algorithms to achieve comprehensive online monitoring of photovoltaic modules, increasing efficiency by approximately 200% compared to manual inspections. The accuracy of detecting microcracks can reach up to 99%, effectively promoting the sustainable development of the photovoltaic industry.

[0139] Example 2:

[0140] A photovoltaic module online defect monitoring system is used to run a photovoltaic module online defect monitoring method, wherein the photovoltaic module online defect monitoring system includes a high-resolution visible light camera, an infrared thermal imager, a spectrometer, a PLC controller, a servo motor, a contactor, a frequency converter, a transmission belt, a 5G NR module, a human-machine interface HMI, and an alarm module; the high-resolution visible light camera is divided into a first high-resolution visible light camera and a second high-resolution visible light camera; the PLC controller includes a conventional transmission PLC controller and a fault transmission PLC controller; the servo motor also includes a first servo motor and a second servo motor; the contactor also includes a first contactor and a second contactor; and the transmission belt includes a conventional transmission belt and a fault transmission belt; the 5G NR module uses a Qualcomm 5G module.

[0141] The specific settings also include: a first high-resolution visible light camera and a second high-resolution visible light camera are installed in parallel on the side of the conventional conveyor belt to monitor the appearance status of the photovoltaic modules in real time; an infrared thermal imager is installed on the other side of the conventional conveyor belt to monitor the temperature changes of the photovoltaic modules; a spectrometer is set in the middle of the conventional conveyor belt and is located between the conventional conveyor belt and the fault conveyor belt for photoluminescence detection; the conventional transmission PLC controller is responsible for the transmission control of the photovoltaic modules under normal conditions to ensure the smooth transmission of the photovoltaic modules; the fault transmission PLC controller is dedicated to handling faulty photovoltaic modules to ensure that the faulty photovoltaic modules are effectively guided to the faulty conveyor belt; the first servo motor is used to drive the conventional conveyor belt to operate, and the second servo motor is used to drive the conventional conveyor belt to operate. The servo motor is used to drive the fault transmission belt to operate; the first servo motor is controlled by the conventional transmission PLC controller, and the second servo motor is controlled by the fault transmission PLC controller; the first contactor is used to control the start and stop of the conventional transmission belt, and the second contactor is used to control the start and stop of the fault transmission belt; the frequency converter is installed in the control cabinet to adjust the speed of the servo motor; the conventional transmission belt is used for the transmission of normal photovoltaic modules, and the fault transmission belt is used for the transmission of faulty photovoltaic modules; the 5GNR module is used to realize real-time transmission and remote monitoring of data; the human-machine interface HMI is used to provide the operator with an interactive interface of the system, and display the system status, fault alarm and parameter settings in real time; the alarm module is responsible for monitoring the system operation status and issuing alarms in time to notify the operator to intervene.

[0142] The system also includes a user feedback module to allow operators to manually verify and provide feedback on the detection results. The feedback information is used to continuously optimize the training process of the convolutional long short-term memory network and improve the system's adaptability and accuracy.

[0143] The photovoltaic module online defect monitoring system disclosed in the present invention is applicable to the field of intelligent manufacturing of photovoltaic modules. By using a high-resolution visible light camera to obtain the structural image of the photovoltaic module in real time, it can preliminarily diagnose physical defects such as cracks and broken grids of the photovoltaic module. Then, an infrared thermal imager is used to monitor the temperature state of the photovoltaic module in real time to further identify faults such as hidden cracks, poor soldering of solder ribbons, diode failure and hot spots. Subsequently, a spectrometer is used to perform photoluminescence detection to evaluate material quality and electrical performance. Finally, the real-time collected image data, temperature data and spectral data are transmitted to the monitoring center through the 5G network to achieve centralized monitoring and analysis. The system makes full use of the trained convolutional neural network to automatically detect and locate faults on the received feature information, and combines the clustering algorithm to classify faults and identify similar fault modes. In addition, a long short-term memory network is used to process time series data, analyze temperature change trends, predict potential fault risks, and automatically control the production line conveyor belt according to the detection results, so that normal components continue to be transmitted, while faulty components are diverted to the fault channel or scrap channel for scrapping or recycling.

[0144] The photovoltaic module online defect monitoring system and method provided by the present invention aims to solve the problems of fault identification limitations and insufficient traditional data processing capabilities caused by a single detection method in the existing technology; by combining visible light, infrared and spectral detection methods, it can achieve comprehensive monitoring of photovoltaic modules and ensure accuracy and reliability in complex environments; at the same time, it uses convolutional neural networks and long short-term memory networks for efficient data processing and intelligent analysis, significantly improving the real-time and accuracy of fault detection, thereby effectively promoting the sustainable development of photovoltaic power generation.

Claims

1. A photovoltaic module online defect monitoring method, characterized in that: The following steps are involved: Step S1: using a visible light camera device, an infrared thermal imaging device, and a spectrum analysis device to respectively acquire visible light image data, infrared thermal image data, and spectrum data of the photovoltaic module, and performing corresponding preprocessing on the visible light image data, infrared thermal image data, and spectrum data; Step S2: performing feature enhancement and extraction on the pre-processed visible light image data, infrared thermal image data, and spectral data, respectively, to obtain extracted visible light image feature information, infrared thermal image feature information, and spectral image feature information; Step S3, graphically processing the extracted visible light image feature information, infrared thermal image feature information, and spectral image feature information respectively; Step S4: using a convolutional long short-term memory network to automatically identify and locate faults on the graphically processed visible light image feature information, infrared thermal image feature information, and spectral image feature information, respectively, to obtain detection results corresponding to the visible light image feature information, infrared thermal image feature information, and spectral image feature information; If any test result is abnormal, the detected faulty component will be automatically diverted to the scrap channel or recycling channel; The steps of automatically identifying and locating the fault in step S4 include: Step S41: classify the received feature information and use the fault probability formula output by the convolutional long short-term memory network to perform threshold judgment to determine the fault type; The failure probability formula is as follows: Formula 1 Where P(y=1|X) represents the probability of a fault occurring given the input feature X; σ is the activation function; h i is the hidden state at the previous moment; x i is the input feature at the current moment; W h and W x are the weight matrices of hidden state and input features respectively; b is the bias term; Step S42: Based on the historical detection data of the photovoltaic module and the preset empirical threshold θ, the probability of a fault occurring is compared with the preset empirical threshold θ using the following formula 2 to determine whether a fault exists; Formula 2 Where P(y=1|X) represents the probability of a fault occurring given the input feature X. When the inequality holds, it is determined that the PV module is faulty. Step S43: Classify similar faults using a clustering algorithm.

2. A photovoltaic module online defect monitoring method according to claim 1, characterized in that: The specific steps of preprocessing the visible light image data in step S1 include: Step S111: using Gaussian filtering technology to perform denoising on the visible light image data; Step S112: performing image contrast enhancement processing on the visible light image data using a histogram equalization technique; Step S113: normalize the visible light image data and standardize the pixel values ​​to a preset uniform range.

3. A photovoltaic module online defect monitoring method according to claim 1, characterized in that: The specific steps of pre-processing the infrared thermal image data in step S1 include: Step S121: Smoothing the infrared thermal image data and using a median filter technique to remove isolated noise points in the infrared thermal image data; Step S122, converting the infrared thermal image data processed in step S121 into a grayscale image; Step S123: Convert the grayscale image into a binary image using Formula 3: Formula 3 Among them, T(x,y) represents the binary image value at the position (x,y), I(x,y) is the grayscale value of the infrared thermal image data at the position (x,y), T a (x,y) is the adaptive threshold, calculated by formula 4: Among them, N(x,y) is the neighborhood pixel set centered at position (x,y), and N is the total number of pixels in the neighborhood.

4. A photovoltaic module online defect monitoring method according to claim 1, characterized in that: The step of preprocessing the spectral data in step S1 includes: Step S131: performing background correction processing on the spectral data and subtracting the background signal to eliminate environmental interference; Step S132: normalizing the spectral data by normalization processing; Step S133: performing characteristic peak identification on the standardized spectral data.

5. The photovoltaic module online defect monitoring method according to claim 1, characterized in that: The step of feature enhancement in step S2 is: using morphological processing technology to perform expansion and corrosion operations on the features of visible light image data, infrared thermal image data and spectral data, and using convolutional neural networks to perform depth extraction on the visible light image data, infrared thermal image data and spectral data respectively to obtain the depth feature information corresponding to each of the visible light image data, infrared thermal image data and spectral data.

6. A photovoltaic module online defect monitoring method according to claim 1, characterized in that: The step S3 also includes transmitting the extracted feature information back to the remote monitoring station via 5G communication and performing graphical processing using a host computer. The specific operation steps of the transmission include: The extracted feature information is compressed, and the compressed feature information is transmitted using the MQTT protocol through the 5G NR module and sent back to the remote monitoring station.

7. A photovoltaic module online defect monitoring method according to claim 6, characterized in that: The data transmission process further includes setting security measures, and the step of setting security measures includes: Step S31: Before data transmission, the compressed feature information is encrypted using AES; Step S32: Implement an authentication mechanism for the device connected to the 5G NR module for data transmission; Step S33: During data transmission, use digital certificates for security authentication; Step S34: regularly update the AES encryption key in step S31 and the identity authentication mechanism in step S32.

8. The photovoltaic module online defect monitoring method according to claim 1, characterized in that: The graphical processing method in step S3 is: using a data visualization tool to display image feature information in real time.

9. A photovoltaic module online defect monitoring method according to claim 8, characterized in that: The data visualization tools include real-time charts, heat maps, and trend lines.

10. The photovoltaic module online defect monitoring method according to claim 1, characterized in that: The classification method in step S43 includes: using the K-means algorithm to perform cluster analysis on the feature information, and classifying similar fault modes into the same category by calculating the Euclidean distance between the feature information.

11. The photovoltaic module online defect monitoring method according to claim 1, characterized in that: The photovoltaic module online defect monitoring method further includes, before step S4: Step S3': train the convolutional long short-term memory network.

12. A photovoltaic module online defect monitoring method according to claim 11, characterized in that: The training steps specifically include: The photovoltaic module failure dataset is used for supervised learning. The convolutional long short-term memory network is optimized using the cross entropy loss function according to the preset learning rate and batch size.

13. The photovoltaic module online defect monitoring method according to claim 1, characterized in that: The step S4 further includes: The conveyor belt is controlled based on the test results to allow normal photovoltaic modules to continue to pass through the production line; the detected faulty modules are automatically diverted to the scrap channel or recycling channel.

14. A photovoltaic module online defect monitoring method according to claim 13, characterized in that: The method of controlling the conveyor belt includes: Use PLC to receive fault detection results and control the switch of the conveyor belt according to the fault type; For normal components, the conveyor belt remains operational; For faulty components, the conveyor belt automatically stops and diverts the faulty components to designated scrap channels or recycling channels.

15. A photovoltaic module online defect monitoring system, used for running the photovoltaic module online defect monitoring method according to any one of claims 1 to 14, characterized in that: The photovoltaic module online defect monitoring system includes a high-resolution visible light camera, an infrared thermal imager, a spectrometer, a PLC controller, a servo motor, a contactor, a frequency converter, a transmission belt, a 5G NR module, a human-machine interface HMI and an alarm module; the high-resolution visible light camera is divided into a first high-resolution visible light camera and a second high-resolution visible light camera; the PLC controller includes a conventional transmission PLC controller and a fault transmission PLC controller; the servo motor includes a first servo motor and a second servo motor; the contactor includes a first contactor and a second contactor; the transmission belt includes a conventional transmission belt and a fault transmission belt; the 5G NR module uses a Qualcomm 5G module.

16. The photovoltaic module online defect monitoring system according to claim 15, characterized in that: The first high-resolution visible light camera and the second high-resolution visible light camera are installed side by side on the side of the conventional conveyor belt for real-time monitoring of the appearance of the photovoltaic modules; the infrared thermal imager is installed on the other side of the conventional conveyor belt for monitoring the temperature changes of the photovoltaic modules; the spectrometer is arranged in the middle of the conventional conveyor belt and between the conventional conveyor belt and the fault conveyor belt for photoluminescence detection; the conventional transmission PLC controller is responsible for the transmission control of the photovoltaic modules under normal conditions to ensure the smooth transmission of the photovoltaic modules; the fault transmission PLC controller is dedicated to handling faulty photovoltaic modules to ensure that the faulty photovoltaic modules are effectively guided to the faulty conveyor belt; the first servo motor is used to drive the conventional conveyor belt to operate, and the second servo motor is used to drive the faulty conveyor belt to operate; the first servo motor is controlled by the conventional transmission PLC controller, and the second servo motor is controlled by the fault transmission PLC controller; The first contactor is used to control the start and stop of the conventional transmission belt, and the second contactor is used to control the start and stop of the fault transmission belt; the frequency converter is installed in the control cabinet and is used to adjust the speed of the servo motor; the conventional transmission belt is used for the transmission of normal photovoltaic modules, and the fault transmission belt is used for the transmission of faulty photovoltaic modules; the 5G NR module is used to realize real-time transmission and remote monitoring of data; the human-machine interface HMI is used to provide the operator with an interactive interface of the system and display the system status, fault alarm and parameter settings in real time; the alarm module is responsible for monitoring the system operation status and issuing an alarm in time to notify the operator to intervene.

Images