A method and system for online defect detection of an unaddressed photovoltaic panel

Abstract

This invention discloses an online defect detection method for gridless photovoltaic (PV) panels, comprising the following steps: acquiring a lifetime distribution map of the gridless PV cell using time-resolved photoluminescence; acquiring a heat dissipation map of the gridless PV cell using optically induced phase-locked thermal imaging; and acquiring an ultrasonic image of the gridless PV cell using laser ultrasound. The lifetime distribution map, heat dissipation map, and ultrasonic image are preprocessed and registered to be uniformly mapped to the cell coordinate system, and minority carrier lifetime, heat dissipation power, ultrasonic propagation time delay offset, and scattering energy are extracted, which, together with the thermal phase, constitute multimodal features. Based on anomaly scores, defects are detected and classified according to the multimodal features of potential defect areas, yielding classification results for electrical defects, thermal defects, and mechanical defects. This invention also discloses a system for implementing this method. This invention achieves full-coverage detection of electrical, thermal, and mechanical defects.

Description

Technical Field

[0001] This invention relates to a method and system for testing photovoltaic panels, belonging to the field of battery module testing technology. Background Technology

[0002] Under high-speed production lines and complex process conditions, gridless solar cells are highly susceptible to various defects during manufacturing, interconnection, and packaging, including microcracks, increased local series resistance, grid breakage, voids in the back silver paste, surface contamination, and interlayer debonding. These defects not only reduce the photoelectric conversion efficiency of the cells but may also cause hot spot effects during long-term operation, accelerating performance degradation and directly threatening the 25-30 year service life target of photovoltaic modules.

[0003] Currently, commonly used detection methods mainly include electroluminescence (EL) and photoluminescence (PL) imaging. In existing technologies, EL imaging is widely used for module-level and finished cell manufacturing and maintenance testing because it can directly reveal structural defects such as microcracks, broken grids, and PID (Potential Influenced Defects). Chinese patent CN113257696A proposes a continuous EL testing scheme for modules of different sizes, which improves production line adaptability and testing efficiency by measuring the module length and automatically adjusting the EL camera position. However, this method requires electrical excitation, and its detection efficiency is limited by electrode contact conditions, making it difficult to adapt to gridless cell structures. US patent US20120142125A1 discloses a PL imaging system for silicon photovoltaic cell production lines. This system uses the above-mentioned bandgap light source excitation and PL emission imaging to spatially characterize the minority carrier lifetime, diffusion length, poor crystal structure, or impurity distribution within the crystal. This scheme demonstrates good performance in identifying microcracks, crystal defects, and regions that affect diffusion length. Although the PL method can reflect the minority carrier lifetime and recombination defects of materials to some extent, its results depend on brightness information and are easily affected by light source stability, surface texture, and environmental noise. It also has insufficient ability to detect microcracks and subsurface defects.

[0004] Furthermore, single-modal detection methods often struggle to achieve full coverage of electrical, thermal, and mechanical defects. For instance, lifetime decay defects are relatively easy to identify in physical modal analysis (PL), but microcracks are often difficult to detect using optical methods; while ultrasonic testing is sensitive to cracks, it is difficult to quantify the corresponding electrical losses. Existing detection technologies have shortcomings in multimodal fusion, quantitative defect assessment, and high-speed online adaptation. Summary of the Invention

[0005] To address the shortcomings of the prior art, the present invention aims to provide an online defect detection method for grid-free photovoltaic panels and an online defect detection system for grid-free photovoltaic panels. The purpose is to solve the problems that existing methods are difficult to adapt to grid-free photovoltaic panels and that single-mode detection methods are insufficient to fully cover possible defects, resulting in missed detections.

[0006] The technical solution of this invention is as follows: An online defect detection method for gridless photovoltaic panels, comprising the following steps:

[0007] The lifetime distribution map of the gridless photovoltaic cell was obtained by time-resolved photoluminescence, the heat dissipation map of the gridless photovoltaic cell was obtained by optically induced phase-locked thermal imaging, and the ultrasonic map of the gridless photovoltaic cell was obtained by laser ultrasound. The lifetime distribution map is composed of the curve of the luminous intensity of each pixel decaying over time. The heat dissipation map is composed of the phase-locked temperature amplitude and thermal phase of each pixel. The ultrasonic map is composed of the ultrasonic time-domain signal of each receiving point.

[0008] The phase-locked temperature amplitude and ultrasonic wave of the lifetime distribution map, heat dissipation map and the ultrasonic wave are preprocessed and registered to be uniformly mapped to the cell coordinate system.

[0009] Minority carrier lifetime is extracted from the lifetime distribution map after preprocessing and registration; heat dissipation power density is extracted from the phase-locked temperature amplitude of the heat dissipation map after preprocessing and registration; and ultrasonic propagation time delay shift and scattered energy are extracted from the ultrasonic map after preprocessing and registration. The minority carrier lifetime, heat dissipation power density, thermal phase, ultrasonic propagation time delay shift and scattered energy constitute multimodal features.

[0010] Anomaly scores are calculated based on minority carrier lifetime, heat dissipation power density, ultrasonic propagation time delay offset, and scattering energy. Regions with anomaly scores exceeding a set threshold are considered potential anomaly regions. A multimodal attention convolutional neural network is used to detect and classify the multimodal features of potential anomaly regions, resulting in classifications of electrical, thermal, and mechanical defects.

[0011] Furthermore, the preprocessing includes normalizing the curve of luminescence intensity decay over time in the lifetime distribution map using the following formula:

[0012] ,

[0013] In the formula, It is the curve of luminescence intensity decaying over time in the lifetime distribution map. This is the corrected TRPL curve. Dark field, A draw. For time sample index;

[0014] The preprocessing includes bandpass filtering and normalization of the ultrasonic time-domain signal of the ultrasound image using the following formula:

[0015] ,

[0016] In the formula, This is the ultrasonic time-domain signal of the ultrasound image. The normalized waveform after bandpass filtering. It is a bandpass filter operator for 0.2–2 MHz. This represents the laser energy reading for each excitation.

[0017] Furthermore, the registration involves mapping the coordinates of the preprocessed lifetime distribution map, the phase-locked temperature amplitude of the heat dissipation map, and the ultrasonic map to the cell coordinates. Then, through encoder pulses, the discrete points collected during the rolling process are resampled to an equidistant grid by the delivery encoder, so that different modes are unified in time and space dimensions.

[0018] Furthermore, the extraction of minority carrier lifetime from the preprocessed and registered lifetime distribution map is achieved by estimating the minority carrier lifetime using the first-order moment method, with the following formula:

[0019] , ,

[0020] In the formula, For the estimated minority carrier lifetime, The curves showing the decay of luminescence intensity over time after preprocessing and registration are shown.

[0021] The heat dissipation power density is extracted from the phase-locked temperature amplitude of the preprocessed and registered heat dissipation map and calculated according to the following formula:

[0022] , ,

[0023] In the formula, For heat dissipation power density, The phase-locked loop temperature amplitude after registration. Background grayscale offset. The grayscale-temperature conversion factor obtained from standard heat source calibration. For thermal conductivity, For thermal diffusivity, The modulation angular frequency;

[0024] The formula for extracting the ultrasonic propagation time delay offset and calculating the ultrasonic propagation time delay offset in the scattered energy from the preprocessed and registered ultrasonic image is as follows:

[0025] , ,

[0026] In the formula, It is the ultrasonic propagation time delay offset (s). The standardized waveform for the registered ultrasound image. It is the reference waveform for a defect-free baseline. It is a related search displacement. This is a reference arrival time. It is the integration time;

[0027] The formula for calculating scattering energy is:

[0028] ,

[0029] In the formula It is the scattered energy (relative dimension). These are the upper and lower limits of the selected integration time window.

[0030] Furthermore, the anomaly score is determined by a weighted sum of robust normalization of minority carrier lifetime, robust normalization of thermal dissipation power density, and joint normalization of ultrasonic propagation time delay offset and scattered energy; the multimodal attention convolutional neural network performs defect detection and classification on the multimodal features of potentially anomalous regions, including:

[0031] Convolutional feature extraction is performed on the minority carrier lifetime, heat dissipation power density, thermal phase, and ultrasonic propagation time delay offset combined scattering energy; the convolutional feature extraction results of heat dissipation power density and thermal phase are convolutionally mapped and encoded, and the features are concatenated in the channel dimension, and then fused convolution to obtain thermal feature mapping;

[0032] The convolutional feature extraction results of minority carrier lifetime, the convolutional feature extraction results of ultrasonic propagation time delay offset and scattering energy, and the thermal feature mapping are used as input modes of multi-head attention. The enhanced features output by multi-head attention are aggregated and convolved to obtain a comprehensive feature tensor.

[0033] Based on the comprehensive feature tensor, a probability map of defect classification is output by the fully connected classification layer. The defect region in the probability map is judged by a set threshold, and the classification results of electrical defects, thermal defects and mechanical defects are determined.

[0034] Furthermore, to achieve the sorting and lifetime prediction of gridless photovoltaic panels, the method further includes the following steps: mapping the multimodal characteristics to the increment of the equivalent series resistance and the increment of the equivalent composite current density of the cell; calculating the effective power generation area loss based on the mechanical defect classification results; and estimating the power loss based on the increment of the equivalent series resistance, the increment of the equivalent composite current density, and the effective power generation area loss.

[0035] This scheme also achieves a direct quantitative correlation between defect information and power generation performance loss through feature mapping, so that the detection results can be directly used for cell sorting and life prediction.

[0036] Furthermore, the estimated power loss is a weighted sum of the increment of the equivalent series resistance, the increment of the equivalent composite current density, and the effective power generation area loss. The weighting coefficients are obtained by numerically calculating the partial derivatives of the diode model under nominal operating conditions.

[0037] Furthermore, the mapping relationship for the increment of the equivalent series resistance is as follows:

[0038] ,

[0039] The mapping relationship for the increment of the equivalent composite current density is as follows:

[0040] ,

[0041] In the formula, It is the increment of the equivalent series resistance. , The coefficients are obtained from linear regression of the calibration sample. The mean of the heat dissipation power density is a robustly standardized value. It is the increment of the equivalent composite current density. , These are the coefficients obtained from calibration. This represents the mean of the robustly standardized minority carrier lifetime.

[0042] The effective power generation area loss is the ratio of the area of ​​the mechanical defect region of the solar cell to the total area of ​​the solar cell.

[0043] Another technical solution of the present invention is: an online defect detection system for gridless photovoltaic panels, comprising:

[0044] A multimodal signal acquisition unit is used to acquire a lifetime distribution map of a gridless photovoltaic cell by time-resolved photoluminescence, a heat dissipation map of a gridless photovoltaic cell by optically induced phase-locked thermal imaging, and an ultrasonic map of a gridless photovoltaic cell by laser ultrasound. The lifetime distribution map is composed of the curve of the luminous intensity of each pixel decaying over time. The heat dissipation map is composed of the phase-locked temperature amplitude and thermal phase of each pixel. The ultrasonic map is composed of the ultrasonic time-domain signal of each receiving point.

[0045] The data preprocessing and registration unit is used to preprocess and register the phase-locked temperature amplitude and ultrasonic wave of the lifetime distribution map and heat dissipation map to uniformly map them to the cell coordinate system.

[0046] The feature extraction unit is used to extract minority carrier lifetime from the preprocessed and registered lifetime distribution map, extract heat dissipation power density from the phase-locked temperature amplitude of the preprocessed and registered heat dissipation map, and extract ultrasonic propagation time delay shift and scattered energy from the preprocessed and registered ultrasonic map. The minority carrier lifetime, heat dissipation power density, thermal phase, ultrasonic propagation time delay shift and scattered energy constitute multimodal features.

[0047] The fusion decision unit is used to calculate anomaly scores based on minority carrier lifetime, heat dissipation power density, ultrasonic propagation time delay offset, and scattering energy. Regions with anomaly scores exceeding a set threshold are considered potential anomaly regions. A multimodal attention convolutional neural network is used to detect and classify defects in the multimodal features of potential anomaly regions, resulting in classification results for electrical defects, thermal defects, and mechanical defects.

[0048] Furthermore, it includes a power assessment unit for mapping the multimodal characteristics to the increment of the equivalent series resistance of the battery cell and the increment of the equivalent composite current density, calculating the effective power generation area loss based on the mechanical defect classification results, and estimating the power loss based on the increment of the equivalent series resistance, the increment of the equivalent composite current density, and the effective power generation area loss.

[0049] The advantages of the invention compared to existing technologies are:

[0050] This invention proposes an online defect detection method for gridless photovoltaic cells based on time-resolved photoluminescence (TRPL), optically induced phase-locked-in thermal imaging (OP-LIT), and laser ultrasound (LU). TRPL captures the carrier lifetime decay process, quantitatively correlating electrical performance with optical information to achieve high-precision electrical diagnosis of complex defects. OP-LIT locates thermal power and phase anomalies, enabling not only quantitative identification of thermal defects such as hot spots and series resistance but also providing direct thermal input for power loss quantification. LU detects wave propagation delay and scattering energy, accurately identifying mechanical defects such as microcracks and debonding, effectively supplementing the detection of subsurface defects that optical methods cannot penetrate, achieving comprehensive diagnosis from surface to interior.

[0051] Simultaneously, a multimodal registration and cross-modal attention deep fusion framework is adopted to collaboratively model electrical, thermal, and mechanical information in a unified coordinate system, achieving accurate defect localization and type identification. Compared with single-modal or simple splicing methods, this invention uses a multimodal attention convolutional neural network to detect and classify defects based on multimodal features of potentially abnormal regions. It maintains high robustness under noise, illumination interference, and asynchronous sampling conditions, and introduces a joint decision mechanism of physical rules and network output to improve model interpretability and discrimination robustness. In the preferred technical solution, multimodal features are mapped to equivalent parameters ( , , This enables a quantitative assessment of the power loss caused by defects.

[0052] This invention employs a non-contact, online design, eliminating the need for physical probes and coupling media, thus avoiding any mechanical stress or damage to the solar cells. It achieves comprehensive detection and performance prediction of electrical, thermal, and mechanical defects, and can be integrated into high-speed production lines without disrupting cycle time, demonstrating significant industrial application value. Attached Figure Description

[0053] Figure 1 This is a schematic flowchart of an online defect detection method for a gridless photovoltaic panel, as illustrated in this embodiment.

[0054] Figure 2 This is a schematic diagram of the modules of an online defect detection system for gridless photovoltaic panels, as shown in the embodiment.

[0055] Figure 3 This is a schematic diagram illustrating the relationship between various data in the online defect detection method for gridless photovoltaic panels, as shown in this embodiment. Detailed Implementation

[0056] The present invention will be further described below with reference to embodiments, but these are not intended to limit the scope of the invention.

[0057] Please combine Figure 1 , Figure 3 As shown, the online defect detection method for gridless photovoltaic panels in this embodiment includes the following steps:

[0058] Step 1: Obtain the lifetime distribution map of the gridless photovoltaic cell by time-resolved photoluminescence, obtain the heat dissipation map of the gridless photovoltaic cell by optically induced phase-locked thermal imaging, and obtain the ultrasonic map of the gridless photovoltaic cell by laser ultrasound.

[0059] Specifically, step 1.1 involves using a pulsed laser to excite the solar cell (picosecond / nanosecond). With each pulse, the pixel emits light over tens of nanoseconds to tens of microseconds, gradually dimming. A high-speed camera is then used to collect and record the light intensity decay curve over time, creating a time curve for each pixel. A standard scattering sheet is first measured to obtain... The sample was then tested again. Finally, fitting and Its computational physical model is as follows:

[0060] ,

[0061] In the formula, The curve showing the decay of luminescence intensity over time as observed by TRPL. Sampling time (in seconds), For the instrument response function, Let m be the magnitude of the exponential component. Let m be the lifetime constant (in seconds) of the m-th exponential component. The number of exponential components. The background count is used. The formula is a physical model of the instrument response convolution, where multiple exponential decay terms characterize the minority carrier lifetime process, and the convolution term eliminates the influence of the instrument response function. The actual lifetime decay, after being time-blurred by the instrument, becomes the curve we see. This lays the foundation for subsequent lifetime estimation.

[0062] Step 1.2: The solar cell is excited by being illuminated with a modulated 20 Hz alternating light source. If there is a current loop, increased series resistance, or micro-short circuit in the illuminated area, it will periodically heat up. Therefore, an infrared camera is used to capture N frames of temperature images in one cycle. The N frames in one cycle are substituted into the following formula and two weighted sums are performed to obtain the phase-locked temperature amplitude and thermal phase of each pixel. The phase-locked demodulation calculation formula is as follows:

[0063] ,

[0064] ,

[0065] In the formula, The phase-locked temperature amplitude, The number of sampling frames within one modulation period. This represents the temperature perturbation value of the k-th frame after emissivity correction. Let k be the sampling time of the k-th frame. Modulate the angular frequency (rad / s) of the light source. This refers to the thermal phase. The formula extracts the amplitude and phase of the thermal response using Fourier components. , Two quantities can stably represent localized heating, thus enabling quantitative characterization of localized hotspots and series resistance. Large amplitude indicates strong heating, while phase lag allows for the deduction of heat diffusion paths and depth information.

[0066] Step 1.3: An elastic wave is excited using a pulsed laser. The wave propagates within the sheet and deforms or deflects when it encounters cracks or debonding. The waveform is received at different points using an LDV (Laser-Dependent Laser) and then converted into a time-domain signal through Fourier transform. The formula for synthesizing the time-domain waveform is as follows:

[0067] ,

[0068] In the formula, In position The ultrasonic time-domain signal received at the location, It is the sensor coordinate position. It is the sampling time of the waveform signal in the time domain. This is the frequency domain response at that location. This formula converts the frequency domain response into a time domain signal to preserve the scattering characteristics of cracks and debonding defects.

[0069] Step 2: Preprocess and register the phase-locked temperature amplitude and ultrasonic waveform of the lifetime distribution map, heat dissipation map to uniformly map them to the cell coordinate system.

[0070] This step includes, step 2.1: Denoising and correcting the lifetime distribution map (the curve of luminescence intensity decaying over time) and the ultrasonic map (ultrasonic time-domain signal).

[0071] Specifically, for the curve of luminous intensity decaying over time, by acquiring a completely dark field and a uniformly bright flat field, the curves are linearly normalized using the following formula, thus achieving curve standardization. The calculation formula is as follows:

[0072] ,

[0073] In the formula, This is the corrected TRPL curve. Dark field, A draw. This is a time sample index.

[0074] For ultrasonic time-domain signals, the original waveform is passed through a bandpass filter and then divided by the laser energy reading to obtain a waveform on the same scale, ensuring the comparability of ultrasonic signals under different experimental conditions. The calculation formula is as follows:

[0075] ,

[0076] In the formula, The normalized waveform after bandpass filtering. It is a bandpass filter operator for 0.2–2 MHz. This represents the laser energy reading for each excitation.

[0077] Step 2.2: Cross-modal geometric registration to cell coordinates. First, take a picture of a calibration board with corner points, and use least squares to obtain the coordinates. The processed lifetime distribution map, heat dissipation map, and ultrasonic map are all projected onto the same plane coordinate system, that is, the coordinates of any modal image are mapped to the cell coordinates. The calculation formula is as follows:

[0078] ,

[0079] In the formula, It is the pixel position in the cell coordinate system. It is the pixel position in the original coordinates of the sensor.

[0080] Then, through encoder pulses, the discrete points collected during rolling are resampled to an equidistant grid by the transport encoder, so that different modes are unified in time and space dimensions. The calculation formula is as follows:

[0081] ,

[0082] In the formula, The original sequence is, that is , or ; Equidistant position index The resampled value on, that is, the corresponding value , or , The coordinates of the measurement point in the solar cell coordinate system; It's encoder counting. It is a counting scale for unit displacement.

[0083] Step 3: Extract minority carrier lifetime from the preprocessed and registered lifetime distribution map, extract heat dissipation power density from the phase-locked temperature amplitude of the preprocessed and registered heat dissipation map, and extract ultrasonic propagation time delay shift and scattered energy from the preprocessed and registered ultrasonic map. The multimodal features are composed of minority carrier lifetime, heat dissipation power density, ultrasonic propagation time delay shift, scattered energy, and thermal phase of the heat dissipation map.

[0084] Specifically, step 3.1 involves estimating the minority carrier lifetime using the intensity-time weighted average, i.e., the first-order moment method, based on the preprocessed and registered lifetime distribution map. , The smaller the value, the more severe the composite effect and the worse the material quality. The formula is as follows:

[0085] , ,

[0086] In the formula, To estimate the minority carrier lifetime, no complex fitting was performed here. Instead, a robust lifetime index was calculated directly: the greater the brightness in later time periods, the longer the lifetime.

[0087] Step 3.2: Convert the phase-locked temperature amplitude into heat dissipation power density. This is used to quantify energy dissipation caused by series resistance or hot spots, and its calculation formula is as follows:

[0088] , ,

[0089] In the formula, the registered phase-locked temperature amplitude True temperature amplitude . Background grayscale offset. The gray-temperature conversion factor obtained for standard heat source calibration ( (grayscale) For thermal conductivity, For thermal diffusivity, This is the modulation angular frequency.

[0090] Step 3.3: Select the measurement point in the cell coordinate system. Standardized waveform of ultrasound After undergoing time synchronization and spatial registration processing in step 2.2, the waveform's time axis is now aligned with the laser excitation trigger signal. Because the time reference is consistent, the time variable can be... The equivalent is denoted as the time variable used in the cross-correlation calculation. Therefore, after unifying the time reference, a signal is obtained for cross-correlation calculation: by calculating the shift in wave arrival time through correlation, the propagation delay caused by cracks or debonding is detected, and the ultrasonic propagation time delay shift is obtained, the calculation formula of which is as follows:

[0091] , ,

[0092] In the formula, It is the ultrasonic propagation time delay offset (s). It is a test waveform. It is the reference waveform for a defect-free baseline. It is a related search displacement. This is a reference arrival time. It is the integration time. The arrival time is evidence of crack circumference and velocity variation, and how much difference remains after alignment reflects scattering and attenuation intensity.

[0093] The residual energy between the reference wave and the defect wave is measured using the scattering energy formula, which quantifies the crack intensity. The calculation formula is as follows:

[0094] ,

[0095] In the formula It is the scattered energy (relative dimension). These are the upper and lower limits of the selected integration time window. The larger the value, the more obvious the difference in waveform shape, the stronger the scattering and attenuation, which can lead to the conclusion that the defect is more serious.

[0096] Step 4: Calculate the anomaly score based on minority carrier lifetime, heat dissipation power density, ultrasonic propagation time delay offset, and scattering energy. Regions with anomaly scores exceeding a set threshold are considered potential anomaly regions. A multimodal attention convolutional neural network is used to detect and classify the multimodal features of the potential anomaly regions to obtain classification results for electrical defects, thermal defects, and mechanical defects, thus completing the defect diagnosis.

[0097] Specifically, step 4.1 involves constructing anomaly scores using weighted and standardized features. Based on the statistical significance of the multimodal features, defect areas are preliminarily identified, calculated using the following formula:

[0098] ,

[0099] In the formula, It is an anomaly score (the higher the score, the more abnormal). It is a robust standardization of low birth rate life expectancy. It is a robust standardization of heat dissipation power density. yes Joint normalization with scattering energy, , , It has three-branch weights. Before entering the multimodal attention convolutional neural network, compare... The value and the size of the set threshold, Areas where the value exceeds the set threshold are considered potentially abnormal areas. This step allows for the rapid screening of these potentially abnormal areas.

[0100] Step 4.2: Based on the initial screening, a multimodal attention convolutional neural network is used to perform deep fusion and classification of multimodal features in potentially anomalous regions. The input to the multimodal attention convolutional neural network includes five modal features: minority carrier lifetime. Heat dissipation power density thermal phase Ultrasonic propagation time delay offset and scattered energy Each modal input is processed by an independent convolutional encoder to extract spatial semantic features, which are represented as follows:

[0101] , , , ,

[0102] , , ,

[0103] in, For convolutional feature extraction functions, , They are respectively for , The encoding layer that performs convolutional mapping, This represents feature concatenation along the channel dimension. To fuse convolutional layers, thermal channels are constructed using... and The thermal feature map obtained by joint encoding is generated This also reflects the surface heating intensity (from Dominant) and thermal diffusion delay (by (Reflection), used to comprehensively characterize series resistance and hot spot defects.

[0104] Features from each modality are fed into a multi-head cross-modal attention mechanism for information exchange. The single-head form of the multi-head attention mechanism is calculated as follows:

[0105] ,

[0106] In the formula, It is a cross-modal attention output tensor. It is a query feature derived from one of the minority carrier lifetime, heat dissipation power, ultrasonic propagation time delay shift, and scattering energy. , These are the value features from the other modes. Each time, one mode is selected as Q, and the other two are selected as K / V (input modes: { , , Output the enhanced features of each modality. , , . The channel dimension is used for scaling. Then, it is fused into a comprehensive feature tensor through a channel aggregation convolutional layer.

[0107] ,

[0108] In the formula This is an aggregation convolution operation used to semantically integrate enhanced features from various modalities.

[0109] Subsequently, a fully connected classification layer is used to predict the probability of each pixel across multiple classes, yielding the probability distributions for three types of defects: electrical, thermal, and mechanical. The calculation formula is as follows:

[0110] ,

[0111] In the formula For pixels The defect category probability vector, It is the last thing the network passes through The three defect probability maps output by the classification layer correspond to the three physical domains of electricity, heat, and mechanics, respectively. This is the classification layer weight matrix. It is the bias vector;

[0112] ,

[0113] In the formula Characterizes the spatial distribution of defects of various physical types, where 1 represents a defect area and 0 represents a normal area. For the set threshold, If lifespan is extremely short in a certain place ( High), judged as an electrical defect area; if the local thermal amplitude is large and the phase lag is large ( High), judged as a hotspot; if the wave propagation delay is significant ( (High), judged as a crack.

[0114] The above process, through the combination of multimodal features and cross-modal deep learning, ultimately achieves accurate identification of electrical, thermal, and mechanical defects in gridless photovoltaic cells.

[0115] In order to enable the test results to be directly applied to cell sorting and life prediction, as a preferred embodiment, the method of this embodiment further includes step 5: mapping the multimodal features to the increment of the equivalent series resistance and the increment of the equivalent composite current density of the cell, calculating the effective power generation area loss based on the mechanical defect classification results, estimating the power loss based on the increment of the equivalent series resistance, the increment of the equivalent composite current density and the effective power generation area loss, and evaluating the power generation performance.

[0116] Specifically, step 5.1 involves mapping the average heat dissipation power density to the increment of the series resistance; and mapping the average minority carrier lifetime to the increment of the recombination current density, using the following formula:

[0117] , ,

[0118] In the formula, It is the increment of the equivalent series resistance. , The coefficients are obtained from linear regression of the calibration sample. To achieve a robust and standardized average heat dissipation power, It is the increment of the equivalent composite current density. , The coefficients obtained from calibration This represents the mean of the minority carrier lifetime, which is robustly standardized. A higher heat dissipation power density results in a greater increase in the equivalent series resistance. The larger the value, the worse the minority carrier lifetime, the stronger the recombination, and the greater the increase in the equivalent recombination current density. The larger.

[0119] Then, the crack mask ratio is calculated, representing the effective power generation area loss. The defect area is calculated from the defect detection results in step 4. Specifically, the probability map output by the mechanical defect branch of the multimodal attention convolutional neural network in step 4 is... Its value range is [0,1]. A threshold is set... A binary mask for defects can be obtained. :

[0120] ,

[0121] In the formula This is the threshold, typically set to 0.4–0.6, calibrated based on the ROC curve. Defect binary mask. Used to identify defective pixels, pixels with a value of 1 in the mask represent defective areas, and pixels with a value of 0 represent normal areas. Let the total number of defective pixels in the binary defect mask be... The actual area corresponding to each pixel in the battery cell coordinate system is (Determined by imaging calibration or coordinate mapping), the area of ​​the defect region can then be calculated as:

[0122] ,

[0123] Since the cracked area generates virtually no electricity, the effective area loss is estimated based on the area ratio. The calculation formula is:

[0124] .

[0125] Step 5.2: Simultaneous Combination , , The power loss under standard operating conditions is obtained. This is used to directly evaluate the grade of solar cells, and its calculation formula is as follows:

[0126] ,

[0127] In the formula This refers to the output power loss under standard operating conditions (STC). , , The sensitivity coefficient is obtained by taking the partial derivative of a diode model under nominal operating conditions.

[0128] Please combine Figure 2 As shown, another embodiment of the present invention is an online defect detection system for gridless photovoltaic panels that implements the method of the aforementioned embodiments, which specifically includes: a multimodal signal acquisition unit 1, a data preprocessing and registration unit 2, a feature extraction unit 3, a fusion decision unit 4, and a power evaluation unit 5.

[0129] The multimodal signal acquisition unit 1 includes a time-resolved photoluminescence (TRPL) detection module 101, an optically induced phase-locked thermal imaging (OP-LIT) detection module 102, and a laser ultrasound (LU) detection module 103.

[0130] The time-resolved photoluminescence (TRPL) detection module 101 is used to obtain a lifetime distribution map of the grid-less photovoltaic cell through time-resolved photoluminescence. The lifetime distribution map is composed of the curves showing the decay of the luminous intensity of each pixel over time. The optically induced phase-locked-in thermal imaging (OP-LIT) detection module 102 is used to obtain a heat dissipation map of the grid-less photovoltaic cell through optically induced phase-locked-in thermal imaging. The heat dissipation map is composed of the phase-locked-in temperature amplitude and phase demodulation map of each pixel. The laser ultrasound (LU) detection module 103 is used to obtain an ultrasonic image of the grid-less photovoltaic cell through laser ultrasound. The ultrasonic image is composed of the ultrasonic time-domain signals from each receiving point.

[0131] For details on the implementation of the multimodal signal acquisition unit 1, please refer to step 1 of the aforementioned embodiment.

[0132] The data preprocessing and registration unit 2 is used to preprocess and register the phase-locked temperature amplitude and ultrasonic wave of the lifetime distribution map and heat dissipation map to uniformly map them to the cell coordinate system. For the specific functional implementation process of the data preprocessing and registration unit 2, please refer to step 2 of the aforementioned embodiment.

[0133] Feature extraction unit 3 is used to extract minority carrier lifetime from the preprocessed and registered lifetime distribution map, extract heat dissipation power density from the phase-locked temperature amplitude of the preprocessed and registered heat dissipation map, and extract ultrasonic propagation time delay shift and scattered energy from the preprocessed and registered ultrasonic map. Minority carrier lifetime, heat dissipation power density, thermal phase, ultrasonic propagation time delay shift, and scattered energy constitute multimodal features. For the specific functional implementation process of feature extraction unit 3, please refer to step 3 of the aforementioned embodiment.

[0134] The fusion decision unit 4 is used to calculate anomaly scores based on minority carrier lifetime, heat dissipation power density, ultrasonic propagation time delay offset, and scattering energy. Regions with anomaly scores exceeding a set threshold are considered potential anomaly regions. A multimodal attention convolutional neural network is used to detect and classify defects in the multimodal features of potential anomaly regions, resulting in classifications of electrical, thermal, and mechanical defects. For a detailed explanation of the specific functional implementation of the fusion decision unit 4, please refer to step 4 of the aforementioned embodiment.

[0135] The power assessment unit 5 is used to map multimodal characteristics to the increments of the equivalent series resistance and the equivalent composite current density of the solar cell, calculate the effective power generation area loss based on the mechanical defect classification results, and estimate the power loss based on the increments of the equivalent series resistance, the equivalent composite current density, and the effective power generation area loss. For the specific functional implementation process of the power assessment unit 5, please refer to step 5 of the aforementioned embodiment.

[0136] This invention solves the problem that traditional methods cannot locate multiple types of defects. In complex process environments, it enables real-time online defect detection of grid-less photovoltaic cells, offering advantages such as high detection accuracy and strong adaptability, significantly improving the quality control and production line sorting efficiency of photovoltaic modules.

Claims

1. A method for online defect detection of gridless photovoltaic panels, characterized in that, Includes the following steps: The lifetime distribution map of the gridless photovoltaic cell was obtained by time-resolved photoluminescence, the heat dissipation map of the gridless photovoltaic cell was obtained by optically induced phase-locked thermal imaging, and the ultrasonic map of the gridless photovoltaic cell was obtained by laser ultrasound. The lifetime distribution map is composed of the curve of the luminous intensity of each pixel decaying over time. The heat dissipation map is composed of the phase-locked temperature amplitude and thermal phase of each pixel. The ultrasonic map is composed of the ultrasonic time-domain signal of each receiving point. The phase-locked temperature amplitude and ultrasonic wave of the lifetime distribution map, heat dissipation map and the ultrasonic wave are preprocessed and registered to be uniformly mapped to the cell coordinate system. Minority carrier lifetime is extracted from the lifetime distribution map after preprocessing and registration; heat dissipation power density is extracted from the phase-locked temperature amplitude of the heat dissipation map after preprocessing and registration; and ultrasonic propagation time delay shift and scattered energy are extracted from the ultrasonic map after preprocessing and registration. The minority carrier lifetime, heat dissipation power density, thermal phase, ultrasonic propagation time delay shift and scattered energy constitute multimodal features. Anomaly scores are calculated based on minority carrier lifetime, heat dissipation power density, ultrasonic propagation time delay offset, and scattering energy. Regions with anomaly scores exceeding a set threshold are considered potential anomaly regions. A multimodal attention convolutional neural network is used to detect and classify the multimodal features of potential anomaly regions, resulting in classifications of electrical, thermal, and mechanical defects. The multimodal features are mapped to the increments of the equivalent series resistance and the equivalent composite current density of the solar cell. The effective power generation area loss is calculated based on the mechanical defect classification results. The power loss is estimated based on the increments of the equivalent series resistance, the equivalent composite current density, and the effective power generation area loss. The estimated power loss is a weighted sum of the increments of the equivalent series resistance, the equivalent composite current density, and the effective power generation area loss. The weighting coefficients are obtained by numerically calculating the partial derivatives of the diode model under nominal operating conditions.

2. The online defect detection method for gridless photovoltaic panels according to claim 1, characterized in that, The preprocessing includes normalizing the curve of luminescence intensity decay over time in the lifetime distribution map using the following formula: ， In the formula, It is the curve of luminescence intensity decaying over time in the lifetime distribution map. This is the corrected TRPL curve. Dark field, A draw. For time sample index; The preprocessing includes bandpass filtering and normalization of the ultrasonic time-domain signal of the ultrasound image using the following formula: ， In the formula, This is the ultrasonic time-domain signal of the ultrasound image. The normalized waveform after bandpass filtering. It is a bandpass filter operator for 0.2–2 MHz. This represents the laser energy reading for each excitation.

3. The online defect detection method for gridless photovoltaic panels according to claim 1, characterized in that, The registration process involves mapping the coordinates of the preprocessed lifetime distribution map, the phase-locked temperature amplitude of the heat dissipation map, and the ultrasonic map to the cell coordinates. Then, through encoder pulses, the discrete points collected during the rolling process are resampled to an equidistant grid by the delivery encoder, so that different modes are unified in time and space dimensions.

4. The online defect detection method for gridless photovoltaic panels according to claim 1, characterized in that, The extraction of minority carrier lifetime from the preprocessed and registered lifetime distribution map is achieved by estimating the minority carrier lifetime using the first-order moment method, with the following formula: ， ， In the formula, For the estimated minority carrier lifetime, The curves showing the decay of luminescence intensity over time after preprocessing and registration are shown. The heat dissipation power density is extracted from the phase-locked temperature amplitude of the preprocessed and registered heat dissipation map and calculated according to the following formula: ， ， In the formula, For heat dissipation power density, The phase-locked loop temperature amplitude after registration. Background grayscale offset. The grayscale-temperature conversion factor obtained from standard heat source calibration. For thermal conductivity, For thermal diffusivity, The modulation angular frequency; The formula for extracting the ultrasonic propagation time delay offset and calculating the ultrasonic propagation time delay offset in the scattered energy from the preprocessed and registered ultrasonic image is as follows: ， ， In the formula, It is the ultrasonic propagation time delay offset (s). The standardized waveform for the registered ultrasound image. It is the reference waveform for a defect-free baseline. It is a related search displacement. This is a reference arrival time. It is the integration time; The formula for calculating scattering energy is: ， In the formula It is scattered energy. These are the upper and lower limits of the selected integration time window.

5. The online defect detection method for gridless photovoltaic panels according to claim 1, characterized in that, The anomaly score is determined by a weighted sum of robust normalization of minority carrier lifetime, robust normalization of heat dissipation power density, and joint normalization of ultrasonic propagation time delay offset and scattered energy; the multimodal attention convolutional neural network performs defect detection and classification on the multimodal features of potentially anomalous regions, including: Convolutional feature extraction is performed on the minority carrier lifetime, heat dissipation power density, thermal phase, and ultrasonic propagation time delay offset combined scattering energy; the convolutional feature extraction results of heat dissipation power density and thermal phase are convolutionally mapped and encoded, and the features are concatenated in the channel dimension, and then fused convolution to obtain thermal feature mapping; The convolutional feature extraction results of minority carrier lifetime, the convolutional feature extraction results of ultrasonic propagation time delay offset and scattering energy, and the thermal feature mapping are used as input modes of multi-head attention. The enhanced features output by multi-head attention are aggregated and convolved to obtain a comprehensive feature tensor. Based on the comprehensive feature tensor, a probability map of defect classification is output by the fully connected classification layer. The defect region in the probability map is judged by a set threshold, and the classification results of electrical defects, thermal defects and mechanical defects are determined.

6. The online defect detection method for gridless photovoltaic panels according to claim 1, characterized in that, The mapping relationship for the increment of the equivalent series resistance is as follows: ， The mapping relationship for the increment of the equivalent composite current density is as follows: ， In the formula, It is the increment of the equivalent series resistance. , The coefficients are obtained from linear regression of the calibration sample. The mean of the heat dissipation power density is a robustly standardized value. It is the increment of the equivalent composite current density. , These are the coefficients obtained from calibration. This represents the mean of the robustly standardized minority carrier lifetime. The effective power generation area loss is the ratio of the area of ​​the mechanical defect region of the solar cell to the total area of ​​the solar cell.

7. An online defect detection system for gridless photovoltaic panels, characterized in that, include: A multimodal signal acquisition unit is used to acquire a lifetime distribution map of a gridless photovoltaic cell by time-resolved photoluminescence, a heat dissipation map of a gridless photovoltaic cell by optically induced phase-locked thermal imaging, and an ultrasonic map of a gridless photovoltaic cell by laser ultrasound. The lifetime distribution map is composed of the curve of the luminous intensity of each pixel decaying over time. The heat dissipation map is composed of the phase-locked temperature amplitude and thermal phase of each pixel. The ultrasonic map is composed of the ultrasonic time-domain signal of each receiving point. The data preprocessing and registration unit is used to preprocess and register the phase-locked temperature amplitude and ultrasonic wave of the lifetime distribution map and heat dissipation map to uniformly map them to the cell coordinate system. The feature extraction unit is used to extract minority carrier lifetime from the preprocessed and registered lifetime distribution map, extract heat dissipation power density from the phase-locked temperature amplitude of the preprocessed and registered heat dissipation map, and extract ultrasonic propagation time delay shift and scattered energy from the preprocessed and registered ultrasonic map. The minority carrier lifetime, heat dissipation power density, thermal phase, ultrasonic propagation time delay shift and scattered energy constitute multimodal features. The fusion decision unit is used to calculate the anomaly score based on minority carrier lifetime, heat dissipation power density, ultrasonic propagation time delay offset and scattering energy. Regions with anomaly scores exceeding a set threshold are considered possible anomaly regions. A multimodal attention convolutional neural network is used to detect and classify the multimodal features of the possible anomaly regions to obtain classification results for electrical defects, thermal defects and mechanical defects. The power assessment unit is used to map the multimodal characteristics to the increment of the equivalent series resistance and the increment of the equivalent composite current density of the battery cell, calculate the effective power generation area loss based on the mechanical defect classification results, and estimate the power loss based on the increment of the equivalent series resistance, the increment of the equivalent composite current density, and the effective power generation area loss. The estimated power loss is a weighted sum of the increment of the equivalent series resistance, the increment of the equivalent composite current density, and the effective power generation area loss. The weighting coefficients are obtained by numerically calculating the partial derivatives of the diode model under nominal operating conditions.

Images