A photovoltaic panel surface defect detection method and system based on physical property analysis

Abstract

The present invention belongs to the technical field of photovoltaic panel surface defect detection and discloses a photovoltaic panel surface defect detection method and system based on physical property analysis. The method first utilizes an infrared thermal imager, a laser speckle interferometer, an ultrasonic flaw detector, a visible light multi-band imager, and eddy current detection equipment to obtain a surface temperature distribution image, surface deformation data, ultrasonic echo data, a spectral image and spectral feature data, and eddy current signal feature data of the photovoltaic panel, and then determines various abnormal areas such as temperature, deformation, ultrasonic echo, spectral feature, and eddy current signal. The defect location is then determined by combining various abnormal areas, and the type and size of the photovoltaic panel surface defect are determined by combining multiple data corresponding to the defect location. This method achieves precise positioning and identification of photovoltaic panel surface defects through multiple physical property detection methods, effectively improving detection accuracy and reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic panel surface defect detection, and in particular to a photovoltaic panel surface defect detection method and system based on physical property analysis. Background Art

[0002] With the growing global demand for clean energy, photovoltaic power generation, as a green and environmentally friendly energy source, has gained widespread application and rapid development. As the core component of photovoltaic power generation systems, the performance and reliability of photovoltaic panels directly impact power generation efficiency and system lifespan. However, during production, transportation, and use, photovoltaic panels are prone to various surface defects, making effective detection of surface defects in photovoltaic panels crucial.

[0003] Currently, photovoltaic panel surface defect detection primarily relies on a single inspection technology. For example, some companies use visual inspection, relying on direct observation of the panel surface to identify obvious defects such as cracks and stains. Others employ simple optical imaging, capturing images of the panel surface with a standard camera and using image processing algorithms to identify defects. Furthermore, some companies rely solely on physical inspection methods.

[0004] However, these existing single-item inspection technologies present numerous challenges. Visual inspection methods are significantly affected by human factors, resulting in low efficiency and a high risk of missed inspections. Simple optical imaging inspections struggle to identify minor and internal defects. Single physical inspection methods can only detect specific types of defects and fail to fully cover the full range of surface defects that may occur in photovoltaic panels. Consequently, their accuracy and completeness are insufficient to meet the photovoltaic industry's demand for high-quality inspections. Summary of the Invention

[0005] In view of this, an embodiment of the present invention provides a photovoltaic panel surface defect detection method and system based on physical property analysis to solve the above technical problems.

[0006] To achieve the above objectives, in a first aspect, a method for detecting photovoltaic panel surface defects based on physical property analysis is provided, which comprises the following steps:

[0007] Scanning the entire photovoltaic panel using an infrared thermal imager to obtain a surface temperature distribution image of the photovoltaic panel, and determining temperature abnormality areas based on the surface temperature distribution image;

[0008] Scanning the surface of the photovoltaic panel using a laser speckle interferometer to obtain surface deformation data of the photovoltaic panel, and determining an abnormal deformation area based on the surface deformation data;

[0009] Using an ultrasonic flaw detector to ultrasonically scan the surface of the photovoltaic panel to obtain ultrasonic echo data of the surface of the photovoltaic panel, and determining an ultrasonic echo abnormal area based on the ultrasonic echo data;

[0010] A visible light multi-band imager is used to obtain a spectral image of the photovoltaic panel surface, spectral feature data is extracted from the spectral image based on a spectral unmixing algorithm, and spectral feature abnormality areas are determined based on the spectral feature data;

[0011] Using eddy current detection equipment to detect the surface area of ​​the photovoltaic panel to obtain eddy current signal characteristic data, and determining the eddy current signal abnormal area based on the eddy current signal characteristic data;

[0012] Determining a defect location based on the abnormal temperature area, the abnormal deformation area, the abnormal ultrasonic echo area, the abnormal spectrum feature area, and the abnormal eddy current signal area;

[0013] The photovoltaic panel surface defect recognition result including the type and size of the surface defect is determined based on the surface temperature distribution image, surface deformation data, ultrasonic echo data, spectral feature data and eddy current signal feature data corresponding to the defect position.

[0014] In a second aspect, a photovoltaic panel surface defect detection system based on physical property analysis is provided, wherein the system is configured to execute the method described in the first aspect, and the system comprises:

[0015] An infrared thermal imaging detection module is used to scan the entire photovoltaic panel using an infrared thermal imager to obtain a surface temperature distribution image of the photovoltaic panel and determine temperature abnormality areas based on the surface temperature distribution image;

[0016] A laser speckle deformation detection module is used to scan the surface of the photovoltaic panel using a laser speckle interferometer to obtain surface deformation data of the photovoltaic panel and determine the abnormal deformation area based on the surface deformation data;

[0017] An ultrasonic flaw detection module is used to perform ultrasonic scanning on the surface of the photovoltaic panel using an ultrasonic flaw detector, obtain ultrasonic echo data on the surface of the photovoltaic panel, and determine an ultrasonic echo abnormal area based on the ultrasonic echo data;

[0018] A spectral feature analysis module is used to obtain a spectral image of the photovoltaic panel surface using a visible light multi-band imager, extract spectral feature data from the spectral image based on a spectral unmixing algorithm, and determine spectral feature abnormality areas based on the spectral feature data;

[0019] An eddy current signal detection module is used to detect the surface area of ​​the photovoltaic panel using eddy current detection equipment to obtain eddy current signal characteristic data, and determine the eddy current signal abnormal area based on the eddy current signal characteristic data;

[0020] a defect location determination module, configured to determine the defect location based on the abnormal temperature area, the abnormal deformation area, the abnormal ultrasonic echo area, the abnormal spectrum feature area, and the abnormal eddy current signal area;

[0021] The defect recognition and analysis module is used to determine the photovoltaic panel surface defect recognition results, including the type and size of the surface defects, based on the surface temperature distribution image, surface deformation data, ultrasonic echo data, spectral feature data and eddy current signal feature data corresponding to the defect location.

[0022] The above technical solution has the following beneficial technical effects:

[0023] The present invention can comprehensively obtain multi-source data information such as photovoltaic panel surface temperature, deformation, ultrasonic echo, spectrum and eddy current signal through the collaborative operation of infrared thermal imager, laser speckle interferometer, ultrasonic flaw detector, visible light multi-band imager and eddy current detection equipment. Compared with the disadvantage that a single detection technology can only capture a specific type of defect, the organic combination of this multi-physical property detection method can fully cover all kinds of defects that may appear on the surface of photovoltaic panels, avoiding missed detection and false detection caused by a single detection dimension. On the basis of accurately locating the defect position, based on the comprehensive analysis of multiple data, it can accurately determine the type and size of surface defects, providing a comprehensive and reliable basis for photovoltaic panel quality assessment. In addition, the multi-device collaborative detection mode realizes the automation and efficiency of the detection process, reduces manual intervention, reduces the impact of human factors on the detection results, and improves detection efficiency and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings are provided for a better understanding of the present invention and are not intended to limit the present invention.

[0025] Figure 1 This is an overall flow chart of a photovoltaic panel surface defect detection method based on physical property analysis according to an embodiment of the present invention;

[0026] Figure 2 is a specific flow chart of step S10 in an embodiment of the present invention;

[0027] Figure 3 is a specific flow chart of step S20 in an embodiment of the present invention;

[0028] Figure 4 is a specific flow chart of step S30 of an embodiment of the present invention;

[0029] Figure 5 is a specific flow chart of step S40 in an embodiment of the present invention;

[0030] Figure 6 is a specific flow chart of step S50 in an embodiment of the present invention;

[0031] Figure 7 is a specific flow chart of step S60 in an embodiment of the present invention;

[0032] Figure 8 is a specific flow chart of step S70 in an embodiment of the present invention;

[0033] Figure 9 This is a block diagram of a photovoltaic panel surface defect detection system based on physical property analysis according to an embodiment of the present invention;

[0034] Figure 10 Schematic diagram of the structure of a computer system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following description of exemplary embodiments of the present invention is made in conjunction with the accompanying drawings, in which various details of the embodiments of the present invention are included to facilitate understanding. These details should be considered as merely exemplary. Therefore, it should be appreciated by those skilled in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0036] like Figure 1 As shown, this embodiment provides a photovoltaic panel surface defect detection method based on physical property analysis, which includes the following steps:

[0037] S10: Scanning the entire photovoltaic panel using an infrared thermal imager to obtain a surface temperature distribution image of the photovoltaic panel, and determining an abnormal temperature area based on the surface temperature distribution image;

[0038] First, the photovoltaic panels to be tested were placed in a constant temperature and humidity environment at 25°C and 40% for 30 minutes to allow the panel surface temperature to reach a stable state and prevent environmental factors from interfering with the temperature measurement results. An infrared thermal imager was used to measure the surface temperature distribution of the panels. Before testing, the camera was calibrated by measuring a blackbody radiation source of known temperature and adjusting the device parameters to ensure accurate temperature measurements.

[0039] During testing, the infrared thermal imager was mounted on a tripod and adjusted using a level to maintain a perpendicular position to the photovoltaic panel surface, with a distance of 0.5 meters between them. The infrared thermal imager was turned on, with a resolution of 640 × 480 pixels, a temperature measurement range of -20°C to 150°C, an accuracy of ±1°C, and a scanning speed of 10 frames per second. The entire photovoltaic panel was scanned to obtain an image of the surface temperature distribution.

[0040] After acquiring the image, the computer's onboard image processing software uses a region growing algorithm based on threshold segmentation to process the image. This algorithm uses the average normal operating temperature as a benchmark, sets a temperature deviation threshold of ±5°C, and marks areas where the temperature exceeds this range as seed points. Starting from this seed point, adjacent pixels are merged according to a similarity criterion (temperature differences within a certain range), ultimately forming a temperature anomaly region.

[0041] S20: Scanning the surface of the photovoltaic panel using a laser speckle interferometer to obtain surface deformation data of the photovoltaic panel, and determining an abnormal deformation area based on the surface deformation data;

[0042] This step uses a laser speckle interferometer, which features high resolution and sensitivity, capable of detecting minute surface deformations on photovoltaic panels. Prior to testing, the laser speckle interferometer undergoes optical path calibration. By adjusting the positions of the laser emitter and receiver, the laser beam is accurately illuminated on the photovoltaic panel surface and effectively receives reflected light. The photovoltaic panel is mounted on a vibration table and a sinusoidal excitation with a frequency of 50 Hz and an amplitude of 0.01 mm is applied to induce minute surface vibrations, facilitating the detection of surface deformation. The laser speckle interferometer's laser wavelength is set to 532 nm, power is 20 mW, and the scanning area is set to a 10 cm × 10 cm grid on the photovoltaic panel surface. During the scanning process, phase shifting technology is used to acquire interference fringe images. A four-step phase shifting algorithm is used, which sequentially changes the laser phase to acquire four interference images with different phases. Surface deformation data is then calculated using the four-step phase shifting algorithm. After the calculated deformation data is obtained, the deformation value is compared with a preset threshold of 0.1 μm. Areas exceeding this threshold are identified as areas of abnormal deformation.

[0043] S30: performing ultrasonic scanning on the surface of the photovoltaic panel using an ultrasonic flaw detector to obtain ultrasonic echo data of the surface of the photovoltaic panel, and determining an ultrasonic echo abnormal area based on the ultrasonic echo data;

[0044] This step uses an ultrasonic flaw detector equipped with a 5MHz dual-element probe. This probe effectively reduces near-field interference and improves detection accuracy and resolution. Before testing, the ultrasonic flaw detector is calibrated for sound velocity. A standard test block made of the same material as the photovoltaic panel is selected. The propagation time of the ultrasonic wave in the test block is measured to calculate and set the correct sound velocity parameters. During testing, a water-based coupling agent is evenly applied to the photovoltaic panel surface to ensure smooth transmission of the ultrasonic wave into the panel. The ultrasonic flaw detector probe is scanned along the panel surface at a speed of 0.5m / s, with a scan interval of 2mm to ensure full coverage of the test area. The ultrasonic transmit voltage is set to 100V and the gain to 40dB to obtain clear ultrasonic echo signals. The collected echo signals are processed using a wavelet transform algorithm using computer signal processing software to remove noise interference. The amplitude, time, and frequency characteristics of the echo signals are then extracted. Areas with an echo signal amplitude attenuation exceeding 50% or abnormal reflection peaks are marked as abnormal ultrasonic echo areas.

[0045] S40: Acquire a spectral image of the photovoltaic panel surface using a visible light multi-band imager, extract spectral feature data from the spectral image based on a spectral unmixing algorithm, and determine a spectral feature abnormality region based on the spectral feature data;

[0046] The visible light multi-band imager covers a spectral range of 400-1000nm with a spectral resolution of 5nm, capable of capturing rich spectral information. Testing was conducted under a standard D65 illuminant, which simulates the spectral distribution of average daylight and provides stable, uniform lighting conditions. Prior to testing, the imager underwent dark current and radiometric calibration to eliminate inherent device noise and improve the accuracy of the spectral data.

[0047] The imager was mounted on a movable detection bracket, and its position and angle were adjusted to capture the entire surface of the photovoltaic panel. An exposure time of 100ms was set, and a surface scan of the photovoltaic panel was performed to obtain a spectral image. Spectral unmixing was performed using a linear mixture model. This model, based on the principle of linear combination of spectral data, decomposes each pixel in the spectral image into a combination of spectral signatures of different substances.

[0048] The unmixed spectral feature vectors are compared with a pre-set defect spectrum library, which stores standard spectral features for different types of defects. The similarity between the spectral feature vectors is calculated. When the similarity exceeds 85%, the corresponding area is identified as an abnormal spectral feature area.

[0049] S50: Using eddy current detection equipment to detect the surface area of ​​the photovoltaic panel, obtain eddy current signal characteristic data, and determine the eddy current signal abnormal area based on the eddy current signal characteristic data;

[0050] Specifically, the eddy current detector is equipped with an absolute probe with an outer diameter of 8mm. The absolute probe can detect the overall characteristic changes of the object being tested and is suitable for the preliminary screening of surface defects of photovoltaic panels. The probe of the eddy current detector is kept perpendicular to the surface of the photovoltaic panel at a distance of 1mm to ensure that the eddy current can effectively penetrate the surface layer of the photovoltaic panel. The excitation frequency is set to 10kHz, the detection speed is 10mm / s, and the surface area of ​​the photovoltaic panel is grid-scanned with a scanning interval of 1mm. During the scanning process, the eddy current signal is collected in real time, and the impedance change characteristics of the eddy current signal are extracted. The collected impedance change data is compared with the impedance data of the normal area, and the impedance change rate is calculated. When the impedance change rate exceeds 30% of the normal area, the corresponding area is marked as an abnormal eddy current signal area.

[0051] S60: determining a defect location according to the abnormal temperature area, the abnormal deformation area, the abnormal ultrasonic echo area, the abnormal spectrum feature area, and the abnormal eddy current signal area;

[0052] The anomaly areas identified by the five detection methods were projected onto a unified coordinate system using Geographic Information System (GIS) technology. A weighted voting method was used to determine defect locations. To ensure that the voting results more accurately reflect actual detection conditions, different weights were assigned based on the sensitivity and accuracy of each detection method for each defect type: temperature anomaly areas were weighted 0.2, as temperature testing is more sensitive to defects such as hot spots; deformation anomaly areas were weighted 0.25, as deformation testing is effective in detecting surface deformation defects; ultrasonic echo anomaly areas were weighted 0.25, as ultrasonic testing is more effective for detecting internal defects; spectral feature anomaly areas were weighted 0.2, as spectral testing can identify defects such as material changes; and eddy current signal anomaly areas were weighted 0.1, as eddy current testing is primarily used for preliminary screening of changes in surface conductivity. If an area is flagged as an anomaly by at least three detection methods and its combined weight score exceeds 0.6, it is considered a defect location. The specific calculation method is to convert the anomaly flag for each detection method into a score (1 for an anomaly flag, 0 for no anomaly flag), multiply the scores by the corresponding weights, and sum them to obtain the combined weighted score.

[0053] S70: Determine photovoltaic panel surface defect recognition results including the type and size of the surface defect based on the surface temperature distribution image, surface deformation data, ultrasonic echo data, spectral feature data, and eddy current signal feature data corresponding to the defect position.

[0054] For each identified defect location, five types of physical property data are collected: surface temperature distribution images, surface deformation data, ultrasonic echo data, spectral signature data, and eddy current signal signature data. A defect type classification model based on a support vector machine (SVM) is established. Before building the model, the five types of data are normalized and uniformly mapped to the [0, 1] interval to eliminate the influence of data dimension.

[0055] The normalized feature vector is used as input to the SVM model, and the output is the defect type, including cracks, delamination, bubbles, contamination, etc. The SVM model is trained with a large amount of known defect sample data, and the model parameters are adjusted to enable accurate classification capabilities.

[0056] A morphological processing algorithm is used to calculate the area and perimeter of the defective area. The specific process is as follows: the image of the defective area is first binarized to convert it into a black and white image. Morphological operations such as erosion and dilation are then applied to remove noise and fill holes. Finally, the area is calculated by counting the number of white pixels in the image, and the perimeter is calculated using a boundary tracing algorithm. Using a preset conversion factor, the pixel size in the image is converted to actual physical dimensions to determine the actual size of the defect. The final output is a photovoltaic panel surface defect identification report that includes the defect type and size.

[0057] This solution integrates infrared thermal imagers, laser speckle pattern interferometers, ultrasonic flaw detectors, visible light multi-band imagers, and eddy current testing equipment to capture data from various physical characteristics, including temperature, deformation, ultrasonic echoes, spectral characteristics, and eddy current signals. This comprehensive approach covers all possible defects on the photovoltaic panel surface, including cracks, delamination, bubbles, and contamination. This avoids the limitations of single detection technologies that can only target specific defects, reducing the risk of missed detections. For detection accuracy, each detection device undergoes parameter optimization and calibration based on its specific characteristics. Combined with corresponding data processing algorithms, such as the threshold segmentation and region growing algorithm for infrared thermal imaging and the four-step phase shift algorithm for laser speckle pattern interferometry, these algorithms accurately capture defect characteristics. Furthermore, a weighted voting method is used to determine defect location, a support vector machine is used to classify defect types, and a morphological processing algorithm is used to calculate defect size, achieving high-precision identification of defect location, type, and size. In terms of detection efficiency and reliability, the automated inspection process reduces manual intervention and human error. Furthermore, the collaborative operation of multiple devices and the integration of multiple algorithms enhance detection stability and reliability.

[0058] like Figure 2 As shown, in some embodiments, step S10 specifically includes:

[0059] S11: setting the scanning range of the infrared thermal imager to cover the entire surface of the photovoltaic panel, performing a full-area scan of the photovoltaic panel, and obtaining multiple frames of surface temperature distribution images;

[0060] The photovoltaic panels to be inspected were placed in a darkroom or outdoors at a temperature of 25°C ± 0.5°C and a humidity of 40% ± 5% for 30 minutes to ensure uniform surface temperature. An infrared thermal imager (equipped with a macro lens and a 45° × 34° field of view) was used and mounted 1.5 meters above the photovoltaic panels on a motorized lift platform. The pan / tilt stage was adjusted so that the optical axis was perpendicular to the panel surface. Before inspection, a two-point blackbody calibration (25°C / 50°C) was performed. Parameters were set to a resolution of 640 × 480 pixels, a frame rate of 20 Hz, an emissivity of 0.95, a measurement range of 20°C–80°C, and a thermal sensitivity of less than 0.03°C. A spiral scanning mode was used (starting at the panel center, a pitch of 5 cm), a scanning speed of 0.1 m / s, and 60 frames were acquired continuously, ensuring an overlap of ≥30% per frame. Immediately after acquisition, images were time-synchronized and geotagged, and the raw data was stored on an industrial computer.

[0061] S12: performing noise reduction processing on the surface temperature distribution image using an adaptive threshold filtering algorithm to obtain a preprocessed surface temperature distribution image;

[0062] The acquired image sequence was imported into the MATLAB R2023a environment. Temporal filtering (5-frame median filtering) was first performed to remove random noise. An adaptive Wiener filter algorithm was used, implemented as follows: the image was segmented into overlapping 8×8 pixel sub-blocks; the local mean and variance of each sub-block were calculated; and the mean and variance of the image were calculated using the formula H (u, v) = 1 / [1 + σ² / σ n ²(u,v)] calculates the frequency domain transfer function, where σ² is the noise variance (estimated by the dark area of ​​the image), σ n ²(u,v) is the local variance. Frequency domain filtering is performed on each sub-block to reconstruct the image. The final output is a preprocessed image with an SNR ≥ 38dB.

[0063] S13: Calculating the temperature deviation between the temperature value of each pixel in the pre-processed surface temperature distribution image and a preset normal temperature range, and marking the pixel whose temperature deviation exceeds a dynamic threshold as an abnormal point;

[0064] First, a temperature reference model is constructed. The preprocessed image is Gaussian smoothed (σ=1.5) to calculate the global temperature mean μ0=30.2℃ and the standard deviation σ0=2.1℃. The normal temperature range is set to [μ0-2σ0, μ0+2σ0], that is, [26.0℃, 34.4℃]. The dynamic threshold calculation adopts an iterative optimization strategy as follows: (1) Initial threshold τ0=μ0+1.5σ0=33.4℃; (2) Pixels with temperature greater than T0 are marked as candidate outliers; (3) Calculate the temperature mean μ1 and standard deviation σ1 of the candidate points; (4) Update the threshold τ1=μ1+0.8σ1; (5) Repeat steps 2 to 4 until τ converges (the change is less than 0.1℃). Finally, the threshold τ=35.7℃ is determined, and pixels with temperature greater than τ are marked as outliers, generating a binary label map (outlier value is 255, normal point value is 0).

[0065] S14: Connectivity analysis is performed on the pixel points marked as outliers, and adjacent outliers are classified into the same outlier sub-region using a connected region algorithm to generate a preliminary region map of the temperature anomaly region; within each outlier sub-region of the preliminary region map, a region expansion algorithm is used to further connect adjacent but not yet connected outliers based on a preset spatial distance threshold to form a temperature anomaly region with a continuous boundary.

[0066] Connected regions are generated using a parallel queue filling algorithm. The specific steps are as follows: the labeled graph is evenly divided into 16×16 processing blocks, and each processing block is independently processed in parallel; a 4-neighborhood connectivity analysis is performed on the outliers in each processing block to identify local connected regions within the block; cross-block regions are merged using a boundary matching algorithm, and the region ID mapping table is used to record the region merging relationship to ensure that the same connected region in adjacent processing blocks is assigned the same ID.

[0067] The implementation process of the region expansion algorithm includes the following steps: calculating the geometric center and circumscribed rectangle of each connected region for subsequent spatial relationship analysis; screening adjacent region pairs with a distance less than 15 pixels and a temperature gradient greater than 0.5°C / pixel, and constructing a Delaunay triangulation for these qualified region pairs to identify connection paths; searching the triangular mesh for bridge points with a temperature greater than 34.5°C, which serve as key nodes for region connection; and connecting adjacent regions using the region growing method, using the nearest boundary point of the region as the seed point. The growth condition is set as a temperature greater than 34°C and meeting the 8-neighborhood connectivity requirement.

[0068] The resulting anomaly area map contains complete boundary information and calculates 12 geometric characteristic parameters for each area, including area, perimeter, eccentricity, circularity, rectangularity, elongation, density, equivalent diameter, orientation angle, second-order moment, Euler number, and fractal dimension.

[0069] By setting the infrared thermal imager's scanning range to full coverage and acquiring multiple frames of images, blind spots can be effectively avoided. At the same time, multiple frames of data provide redundant information for subsequent analysis, thereby improving detection reliability. The adaptive threshold filtering algorithm can dynamically adjust the noise reduction parameters according to the local characteristics of the image. Compared with the fixed threshold method, it can more accurately remove noise, retain the temperature anomaly details to the greatest extent, and improve image quality. The dynamic threshold marking method of abnormal points fully considers the working environment of the photovoltaic panel and the measurement error of the equipment. Compared with the static threshold, it is more in line with the actual situation and effectively reduces the probability of false detection and missed detection. The connectivity analysis is combined with the region expansion algorithm. The abnormal sub-regions are preliminarily divided through the connected region algorithm, and then the region is expanded according to the spatial distance threshold. It can not only quickly locate the abnormal area, but also integrate the adjacent tiny abnormal points into a complete defect area, accurately define the defect boundary, and ultimately achieve high-precision and high-efficiency detection and identification of abnormal temperature areas on the photovoltaic panel surface.

[0070] like Figure 3 As shown, step S20 specifically includes:

[0071] S21: controlling the laser speckle interferometer to scan the surface of the photovoltaic panel with a pulsed laser of a preset frequency. During the scanning process, using the spatial light modulator in the laser speckle interferometer to modulate the laser speckle pattern in real time, so that the laser speckle pattern forms a dynamically changing modulated speckle field on the surface of the photovoltaic panel;

[0072] The photovoltaic panel to be inspected was mounted horizontally on a vibration-isolating platform (with a vibration suppression rate greater than 90%). The laser speckle interferometer was adjusted to maintain a perpendicular position with a distance of 0.8 meters from the panel surface. The laser transmitter was turned on, with a pulse frequency of 10 kHz, a laser wavelength of 532 nm, and a power of 15 mW. A spatial light modulator was controlled to modulate the laser speckle pattern in real time. The modulation method involved generating a pseudo-random phase mask sequence using an FPGA (Field-Programmable Gate Array) controller. Each phase mask had a size of 1024 × 768 pixels and a phase modulation depth of 0–2π. The phase mask sequence was applied to the spatial light modulator at a frequency of 50 Hz, resulting in a dynamically changing modulated speckle field on the panel surface. During the modulation process, a closed-loop control system monitored the laser power stability in real time to ensure power fluctuations were within ±2%.

[0073] The FPGA controller is connected to the spatial light modulator via a high-speed data interface. Its function is to generate a pseudo-random phase mask sequence and load the phase mask onto the spatial light modulator at a frequency of 50 Hz, realizing real-time modulation of the laser speckle pattern. At the same time, it also receives laser power monitoring data from the photodiode and dynamically adjusts the laser drive current through the built-in PID control algorithm and closed-loop control system to ensure that the laser power fluctuation is less than ±2%. In addition, the FPGA controller also generates synchronization signals to control the pulsed laser and high-speed camera respectively, ensuring the precise timing synchronization of laser emission, speckle field modulation and image acquisition.

[0074] S22: collecting a modulated speckle image sequence corresponding to the modulated speckle field on the surface of the photovoltaic panel by a camera at a frame rate synchronized with the pulsed laser;

[0075] A high-speed camera synchronized with a pulsed laser to acquire modulated speckle images. A trigger signal generator synchronized the camera and laser, setting the camera frame rate to 50 fps (matching the phase mask update frequency), exposure time to 10 μs, and aperture to F5.6. During acquisition, an ambient light suppression device (narrowband filter with a center wavelength of 532 nm and a bandwidth of 10 nm) was used to reduce ambient light interference. 100 frames of modulated speckle images were acquired continuously to form a complete image sequence. Each frame was immediately transmitted to an industrial computer via Gigabit Ethernet for caching, ensuring temporal and spatial consistency of the image sequence.

[0076] S23: Processing the modulated speckle image sequence using a phase unwrapping algorithm based on deep learning to obtain restored phase information, and extracting surface deformation data of the photovoltaic panel from the restored phase information;

[0077] The acquired modulated speckle image sequence is input into the phase unwrapping algorithm module based on deep learning. This module adopts an improved U-Net network architecture and is specifically implemented as follows:

[0078] The data preprocessing unit performs normalization on the input modulated speckle image sequence, converting the pixel values ​​of each frame to the range [0, 1] through a linear mapping. This unit outputs a normalized image sequence with a uniformly constrained pixel value distribution, effectively eliminating data fluctuations caused by differences in lighting conditions and camera parameters between image batches.

[0079] The feature extraction unit extracts features from the normalized image sequence using an encoder network. The network consists of five convolutional blocks, each consisting of two 3×3 convolutional layers and a ReLU (Rectified Linear Unit) activation function. As the network depth increases, the feature map size is halved (from the initial 1280×800 to 80×50), while the number of channels is doubled (from 64 to 1024 channels). This feature extraction unit outputs a set of multi-scale feature maps, including intermediate feature maps from each encoder layer (for subsequent skip connections) and a final high-level abstract feature map. These feature maps capture key information in the speckle image, such as phase gradients and texture variations, providing a rich semantic representation for phase unwrapping. The intermediate feature maps are passed to the corresponding layers of the phase unwrapping unit, and the final feature map serves as the initial input to the decoder.

[0080] The phase unwrapping unit, used in a decoder network with skip connections, concatenates the intermediate feature maps of each encoder layer with the decoder feature maps of the corresponding layer, gradually recovering high-resolution phase information. The specific process involves upsampling the high-level feature maps output by the encoder (via transposed convolution or bilinear interpolation); channel-wise concatenating the upsampled feature maps with the intermediate feature maps of the corresponding encoder layer; fusing the concatenated features through a convolutional layer to generate a more accurate phase representation; and repeating the upsampling, concatenation, and feature fusion process until the original resolution phase map is restored. This unit ultimately outputs a continuous phase map with the same resolution as the input speckle image (1280×800 pixels) and a phase value range of [-π, π]. Each pixel value in the phase map corresponds to the phase information at that point on the photovoltaic panel surface and is directly passed to the deformation data extraction unit for physical quantity conversion.

[0081] During network training, a dataset containing 10,000 pairs of simulated speckle images and real phase maps was used for training. The weighted mean square error (W-MSE) was used as the loss function, and the weights were adaptively adjusted according to the phase gradient.

[0082] The deformation data extraction unit is used to calculate the surface deformation data from the recovered phase information according to the formula Δd = λ・φ / (4π・cosθ), where λ is the laser wavelength, φ is the phase change, and θ is the incident angle (θ=0° in this embodiment).

[0083] S24: Based on a preset deformation threshold, the surface deformation data is analyzed by a cluster analysis algorithm to determine an abnormal deformation area.

[0084] The extracted surface deformation data is imported into the cluster analysis module, specifically implemented as follows: In the data preprocessing step, the deformation data is Gaussian filtered (σ = 1.5) to remove high-frequency noise. In the feature extraction step, statistical features of the deformation data, including mean, standard deviation, skewness, and kurtosis, are calculated. In the cluster analysis step, the DBSCAN algorithm (density-based spatial clustering application) is used for anomaly detection, with parameters set to a neighborhood radius ε = 0.05 μm and a minimum number of points (MinPts) = 10. In the abnormal region identification step, regions in the clustering results that deviate from the main cluster center by more than three standard deviations are marked as potential abnormal regions. In the threshold screening step, the maximum deformation of potential abnormal regions is compared with a preset deformation threshold (0.1 μm in this example). Regions exceeding the threshold are identified as abnormal deformation regions. In the result verification step, morphological operations (dilation and erosion) are used to optimize the boundaries of the identified abnormal deformation regions, ultimately generating a detection report containing the location, area, and maximum deformation of the abnormal regions.

[0085] like Figure 4 As shown, step S30 specifically includes:

[0086] S31: Controlling the ultrasonic flaw detector to perform ultrasonic scanning on the surface of the photovoltaic panel line by line, and synchronously collecting ultrasonic echo data of each scanning point;

[0087] The photovoltaic panel to be inspected was fixed horizontally on a testing platform. The ultrasonic flaw detector was adjusted so that it was perpendicular to the panel surface at a distance of 0.5 meters. A 5MHz dual-element probe (10mm×10mm crystal, 50mm focal length) was used. A suitable amount of water-based coupling agent was applied between the probe and the panel surface to ensure effective transmission of ultrasonic waves into the panel. The ultrasonic flaw detector parameters were set as follows: transmit voltage 100V, gain 40dB, sampling frequency 100MHz, and sampling depth 100mm. An automatic scanning system controlled the probe to scan the panel surface line by line at a speed of 0.5m / s, with a scan pitch of 2mm and a scan width covering the entire panel surface. During the scanning process, ultrasonic echo data was synchronously collected at each scan point. Each echo signal consisted of 1024 sampling points, and the data was formatted as a time-domain waveform. The collected echo data was transmitted in real time to an industrial computer via a USB interface and stored as a binary file with the scan position coordinates included in the file name.

[0088] S32: Input the collected ultrasonic echo data into a pre-trained convolutional neural network model and output the probability value of the presence of a defect at each scanning point;

[0089] After the collected ultrasonic echo data is input into a pre-trained convolutional neural network model, the model extracts feature information from the data through calculations and processing at each layer, and ultimately outputs a probability value for the presence of a defect at each scan point. The probability value ranges from [0, 1], and the closer the value is to 1, the higher the probability of a defect at that scan point. This model can use a modified ResNet-18 architecture, and its specific structure is as follows:

[0090] The data preprocessing module, which serves as the model's input frontend, normalizes the raw ultrasonic echo data, mapping the amplitude of each sampling point to the range [-1, 1]. This operation eliminates signal strength variations between different scanning points caused by factors such as equipment fluctuations and material differences, standardizes the data distribution, and provides a unified and standardized data foundation for subsequent feature extraction, enhancing the stability of model training and prediction.

[0091] The feature extraction module consists of a 1D convolutional layer, a batch normalization layer, and a ReLU activation function connected in sequence. The 1D convolutional layer contains 64 convolution kernels of size 7×1. It performs sliding convolution on the normalized echo data with a stride of 2. Through local calculations between the convolution kernels and the data, it extracts the time domain features of the echo signal and generates 64 feature maps. Subsequently, the batch normalization layer normalizes the feature maps to accelerate model convergence and reduce the risk of overfitting. The ReLU activation function introduces nonlinearity into the model, enhancing its ability to express complex features. The feature maps output by this module are passed to the residual block design module.

[0092] The residual block design module consists of multiple cascaded residual blocks. Each residual block contains two 3×1 convolutional layers, interspersed with a batch normalization layer and a ReLU activation function. Through a residual connection structure, the residual block directly adds the input to the convolutional layer output, allowing the network to skip some layers and directly learn residual information during learning. This design effectively solves the vanishing gradient problem during deep network training, enabling the model to build a deeper network structure, thereby extracting more advanced and complex feature information and improving the model's ability to capture subtle defect characteristics. The residual blocks within the module sequentially process the feature maps, ultimately outputting the feature maps after deep feature extraction to the global pooling and classification module.

[0093] The global pooling and classification module first compresses the feature maps output by the residual block into a one-dimensional vector through a global average pooling layer, reducing the data dimension while retaining key feature information. This is then connected to a fully connected layer, which calculates the one-dimensional vector based on the weight parameters learned from the training data and ultimately outputs a probability value for the presence of a defect at each scan point. This probability value reflects the model's prediction of whether each scan point has a defect and is used to subsequently determine abnormal areas.

[0094] The model was trained on a dataset of 10,000 samples, of which 5,000 were defective samples (covering various defect types, such as cracks, delamination, and bubbles) and 5,000 were normal samples. The Adam (Adaptive Moment Estimation) optimizer was used during training, iteratively updating the model parameters at a learning rate of 0.001. Each training batch consisted of 64 samples, with a total of 100 rounds. A binary cross-entropy loss function was used to quantify the difference between the predicted probability and the true sample label (1 for defective and 0 for normal). The model minimized the loss function through a backpropagation algorithm, continuously adjusting internal parameters to maximize its ability to distinguish between defective and normal samples, thereby accurately identifying and predicting defect characteristics in ultrasonic echo data.

[0095] S33: According to a preset probability threshold, the area where the scanning points having probability values ​​higher than the preset probability threshold are located is determined as an ultrasonic echo abnormal area.

[0096] First, the probability of a defect at each scan point output by the convolutional neural network model is compared against a pre-set probability threshold (set to 0.7 in this example). Scanning points with a probability greater than 0.7 are marked as potential defect points, and their two-dimensional (X, Y) coordinates on the photovoltaic panel surface are recorded to form an initial set of potential defect points. Subsequently, these marked potential defect points are spatially clustered, and the density clustering algorithm (DBSCAN) is used to merge regions. Specifically, a neighborhood radius ε = 5 mm is set around each potential defect point, searching for other potential defect points within a 5 mm radius of the center point. Furthermore, a minimum number of points, MinPts = 3, is defined, meaning that if a point's neighborhood contains at least three potential defect points, these points are grouped together. The DBSCAN algorithm traverses all potential defect points, merging those with similar spatial distances that meet the density threshold to form multiple clusters. After clustering, each cluster is validated. The area of ​​each region is calculated, and areas with an area less than 2mm² are eliminated (this threshold is set based on the minimum size of common defects in photovoltaic panels). Such areas are usually identified as misjudgments caused by noise interference. For the retained areas, their geometric features are further calculated, including parameters such as centroid coordinates, boundary perimeter, and aspect ratio. Finally, the clustered areas that have been screened and feature calculated are identified as ultrasonic echo abnormality areas. To facilitate subsequent analysis and recording, each abnormal area is assigned a unique number, and key information such as the area's coordinate range, area, and centroid coordinates is stored in the test result database. At the same time, in the visual interface, each abnormal area is marked with different colors to intuitively display the defect distribution on the photovoltaic panel surface.

[0097] The advantages of this technical solution are that, at the data acquisition level, it adopts line-by-line ultrasonic scanning and synchronously collects echo data at each scanning point to ensure full coverage of the photovoltaic panel surface. Combined with specific probe parameters and coupling agents, it ensures the effective transmission of ultrasonic waves and improves data quality. In the data analysis stage, it uses a pre-trained convolutional neural network model to process echo data, extracts time domain features through an improved ResNet-18 architecture, and uses residual connections to solve the gradient problem of deep networks. Compared with traditional analysis methods, it can capture subtle defect features and output the defect probability value of each scanning point. In terms of defect judgment, the scanning points are screened and divided into regions through the combination of preset probability thresholds and the DBSCAN clustering algorithm, which can not only effectively eliminate single misjudgment points, but also fully outline the defect area and reduce missed detections and false detections.

[0098] like Figure 5 As shown, step S40 specifically includes:

[0099] S41: Controlling a visible light multi-band imager to perform a regional scan on the surface of the photovoltaic panel to obtain a multi-band spectral image covering a preset spectral range, wherein each pixel corresponds to a local area on the surface of the photovoltaic panel, and recording spectral data of the local area in different bands;

[0100] The photovoltaic panel to be inspected was mounted horizontally on an automated inspection platform. A visible light multi-band imager was used. This device covers a spectral range of 400-1000 nm, achieves a spectral resolution of 5 nm, and a spatial resolution of 0.5 mm / pixel. Before inspection, a standard D65 light source (color temperature 6500K, color rendering index Ra ≥ 90) was used to uniformly illuminate the panel, with the intensity controlled at 1000 ± 50 lux. The imager was positioned so that its optical axis was perpendicular to the panel surface and a distance of 1.2 meters was maintained to ensure full coverage of the panel (2m × 1m). The imager parameters were set as follows: scanning speed 0.1 m / s, exposure time 120 ms, and band selection covering the full spectral range of 400-1000 nm. Spectral images were collected from 121 bands. Progressive scanning was used, with a 15% overlap between adjacent scan lines to avoid blind spots. During the scanning process, the imager automatically records the spectral data of each pixel in different bands. The data format is a 16-bit unsigned integer and is stored in a standard format file with geographic coordinate information.

[0101] S42: preprocessing the acquired multi-band spectral image to obtain a preprocessed spectral image;

[0102] First, radiometric calibration is performed. Using reference data collected from a standard diffuse reflectance panel supplied with the imager, the image's raw DN (Digital Number) values ​​are converted to actual reflectance data, eliminating errors caused by variations in light intensity and device response. Next, a median filter algorithm is used to reduce noise in the image, using a 3×3 filter window to remove randomly distributed salt-and-pepper noise while preserving spectral feature details. For images with banding noise, banding repair is performed using the correlation between adjacent bands. Anomalous banding is corrected by interpolating the difference in mean values ​​between adjacent bands. Finally, the processed image undergoes geometric correction. Based on the coordinates of pre-set control points on the photovoltaic panel surface, bilinear interpolation is used to map image pixels to their correct geographic coordinates. The resulting preprocessed spectral image is output at a uniform resolution of 1000×1000 pixels for subsequent spectral feature extraction.

[0103] S43: extracting spectral feature data from the preprocessed spectral image using a spectral unmixing method;

[0104] A linear mixing model (LMM) was used for spectral unmixing. First, based on the material properties and common defect types of photovoltaic panels, six endmember spectra were selected: photovoltaic panel substrate, glass cover, EVA film, cracks, stains, and bubbles. An endmember library was constructed. For the preprocessed spectral image, the spectral curve of each pixel was considered a linear combination of the endmember spectra. The mixing coefficient was calculated using the non-negative least squares (NNLS) method, as follows:

[0105] R ( λ )=∑ a i E i ( λ )+ ϵ ( λ ),in, R ( λ ) is the pixel spectrum, E i ( λ ) is the i endmember spectra, a i is the mixing coefficient, ϵ ( λ) is the residual, and the value of i ranges from 1 to n. The endmember abundance matrix of each pixel is obtained. Twelve spectral feature data items, including the abundance value of each endmember, the Spectral Angle Mapper (SAM) distance, and the Spectral Information Divergence (SID), are extracted to form the feature vector of each pixel.

[0106] S44: Compare the spectral feature data with a pre-established defect spectral feature library of photovoltaic panels pixel by pixel. When the difference between the spectral feature data and the defect spectral features in the defect spectral feature library exceeds a set threshold, determine the corresponding area as a spectral feature abnormal area.

[0107] The extracted spectral signature data is compared pixel by pixel with a pre-established defect spectral signature library. The defect spectral signature library contains 2,000 sets of manually annotated standard defect spectral data, covering defect samples of different types and severity. During the comparison process, the Spectral Angle Mapper (SAM) algorithm is used to calculate the angle between the pixel spectrum and the defect spectrum in the library. The formula is: . Represents the angle between the pixel spectrum and the defect spectrum in the library, which is used to measure the similarity of the two spectral shapes; It is the spectral vector of the pixel to be detected, which is composed of the spectral data of each band of the pixel; is the spectral vector of the standard defect in the defect spectral feature library, which contains the data of each band of the marked defect sample; It is the dot product of two vectors, that is, the corresponding band data are multiplied and then summed; and are the modulus lengths of the spectrum vector to be detected and the spectrum vector of the standard defect, respectively.

[0108] When the spectral angle θ If the pixel's value is greater than a set threshold (set to 0.15 radians in this example) and the spectral information divergence (SID) is greater than 2.0, the corresponding area is determined to be a spectral abnormality region. The identified abnormal pixels are clustered using the 8-neighbor connectivity algorithm, merging adjacent abnormal pixels to form a complete abnormal region. Geometric parameters such as the area, centroid coordinates, and shape factor of each region are calculated. Finally, a visual inspection result map with defect annotations and a detailed inspection report are generated.

[0109] The advantages of the above technical solution are: it uses multi-band spectral image acquisition to cover the preset spectral range, and each pixel records spectral data in different bands, which can capture subtle spectral feature differences on the surface of photovoltaic panels and provide rich information for defect detection; it eliminates noise and errors through preprocessing to improve image quality; it uses spectral unmixing methods to extract spectral feature data, which can effectively separate the spectral contributions of different material components; it can accurately identify various types of defects by comparing them pixel by pixel with the pre-established defect spectral feature library, and the setting of difference thresholds can flexibly adjust the detection sensitivity to adapt to the detection needs of different types of photovoltaic panels, thereby improving the accuracy and reliability of detection.

[0110] Step S43 may specifically include: S431: calculating the mean and standard deviation of the spectral data of each band, and determining the initial unmixing threshold based on the mean and the standard deviation; S432: based on the spectral similarity of adjacent pixels, iteratively adjusting the initial unmixing threshold through a sliding window algorithm to form a dynamic unmixing threshold; S433: based on the dynamic unmixing threshold, using a non-negatively constrained linear spectral unmixing model to decompose the preprocessed spectral image into spectral feature data of different material components.

[0111] In step S431, the statistical features of the pre-processed multi-band spectral image are calculated according to the band dimension. For each band, the spectral values ​​of all pixels in the image are traversed, and their arithmetic mean is calculated as the mean μ. At the same time, the standard deviation σ is calculated to quantify the degree of data dispersion. For example, in a specific band, the mean is obtained by summing the spectral values ​​of 1000×1000 pixels and dividing it by the total number, and then the data fluctuation range is calculated according to the standard deviation formula. Based on these statistical values, the initial unmixing threshold is set to T0=μ+1.5σ. The principle of setting this threshold is that the mean represents the average level of the spectral data in this band, and 1.5 times the standard deviation excludes extreme values ​​caused by noise while retaining most of the valid signals, thereby preliminarily distinguishing meaningful spectral features from background noise, providing a basic threshold reference for subsequent unmixing processing.

[0112] In step S432, a sliding window algorithm is used to spatially adaptively optimize the initial threshold. A 5×5 pixel window is slid pixel by pixel across the spectral image. For each window's central pixel, the spectral similarity is calculated between it and the other 24 pixels within the window. The specific calculation method is as follows: the pixel's spectral data is treated as a high-dimensional vector, and the spectral similarity is measured by calculating the cosine of the angle between the vectors. A smaller angle indicates greater spectral similarity. If more than 70% of the pixels in the window have a spectral angle less than 0.1 radians with the central pixel (corresponding to higher similarity), the spectral consistency in that region is considered high, and the threshold corresponding to that window is lowered by 10% to enhance sensitivity to subtle spectral differences. Conversely, if more than 30% of the pixels have a spectral angle greater than 0.2 radians with the central pixel (corresponding to lower similarity), the threshold is increased by 10% to enhance robustness to noise. Through this window-by-window iterative adjustment, a dynamic unmixing threshold matrix that changes with spatial position is formed. This matrix can automatically adjust the threshold according to the local spectral characteristics, while maintaining the overall unmixing accuracy and better adapting to the differences in spectral complexity in different regions of the image.

[0113] In step S433, non-negative constrained linear spectral unmixing is performed on each pixel based on the dynamic threshold matrix. The pixel's spectral response is considered to be a mixture of multiple material components (endmembers) in varying proportions, with each endmember corresponding to the characteristic spectrum of a specific material (e.g., glass, silicon substrate, EVA film, crack, or contamination). A system of linear equations is established to represent the pixel spectrum as a linear combination of the endmember spectra, where the combination coefficients are the abundance values ​​of each material. Two constraints are imposed during the solution process: first, all abundance values ​​must be non-negative (consistent with physical reality, i.e., the material content cannot be negative), and second, the sum of all abundance values ​​must be 1 (indicating that the pixel is composed entirely of these endmember materials). The system of equations is solved iteratively using the non-negative least squares method: all abundance values ​​are initialized to 0, then the abundance values ​​are gradually adjusted over 200 iterations. Each iteration, the residual between the fitted spectrum and the actual pixel spectrum for the current abundance combination is calculated, and the abundance value is adjusted based on the comparison of the residual with the dynamic threshold. When the residual is less than the dynamic threshold at the corresponding position, the fit is considered to have achieved satisfactory accuracy and the iteration is stopped. The final output is the abundance matrix of each material component in each pixel, which accurately quantifies the spatial distribution of different materials on the photovoltaic panel surface.

[0114] like Figure 6 As shown, step S50 specifically includes:

[0115] S51: controlling the excitation coil of the eddy current testing device to scan the surface of the photovoltaic panel line by line with a preset multi-frequency excitation signal, wherein the multi-frequency excitation signal includes at least three different frequency components, and the amplitude and phase of each frequency component are pre-configured according to the material characteristics of the photovoltaic panel;

[0116] The photovoltaic panel to be inspected was fixed to a testing platform. A multi-frequency eddy current tester equipped with an absolute coil probe with an outer diameter of 10 mm was selected. Based on the material characteristics of the photovoltaic panel (silicon substrate thickness 0.2 mm, resistivity approximately 10 Ω cm), a three-frequency excitation signal was configured: a low-frequency component of 10 kHz (penetration depth approximately 2.5 mm for detecting deep-seated defects), a mid-frequency component of 50 kHz (penetration depth approximately 1.1 mm for detecting mid-layer defects), and a high-frequency component of 200 kHz (penetration depth approximately 0.5 mm for detecting surface defects). The amplitude ratio of each frequency component was set to 1:0.8:0.6, and the phase differences were set to 0°, 90°, and 180° to enhance the response to defects of different depths. A mechanical scanning platform was controlled to drive the probe along the photovoltaic panel surface line by line at a speed of 0.05 m / s, with a scanning pitch of 1 mm to ensure full inspection coverage. During the scanning process, the tester used DDS (direct digital synthesis) technology to generate a stable multi-frequency excitation signal. The power amplifier drove the excitation coil to ensure output power fluctuations within ±1%.

[0117] S52: collecting eddy current response signals corresponding to each excitation frequency through a receiving coil, performing fast Fourier transform on the collected eddy current response signals, and obtaining complex impedance data at different frequencies;

[0118] The receiving coil synchronously acquires eddy current response signals at each excitation frequency. The sampling frequency is set to 1 MHz, and each sampling duration is 10 ms, ensuring that at least 10 data points are collected during each excitation cycle. The acquired time-domain signals are converted to digital signals using a 24-bit analog-to-digital converter (ADC) and transmitted to an embedded processor for real-time processing. A fast Fourier transform (FFT) is performed on the time-domain signals at each sampling point to convert them to the frequency domain. Specifically, a radix-2 FFT algorithm is used to process the 1024-point data and calculate the complex impedance value (Z = R + jX, where R is the resistance component and X is the reactance component) for each frequency component. To improve the signal-to-noise ratio, the FFT results for each detection point are averaged five times. Finally, complex impedance data at three frequencies, 10 kHz, 50 kHz, and 200 kHz, are obtained and stored in a three-dimensional matrix format.

[0119] S53: A multi-frequency adaptive weight fusion algorithm is used to dynamically adjust the weight according to the signal-to-noise ratio of the complex impedance data at each frequency, and the complex impedance data at different frequencies are fused into a comprehensive characteristic index;

[0120] When processing complex impedance data at different frequencies, an adaptive weight fusion algorithm is employed to dynamically assign weights based on the signal quality at each frequency to enhance the representation of defect features. The specific implementation process is as follows: The signal-to-noise ratio (SNR) is calculated for the complex impedance data at each frequency (10kHz, 50kHz, and 200kHz). Signal power is calculated by calculating the variance of the impedance data at all test points at that frequency, reflecting the overall signal fluctuation strength. Noise power is estimated by analyzing the impedance fluctuations in adjacent known defect-free areas and representing the background noise level. For example, at 10kHz, if the impedance variance of the test area is large while the fluctuations in the defect-free area are small, the SNR at that frequency is higher, indicating greater sensitivity to defects. Weights are then proportionally assigned based on the SNR values ​​of each frequency: the SNR of each frequency is divided by the sum of the SNRs of all frequencies to obtain a normalized weight. For example, if the SNRs for 10kHz, 50kHz, and 200kHz are 8, 5, and 3, respectively, their weights would be 0.5, 0.3125, and 0.1875, respectively. This approach ensures that frequencies with higher signal-to-noise ratios account for a greater proportion of the final fusion, improving the reliability of defect detection. For each inspection point, the modulus difference (i.e., the absolute value of the impedance change) between its impedance value and the defect-free reference value at each frequency is calculated. This difference is then multiplied by the corresponding frequency weight and summed to obtain a comprehensive characteristic index. For example, if the impedance changes at a point at three frequencies are 0.2Ω, 0.15Ω, and 0.1Ω, respectively, and the corresponding weights are 0.5, 0.3125, and 0.1875, the comprehensive index is 0.2×0.5 + 0.15×0.3125 + 0.1×0.1875 = 0.153Ω.

[0121] S54: comparing the comprehensive characteristic index with a preset threshold value, and when the comprehensive characteristic index exceeds the preset threshold value, marking the corresponding detection point as a suspected defect point;

[0122] The calculated comprehensive characteristic index is compared with a preset threshold (set to 3.5 standard deviations in this example). The standard deviation is obtained through statistical analysis of the comprehensive characteristic index of known defect-free areas. When the comprehensive characteristic index of a test point exceeds the threshold, the point is marked as a suspected defect point, and its coordinates are recorded. To reduce misjudgments, marked suspected defect points undergo a secondary verification: their impedance changes at at least two frequencies simultaneously exceed their respective local thresholds (2 standard deviations of the impedance value in the normal area at each frequency). Only points that meet these conditions are retained as valid suspected defect points.

[0123] S55: Perform spatial clustering on all suspected defect points, merge interconnected suspected defect points into the same area, and determine the abnormal area of ​​eddy current signal.

[0124] A spatial cluster analysis is performed on all valid suspected defect points, and they are merged using the 8-neighbor connectivity rule and region growing algorithm. The specific steps are as follows: Step 1: Select an unprocessed suspected defect point as a seed point; Step 2: Check all points within its 8-neighborhood. If a point is also a suspected defect point and is less than 2mm from the seed point, add it to the current region; Step 3: Repeat Step 2 for each newly added point until further expansion is impossible; Mark the current region as a complete eddy current signal anomaly area and record its boundary coordinates, area, perimeter, and other geometric parameters; Repeat Steps 1 to 4 until all suspected defect points have been processed.

[0125] Finally, a detection result map containing all abnormal areas is generated. Different colors in the map distinguish different types of defects (such as cracks, delamination, pores, etc.), and preliminary classification is performed using preset classification rules (based on the shape factor, impedance change pattern, etc. of the abnormal area).

[0126] like Figure 7 As shown, step S60 may specifically include:

[0127] S61: normalizing the original coordinate data of the temperature abnormality region, the deformation abnormality region, the ultrasonic echo abnormality region, the spectral feature abnormality region, and the eddy current signal abnormality region to obtain the normalized coordinate data of each abnormal region;

[0128] The original coordinate data for temperature anomalies detected by infrared thermal imaging, deformation anomalies detected by laser speckle interferometry, echo anomalies identified by ultrasonic testing, spectral feature anomalies determined by multi-band spectral analysis, and signal anomalies detected by eddy current testing were all imported into the data processing platform. Because the coordinate systems of different testing devices differ—for example, infrared thermal imagers measure position in pixels while ultrasonic flaw detectors use millimeters—they require unified processing. Using the minimum-maximum normalization method, the horizontal and vertical coordinates of all anomaly areas are adjusted to a range of 0 to 1. This involves first finding the minimum and maximum values ​​of all coordinate data. For each coordinate value, the minimum value is subtracted from the minimum value and then divided by the difference between the maximum and minimum values ​​to obtain the normalized coordinates. For example, if the original horizontal coordinate of a temperature anomaly area has a minimum value of 100 pixels and a maximum value of 800 pixels, and the original horizontal coordinate of a point is 300 pixels, then the normalized horizontal coordinate of that point is (300 - 100) ÷ (800 - 100), which is approximately 0.286. After such processing, the coordinate data of the abnormal area obtained by different detection methods all have the same measurement scale, which is convenient for subsequent fusion analysis.

[0129] S62: Calculate the geometric characteristic parameters of each abnormal area respectively, and assign corresponding weights to the geometric characteristic parameters of each abnormal area according to a preset weight assignment rule;

[0130] For each abnormal region that has undergone normalization, a series of geometric characteristic parameters are calculated. These parameters include: area (the number of pixels or actual area within the region); perimeter (the sum of the lengths of the region's boundaries); centroid coordinates (representing the geometric center of the region); major and minor axis lengths (describing the region's shape and dimensions); and circularity (comparing the area and perimeter to determine how closely the region's shape resembles a circle). Values ​​closer to 1 indicate a more circular shape. For example, for an abnormal ultrasonic echo region with a calculated area of ​​50 square millimeters and a perimeter of 30 millimeters, its circularity is calculated to be approximately 0.70. These geometric characteristic parameters are weighted according to pre-defined rules. Because area directly reflects defect severity, the area parameter is given a higher weight of 0.4. Perimeter, which reflects the complexity of the defect boundary, is given a weight of 0.2. Centroid coordinates are used to determine defect location and are weighted 0.1. Major and minor axis lengths facilitate defect shape determination and are both weighted 0.1. Circularity, which aids in defect type differentiation, is weighted 0.1.

[0131] S63: determining overlapping portions of multiple abnormal regions through an intersection operation based on the standardized coordinate data, the geometric feature parameters, and weights corresponding to the geometric feature parameters, and marking the overlapping portions as initial defect candidate regions;

[0132] Using GIS (Geographic Information System) spatial analysis tools, such as ArcGIS software or open-source libraries like GDAL, an intersection operation is performed on five abnormal regions: temperature, deformation, ultrasonic echo, spectral signature, and eddy current signal, based on standardized coordinate data and calculated geometric feature parameters. These abnormal regions are treated as polygon layers, and polygon overlay analysis is used to identify overlapping areas. Specifically, each pixel or unit area within each region is individually evaluated. If a region exists in two or more abnormal regions, it is marked as an overlapping region. For example, if an area appears in the temperature abnormality region, the ultrasonic echo abnormality region, and the spectral signature abnormality region, it will be included in the initial defect candidate region. When determining overlapping regions, the confidence level of the overlapping regions is weighted based on the weights previously assigned to each geometric feature parameter. Feature parameters with higher weights have a greater influence on the overlap determination. After this weighted evaluation, the overlapping regions identified become initial defect candidate regions, providing a basis for further precise defect location determination.

[0133] S64: Perform morphological filtering on the initial defect candidate area to obtain the final defect position.

[0134] The initial defect candidate regions are optimized using a morphological filtering algorithm. First, an erosion operation is performed, using a 3×3 square template. This template is then slid sequentially over the candidate region. If all pixels within the area covered by the template belong to the candidate region, the pixel at the center of the template is retained; otherwise, it is deleted. This method removes isolated pixels and small glitches on the candidate region's boundaries, eliminating minor interference areas. After the erosion operation is completed, a dilation operation is performed. Using the same 3×3 square template, the center of the template is aligned with pixels in the candidate region. If at least one pixel within the template belongs to the candidate region, the pixel at the center of the template is marked as part of the candidate region. This restores the area reduced by erosion and connects adjacent small regions. Repeated erosion and dilation operations (in this example, three cycles are performed) gradually remove discrete noise points and discontinuous areas, smoothing the defect boundary. Ultimately, a clearly defined and precisely positioned defect region is obtained, output as polygonal vector data containing detailed coordinate information, area size, shape characteristics, and other attributes. This data can be directly applied to photovoltaic panel defect repair and quality assessment.

[0135] This step accurately integrates data from different detection methods (temperature, deformation, ultrasonic, spectral, and eddy current signals) through standardization and the application of weighting rules, ensuring the proper quantification and fusion of the characteristics of each abnormal area. Intersection operations identify overlapping areas, further improving the accuracy of defect location. Morphological filtering effectively removes noise and misjudgments, ensuring accurate identification of the final defect location.

[0136] like Figure 8 As shown, step S70 specifically includes:

[0137] S71: Obtaining temperature characteristics, deformation characteristics, echo characteristics, spectral characteristics, and eddy current characteristics from the surface temperature distribution image, surface deformation data, ultrasonic echo data, spectral characteristic data, and eddy current signal characteristic data corresponding to the defect position, respectively;

[0138] For temperature characteristics, we extract parameters such as the maximum temperature, average temperature, temperature gradient (the ratio of the maximum temperature difference within the area to the distance), and hot spot area percentage in the defect area from the infrared thermal image. For example, a defect area with a maximum temperature of 65°C, an average temperature of 58°C, and a temperature gradient of 0.8°C / mm indicates localized heating anomalies.

[0139] For deformation characteristics, the maximum deformation, average deformation, and deformation direction consistency index of the defect area (the direction of the principal axis of deformation is determined through principal component analysis) are calculated from the laser speckle pattern interferometry data. For example, if the maximum deformation of a certain area is 0.15mm and the average deformation is 0.08mm, and the deformation direction is concentrated on the X-axis, it may indicate the presence of a linear crack.

[0140] For echo characteristics, wavelet transform is performed on the ultrasonic echo data to extract energy distribution, peak frequency, and time-domain characteristics (such as echo time delay and attenuation coefficient). For example, if the echo energy in a certain area is 30% lower than that in the normal area, and the peak frequency is offset by 20kHz, it may indicate the presence of internal delamination defects.

[0141] For spectral features, we extract the reflectance ratio (e.g., 650nm / 800nm), absorption peak position offset, and spectral angular distance of characteristic bands from multi-band spectral images. For example, a region with a significantly reduced reflectance at 650nm and a 5nm shift in the absorption peak toward longer wavelengths is consistent with material oxidation.

[0142] For eddy current characteristics, principal component analysis (PCA) was performed on the complex impedance data to extract the contribution rate of the first three principal components, the ratio of the real to imaginary impedance parts, and the frequency response characteristics. For example, at a frequency of 10kHz, the real impedance of a certain area increased by 25% and the imaginary impedance decreased by 15%, consistent with the characteristics of a metal material defect.

[0143] S72: Determine weights corresponding to the temperature feature, the deformation feature, the echo feature, the spectral feature, and the eddy current feature, respectively, according to a weight distribution model;

[0144] The weight assignment model is constructed using the Analytic Hierarchy Process (AHP), which includes the following steps: In the judgment matrix construction step, the influence of different features on defect types is determined based on expert experience or historical data. For example, for crack defects, the weight of the deformation feature is 0.35, the weight of the echo feature is 0.3, the weight of the temperature feature is 0.15, the weight of the spectral feature is 0.1, and the weight of the eddy current feature is 0.1. In the consistency verification step, the maximum eigenvalue and consistency index of the judgment matrix are calculated to ensure that the weight assignment is reasonable. For example, when the consistency ratio (CR) is less than 0.1, the judgment matrix is ​​considered to have satisfactory consistency. In the dynamic adjustment step, the weights are dynamically adjusted based on the depth information of the defect location. For example, for surface defects, the weight of the spectral feature is increased to 0.25; for internal defects, the weight of the echo feature is increased to 0.4.

[0145] S73: performing weighted fusion on the temperature feature, the deformation feature, the echo feature, the spectrum feature, the eddy current feature, and their corresponding weights to obtain a comprehensive feature vector;

[0146] Before weighted fusion, the temperature, deformation, echo, spectral, and eddy current features are normalized to eliminate any discrepancies in their values ​​due to different dimensions. For the temperature feature, the lowest temperature detected in all data points is subtracted from the original temperature value, then divided by the temperature range, to map it to the [0, 1] interval. Deformation features (in millimeters) are normalized using the "feature value / maximum deformation" method, based on the maximum deformation within the detection range. Similar minimum-maximum normalization is used for the energy values ​​in the echo features, the reflectivity of the spectral features, and the impedance values ​​of the eddy current features. After normalization, each feature is weighted according to the weights determined by the weight allocation model in step S72. Assume that the temperature feature weight is 0.2, the deformation feature weight is 0.3, the echo feature weight is 0.25, the spectral feature weight is 0.15, and the eddy current feature weight is 0.1. Furthermore, the normalized temperature feature value is 0.7, the deformation feature value is 0.6, the echo feature value is 0.8, the spectral feature value is 0.4, and the eddy current feature value is 0.5. In this case, the weight of the temperature feature is 0.7 × 0.2 = 0.14, the weight of the deformation feature is 0.6 × 0.3 = 0.18, the weight of the echo feature is 0.8 × 0.25 = 0.2, the weight of the spectral feature is 0.4 × 0.15 = 0.06, and the weight of the eddy current feature is 0.5 × 0.1 = 0.05. Finally, the weighted values ​​of each feature are arranged in the order of temperature, deformation, echo, spectrum, and eddy current to form a five-dimensional vector [0.14, 0.18, 0.2, 0.06, 0.05], which is the composite feature vector. This vector, through scientific normalization and weighted calculation, not only preserves the importance differences of features obtained by different detection methods, but also unifies multi-source heterogeneous data into a standard format.

[0147] S74: Analyzing the comprehensive feature vector using a deep learning-based defect type recognition model to determine the type of surface defect of the photovoltaic panel;

[0148] An optimized convolutional neural network (CNN) is used as the defect type recognition model. The model's input layer receives the comprehensive feature vector obtained in step S73. To adapt to the network structure, the vector is reshaped into a three-dimensional tensor of 1×5×1, where 1 represents the number of samples, 5 corresponds to the five feature dimensions, and 1 represents the number of channels.

[0149] The main body of the network consists of three convolutional layers, two pooling layers, and two fully connected layers. The first convolutional layer uses 16 1×3 convolution kernels with a stride of 1 to extract features from the input tensor and explore correlation patterns between different features. This layer is followed by a max pooling layer with a 1×2 pooling window and a stride of 2 to reduce data dimensionality and computational complexity. The second convolutional layer uses 32 1×3 convolution kernels to further extract high-order features. This layer is followed by an average pooling layer with a 1×2 pooling window and a stride of 2. After two convolutional and pooling operations, the output is flattened into a one-dimensional vector and connected to a fully connected layer with 128 neurons and then a fully connected layer with 64 neurons. Reluctant unit (ReLU) activation functions are used between fully connected layers to enhance the network's nonlinear representation capabilities. The final output layer has five neurons corresponding to five common defect types in photovoltaic panels (cracks, delamination, bubbles, stains, and oxidation). A softmax activation function is used to output the probability value for each category.

[0150] During the model training phase, a dataset containing 20,000 labeled samples was used, with 16,000 samples used as the training set, 2,000 samples used as the validation set, and 2,000 samples used as the test set. The Adam optimizer was used, with an initial learning rate of 0.001 and a learning rate decay of 0.9 every 10 training epochs. The cross-entropy loss function was used as the loss function, and the network parameters were updated using the backpropagation algorithm during training, for a total of 50 training epochs. When the validation set loss stopped decreasing and the accuracy stabilized, training was stopped and the optimal model parameters were saved.

[0151] In practical applications, the trained model inputs the comprehensive feature vector to be tested. Among the five probability values ​​output, the category with the highest probability is the type of photovoltaic panel surface defect determined by the model. For example, if the output probability vector is [0.05, 0.85, 0.03, 0.04, 0.03], the photovoltaic panel is determined to have a delamination defect on its surface. The probability value output by the model also serves as a reference for the confidence level of the judgment result.

[0152] S75: Determine the size of the surface defect of the photovoltaic panel according to the temperature characteristics, the deformation characteristics and the echo characteristics.

[0153] First, we use quantification methods tailored to the different characteristic data to obtain preliminary size information. For temperature characteristics, we use the temperature difference between the defective area and the normal area in the infrared thermal image to define the defect. We set the temperature gradient threshold at 0.3°C / mm, then use an edge detection algorithm (Canny algorithm) to identify the boundaries of areas where the temperature gradient exceeds the threshold. We then calculate the equivalent diameter of this area as the defect size based on the temperature characteristic. For example, the calculated equivalent diameter of a defective area is 8mm.

[0154] For deformation characteristics, based on surface displacement data obtained by laser speckle interferometry, the deformation variable is used as the judgment basis. Areas with deformation variables exceeding 0.03mm are considered defects. A contour tracking algorithm is used to extract the regional contour, and the area enclosed by the contour is calculated and converted into an equivalent diameter. Assume that the calculated equivalent diameter of a certain area is 6mm.

[0155] Based on the echo characteristics, the echo data collected by ultrasonic flaw detection is analyzed, and the defect range is determined by the attenuation degree and time delay of the echo signal. When the echo signal energy attenuation exceeds 40% and the time delay exceeds the normal range (set to ±0.5μs), the corresponding area is marked as the defect area. The threshold segmentation algorithm is used to extract the defect area, and its longest straight-line distance is calculated as the defect size based on the echo characteristics. For example, the measured value of a certain area is 7mm.

[0156] Finally, a dynamic weighting strategy is used to fuse the three dimensions, depending on the defect type. Weights are set based on historical data statistics and expert experience: for crack defects, the deformation feature weight is 0.4, the temperature feature weight is 0.3, and the echo feature weight is 0.3; for delamination defects, the echo feature weight is 0.5, the temperature feature weight is 0.3, and the deformation feature weight is 0.2. The dimensions corresponding to each feature are multiplied by the weights and then summed to obtain the final defect size. For example, for a crack defect, the final calculated defect size is: 8 × 0.3 + 6 × 0.4 + 7 × 0.3 = 7.2 mm, which is used as the accurate determination of the surface defect size of the photovoltaic panel.

[0157] like Figure 9 As shown, this embodiment provides a photovoltaic panel surface defect detection system based on physical property analysis, the system is used to perform any of the aforementioned methods, and the system includes:

[0158] An infrared thermal imaging detection module is used to scan the entire photovoltaic panel using an infrared thermal imager to obtain a surface temperature distribution image of the photovoltaic panel and determine temperature abnormality areas based on the surface temperature distribution image;

[0159] A laser speckle deformation detection module is used to scan the surface of the photovoltaic panel using a laser speckle interferometer to obtain surface deformation data of the photovoltaic panel and determine the abnormal deformation area based on the surface deformation data;

[0160] An ultrasonic flaw detection module is used to perform ultrasonic scanning on the surface of the photovoltaic panel using an ultrasonic flaw detector, obtain ultrasonic echo data on the surface of the photovoltaic panel, and determine an ultrasonic echo abnormal area based on the ultrasonic echo data;

[0161] A spectral feature analysis module is used to obtain a spectral image of the photovoltaic panel surface using a visible light multi-band imager, extract spectral feature data from the spectral image based on a spectral unmixing algorithm, and determine spectral feature abnormality areas based on the spectral feature data;

[0162] An eddy current signal detection module is used to detect the surface area of ​​the photovoltaic panel using eddy current detection equipment to obtain eddy current signal characteristic data, and determine the eddy current signal abnormal area based on the eddy current signal characteristic data;

[0163] a defect location determination module, configured to determine the defect location based on the abnormal temperature area, the abnormal deformation area, the abnormal ultrasonic echo area, the abnormal spectrum feature area, and the abnormal eddy current signal area;

[0164] The defect recognition and analysis module is used to determine the photovoltaic panel surface defect recognition results, including the type and size of the surface defects, based on the surface temperature distribution image, surface deformation data, ultrasonic echo data, spectral feature data and eddy current signal feature data corresponding to the defect location.

[0165] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the above-mentioned method embodiment and will not be repeated here.

[0166] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the program implements any one of the above methods when executed by a processor.

[0167] Reference below Figure 10 , which shows a schematic structural diagram of a computer system 800 of an electronic device suitable for implementing an embodiment of the present invention. Figure 10 The electronic device shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention. Figure 10As shown, computer system 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage unit 808 into a random access memory (RAM) 803. Various programs and data required for the operation of computer system 800 are also stored in RAM 803. CPU 801, ROM 802, and RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to bus 804.

[0168] The following components are connected to the I / O interface 805: an input section 806 including a keyboard, mouse, and the like; an output section 807 including devices such as a cathode ray tube (CRT), a liquid crystal display (LCD), and speakers; a storage section 808 including a hard disk; and a communication section 809 including a network interface card such as a LAN card or a modem. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as needed. Removable media 811, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 810 as needed, so that computer programs read from the removable media can be installed in the storage section 808 as needed.

[0169] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A photovoltaic panel surface defect detection method based on physical property analysis, characterized in that: The following steps are involved: S10: Scanning the entire photovoltaic panel using an infrared thermal imager to obtain a surface temperature distribution image of the photovoltaic panel, and determining an abnormal temperature area based on the surface temperature distribution image; S20: Scanning the surface of the photovoltaic panel using a laser speckle interferometer to obtain surface deformation data of the photovoltaic panel, and determining an abnormal deformation area based on the surface deformation data; S30: performing ultrasonic scanning on the surface of the photovoltaic panel using an ultrasonic flaw detector to obtain ultrasonic echo data of the surface of the photovoltaic panel, and determining an ultrasonic echo abnormal area based on the ultrasonic echo data; S40: Acquire a spectral image of the photovoltaic panel surface using a visible light multi-band imager, extract spectral feature data from the spectral image based on a spectral unmixing algorithm, and determine a spectral feature abnormality region based on the spectral feature data; S50: Using eddy current detection equipment to detect the surface area of ​​the photovoltaic panel, obtain eddy current signal characteristic data, and determine the eddy current signal abnormal area based on the eddy current signal characteristic data; S60: determining a defect location according to the abnormal temperature area, the abnormal deformation area, the abnormal ultrasonic echo area, the abnormal spectrum feature area, and the abnormal eddy current signal area; S70: Determine a photovoltaic panel surface defect recognition result including the type and size of the surface defect based on the surface temperature distribution image, surface deformation data, ultrasonic echo data, spectral feature data, and eddy current signal feature data corresponding to the defect location; Step S60 specifically includes: S61: normalizing the original coordinate data of the temperature abnormality region, the deformation abnormality region, the ultrasonic echo abnormality region, the spectral feature abnormality region, and the eddy current signal abnormality region to obtain the normalized coordinate data of each abnormal region; S62: Calculate the geometric characteristic parameters of each abnormal area respectively, and assign corresponding weights to the geometric characteristic parameters of each abnormal area according to a preset weight assignment rule; S63: determining overlapping portions of multiple abnormal regions through an intersection operation based on the standardized coordinate data, the geometric feature parameters, and weights corresponding to the geometric feature parameters, and marking the overlapping portions as initial defect candidate regions; S64: Perform morphological filtering on the initial defect candidate area to obtain the final defect position.

2. The method according to claim 1, characterized in that Step S10 specifically includes: S11: setting the scanning range of the infrared thermal imager to cover the entire surface of the photovoltaic panel, performing a full-area scan of the photovoltaic panel, and obtaining multiple frames of surface temperature distribution images; S12: performing noise reduction processing on the surface temperature distribution image using an adaptive threshold filtering algorithm to obtain a preprocessed surface temperature distribution image; S13: Calculating the temperature deviation between the temperature value of each pixel in the pre-processed surface temperature distribution image and a preset normal temperature range, and marking the pixel whose temperature deviation exceeds a dynamic threshold as an abnormal point; S14: Connectivity analysis is performed on the pixel points marked as outliers, and adjacent outliers are classified into the same outlier sub-region using a connected region algorithm to generate a preliminary region map of the temperature anomaly region; within each outlier sub-region of the preliminary region map, a region expansion algorithm is used to further connect adjacent but not yet connected outliers based on a preset spatial distance threshold to form a temperature anomaly region with a continuous boundary.

3. The method according to claim 1, characterized in that Step S20 specifically includes: S21: controlling the laser speckle interferometer to scan the surface of the photovoltaic panel with a pulsed laser of a preset frequency. During the scanning process, using the spatial light modulator in the laser speckle interferometer to modulate the laser speckle pattern in real time, so that the laser speckle pattern forms a dynamically changing modulated speckle field on the surface of the photovoltaic panel; S22: collecting a modulated speckle image sequence corresponding to the modulated speckle field on the surface of the photovoltaic panel by a camera at a frame rate synchronized with the pulsed laser; S23: Processing the modulated speckle image sequence using a phase unwrapping algorithm based on deep learning to obtain restored phase information, and extracting surface deformation data of the photovoltaic panel from the restored phase information; S24: Based on a preset deformation threshold, the surface deformation data is analyzed by a cluster analysis algorithm to determine an abnormal deformation area.

4. The method according to claim 1, wherein Step S30 specifically includes: S31: Controlling the ultrasonic flaw detector to perform ultrasonic scanning on the surface of the photovoltaic panel line by line, and synchronously collecting ultrasonic echo data of each scanning point; S32: Input the collected ultrasonic echo data into a pre-trained convolutional neural network model and output the probability value of the presence of a defect at each scanning point; S33: According to a preset probability threshold, the area where the scanning points having probability values ​​higher than the preset probability threshold are located is determined as an ultrasonic echo abnormal area.

5. The method according to claim 1, wherein Step S40 specifically includes: S41: Controlling a visible light multi-band imager to perform a regional scan on the surface of the photovoltaic panel to obtain a multi-band spectral image covering a preset spectral range, wherein each pixel corresponds to a local area on the surface of the photovoltaic panel, and recording spectral data of the local area in different bands; S42: preprocessing the acquired multi-band spectral image to obtain a preprocessed spectral image; S43: extracting spectral feature data from the preprocessed spectral image using a spectral unmixing method; S44: Compare the spectral feature data with a pre-established defect spectral feature library of photovoltaic panels pixel by pixel. When the difference between the spectral feature data and the defect spectral features in the defect spectral feature library exceeds a set threshold, determine the corresponding area as a spectral feature abnormal area.

6. The method according to claim 5, characterized in that Step S43 specifically includes: S431: Calculate the mean and standard deviation of the spectral data of each band, and determine an initial unmixing threshold according to the mean and the standard deviation; S432: Based on the spectral similarity of adjacent pixels, iteratively adjust the initial unmixing threshold using a sliding window algorithm to form a dynamic unmixing threshold; S433: Decomposing the preprocessed spectral image into spectral feature data of different material components using a non-negative constrained linear spectral unmixing model according to the dynamic unmixing threshold.

7. The method according to claim 1, characterized in that Step S50 specifically includes: S51: controlling the excitation coil of the eddy current detection device to scan the surface of the photovoltaic panel line by line with a preset multi-frequency excitation signal; S52: collecting eddy current response signals corresponding to each excitation frequency through a receiving coil, performing fast Fourier transform on the collected eddy current response signals, and obtaining complex impedance data at different frequencies; S53: A multi-frequency adaptive weight fusion algorithm is used to dynamically adjust the weight according to the signal-to-noise ratio of the complex impedance data at each frequency, and the complex impedance data at different frequencies are fused into a comprehensive characteristic index; S54: comparing the comprehensive characteristic index with a preset threshold value, and when the comprehensive characteristic index exceeds the preset threshold value, marking the corresponding detection point as a suspected defect point; S55: Perform spatial clustering on all suspected defect points, merge interconnected suspected defect points into the same area, and determine the abnormal area of ​​eddy current signal.

8. The method according to claim 1, characterized in that Step S70 specifically includes: S71: Obtaining temperature characteristics, deformation characteristics, echo characteristics, spectral characteristics, and eddy current characteristics from the surface temperature distribution image, surface deformation data, ultrasonic echo data, spectral characteristic data, and eddy current signal characteristic data corresponding to the defect position, respectively; S72: Determine weights corresponding to the temperature feature, the deformation feature, the echo feature, the spectral feature, and the eddy current feature, respectively, according to a weight distribution model; S73: performing weighted fusion on the temperature feature, the deformation feature, the echo feature, the spectrum feature, the eddy current feature, and their corresponding weights to obtain a comprehensive feature vector; S74: Analyzing the comprehensive feature vector using a deep learning-based defect type recognition model to determine the type of surface defect of the photovoltaic panel; S75: Determine the size of the surface defect of the photovoltaic panel according to the temperature characteristics, the deformation characteristics and the echo characteristics.

9. A photovoltaic panel surface defect detection system based on physical property analysis, characterized in that: The system is used to perform the method according to any one of claims 1 to 8, and the system comprises: An infrared thermal imaging detection module is used to scan the entire photovoltaic panel using an infrared thermal imager to obtain a surface temperature distribution image of the photovoltaic panel and determine temperature abnormality areas based on the surface temperature distribution image; A laser speckle deformation detection module is used to scan the surface of the photovoltaic panel using a laser speckle interferometer to obtain surface deformation data of the photovoltaic panel and determine the abnormal deformation area based on the surface deformation data; An ultrasonic flaw detection module is used to perform ultrasonic scanning on the surface of the photovoltaic panel using an ultrasonic flaw detector, obtain ultrasonic echo data on the surface of the photovoltaic panel, and determine an ultrasonic echo abnormal area based on the ultrasonic echo data; A spectral feature analysis module is used to obtain a spectral image of the photovoltaic panel surface using a visible light multi-band imager, extract spectral feature data from the spectral image based on a spectral unmixing algorithm, and determine spectral feature abnormality areas based on the spectral feature data; An eddy current signal detection module is used to detect the surface area of ​​the photovoltaic panel using eddy current detection equipment to obtain eddy current signal characteristic data, and determine the eddy current signal abnormal area based on the eddy current signal characteristic data; a defect location determination module, configured to determine the defect location based on the abnormal temperature area, the abnormal deformation area, the abnormal ultrasonic echo area, the abnormal spectrum feature area, and the abnormal eddy current signal area; The defect recognition and analysis module is used to determine the photovoltaic panel surface defect recognition results, including the type and size of the surface defects, based on the surface temperature distribution image, surface deformation data, ultrasonic echo data, spectral feature data and eddy current signal feature data corresponding to the defect location.

Images