TOPCon battery piece defect detection device and detection method

Abstract

The invention relates to the technical field of photovoltaic module detection, in particular to a TOPCon battery piece defect detection device and a TOPCon battery piece defect detection method. The optical imaging module comprises a 25,000,000-pixel camera, a three-color integral light source array and a multispectral optical filter group, and the light source array is composed of red, green and blue LEDs and supports pulse synchronous triggering; the traditional algorithm processing module is integrated with a Halcon image processing library and is configured with a VarThreshold operator dynamic threshold segmentation unit, an adjustable kernel size corrosion filtering unit and a light and shade separation unit, and an Erosion Rectangle1 operator is used for corrosion filtering; the AI defect detection module is internally provided with an improved YOLOv8 neural network model, and the model structure comprises a Backbone network embedded with a CBAM attention mechanism, a BiFPN network bidirectional feature pyramid and a dynamic loss function optimizer; according to the technical scheme, the limitation problems of low positioning precision, poor detection efficiency, single defect identification dimension and the like of an artificial vision or traditional probe contact detection means can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic module detection, and specifically relates to a TOPCon cell defect detection device and a detection method. Background Art

[0002] In the current booming photovoltaic industry, the market's demand for high-efficiency and stable photovoltaic cells is increasing day by day. The TOPCon technology has emerged as the times require. Its core goal is to improve the photoelectric conversion efficiency of cells, reduce production costs, and enhance the competitiveness of photovoltaic products in the energy market. With the deepening global dependence on clean energy, large-scale photovoltaic power generation projects urgently need higher-efficiency cell technologies to increase power generation and reduce the cost per kilowatt-hour. Distributed photovoltaic power generation systems, especially in the building field, also require photovoltaic cells to generate more electricity in a limited installation space. The TOPCon technology has become one of the key solutions to meet these needs. There are the following technical bottlenecks in traditional cell detection: Difficulty in identifying complex defects: Traditional detection relies on a single algorithm, and for complex defects such as color spots and scratches, the recognition rate is not high. For example, in the photoluminescence detection technology, when traditional image processing or a single AI model faces such defects, the misjudgment rate is as high as over 0.5%, making it difficult to accurately judge the true defect condition of the cell.

[0003] Half-cell compatibility issue: When detecting two half-cells in parallel, there is optical crosstalk in traditional equipment, resulting in an increase in the missed detection rate and being unable to accurately detect the defects of half-cells. In the current situation where double half-cells are widely used, it seriously affects the reliability of the detection results. Data island phenomenon: Traditional detection equipment lacks deep integration with the manufacturing execution system (MES), and the defect data detected cannot be associated with process parameters. When a cell has a defect, it is impossible to quickly trace back to the process problems in the production process, which is not conducive to timely adjusting the production process and improving product quality.

[0004] Based on the above problems, there is an urgent need for a technical solution that can solve the limitations of existing detection methods such as manual vision or traditional probe contact, including low positioning accuracy, poor detection efficiency, and single defect recognition dimension. Summary of the Invention

[0005] In view of the low positioning accuracy, poor detection efficiency, and single defect recognition dimension of the current detection methods such as manual vision or traditional probe contact, the present application provides a TOPCon cell defect detection device and a detection method to solve the above problems.

[0006] A TOPCon cell defect detection device includes: Optical imaging module: includes a 25-megapixel camera, a three-color integral light source array, and a multi-spectral filter set. The light source array consists of red, green, and blue LEDs and supports pulse synchronization triggering. ‌Traditional algorithm processing module‌: Integrates the Halcon image processing library, configures the VarThreshold operator dynamic threshold segmentation unit, the adjustable kernel size erosion filter unit and the light and dark separation unit, and uses the ErosionRectangle1 operator for erosion filtering; ‌AI Defect Detection Module‌: Built-in improved YOLOv8 neural network model, the model structure includes Backbone network embedded with CBAM attention mechanism, BiFPN network bidirectional feature pyramid and dynamic loss function optimizer; Dynamic Threshold Adjustment Module: Connects to the MES system via the OPC UA interface to obtain diffusion temperature, phosphorus source concentration and annealing rate parameters in real time. It has a built-in nonlinear coupling calculation unit and outputs a dynamic correction coefficient for the defect judgment threshold. ‌Data Integration Module‌: Integrates MySQL database and MQTT protocol communication unit, configures defect distribution heat map generator, process parameter trend analyzer and SPC control chart generator; ‌Calibration module‌: includes a high-precision linear motor driven conveyor, a photoelectric positioning sensor and a white balance calibrator.

[0007] Adopting the above technical solution: The above solution can solve the problems of relatively high false detection rate of image processing algorithms in traditional technologies, single light source and easy islanding of data. The capture of passivation layer reflection anomalies, hidden crack scattering characteristics and grid line deformation provided by this application can improve detection efficiency; the cascade of traditional algorithms and AI models can realize the enhancement of subtle defect features through the CBAM attention mechanism, thereby increasing the hidden crack detection rate from 89% to 99.5%.

[0008] Preferably, the kernel size adjustment formula of the adjustable kernel size erosion filter unit is: ; Among them, K is the size of the corrosion kernel, W and H are the width and height of the effective detection area of the image, and is the maximum grayscale value and the minimum grayscale value of the image.

[0009] The above technical solution is adopted: Traditional corrosion filtering uses a fixed kernel size and cannot effectively remove large-area noise. A large kernel size is prone to mistakenly delete real defects. This solution can achieve a positive correlation between the kernel size and the effective area of the image. A larger kernel is required to cover the noise when adapting to large-size images.

[0010] Further preferably, the formula for calculating the feature weight in the CBAM attention mechanism of the improved YOLOv8 model is as follows: ; where is the attention mechanism for the c-th channel, is the Sigmoid activation function; C is the number of feature channels, and H and W are the height and width of the feature map; is the feature value of the C-th channel at the position .

[0011] Adopting the above technical solution: In the traditional YOLO model for TOPCon defect detection, the surface texture of the cell is similar to the real defect in the shallow feature map, resulting in the model misidentifying the texture as a defect; the channel attention mechanism above can assign weights to each channel, suppress irrelevant channels, such as background texture, and enhance defect-related channels, which can effectively identify defects.

[0012] Further preferably, the formula for calculating the correction coefficient of the dynamic threshold adjustment module is as follows: ; where is the threshold correction coefficient; is the actual diffusion temperature, is the set value of the reference diffusion temperature, is the phosphorus source concentration, is the reference concentration, is the annealing rate.

[0013] Adopting the above technical solution: The above solution can solve the problem that fluctuations in diffusion temperature, phosphorus source concentration, and annealing rate in the traditional TOPCon manufacturing process will significantly affect the defect morphology. Further preferably, the BiFPN network adopts a weighted feature fusion strategy, deletes single input nodes, fuses multi-scale features through a two-way path, and adaptively assigns fusion weights based on feature importance.

[0014] Further preferably, in the traditional algorithm module, the ErosionRectangle1 operator uses a 3×3 rectangular kernel to filter out noise with an area <0.1mm 2 and retain the effective defect contour.

[0015] A detection method is applied to a TOPCon cell defect detection device as described in any one of the above, and is characterized by including: ​​​​‌S1:‌ Load the process parameter configuration file for the day, start the optical imaging module, adjust the intensity of the three-color light source through the white balance calibrator so that the grayscale value deviation of each channel is ≤2%, verify the pulse synchronization between the area array camera and the conveyor, and control the positioning error within ±5μm; S2: The photoelectric sensor is used to trigger the area array camera to collect RGB three-channel images. The Halcon VarThreshold operator is used to perform dynamic threshold segmentation to generate the initial defect area. The corrosion kernel size is dynamically adjusted according to the image effective area size and grayscale distribution. The rectangular kernel corrosion filter is performed to remove the area <0.05mm 2 Noise area; ‌S3:‌ Input the preprocessed image into the improved YOLOv8 model, enhance the hidden crack features through the CBAM attention mechanism, fuse multi-scale features based on the BiFPN network, output the defect category, coordinates and confidence, and the hidden crack recall rate is ≥99.5%; S4: The diffusion temperature, phosphorus source concentration and annealing rate are obtained in real time through the MES system, and the dynamic threshold correction coefficient is calculated according to the nonlinear coupling relationship of the current process parameters. The hidden crack judgment condition is adjusted to confidence ≥ 0.95 × correction coefficient; ‌S5: Bind the defect data with WaferID and store it, generate a defect distribution heat map and a process parameter trend map, and when the same defect occurs in three consecutive wafers of the same batch, push the process optimization instructions to the MES system, update the SPC control chart and trigger the equipment maintenance alarm.

[0016] Preferably, the specific method of dynamically adjusting the size of the corrosion nucleus in S2 is: Based on the width and height of the effective detection area of the current image, the product value in millimeters is divided by 2000 to obtain the benchmark kernel size. The natural logarithm of the ratio of the maximum grayscale value to the minimum grayscale value of the image is taken, and the benchmark kernel size is multiplied by the natural logarithm and rounded to obtain the actual corrosion kernel pixel size used, where the minimum grayscale value is compensated by adding 1 to avoid division by zero errors.

[0017] Preferably, the dynamic threshold correction coefficient in S4 is calculated as follows: The temperature influence factor is obtained by dividing the difference between the actual diffusion temperature and the reference diffusion temperature by 50, multiplying it by the square root of the ratio of the actual value of the phosphorus source concentration to the reference concentration, and then multiplying it by an exponential function of the negative value of the annealing rate with the natural constant as the base divided by 10, and adding 1 to the calculated result as the overall correction coefficient.

[0018] Preferably, the defect distribution heat map in S5 is generated in the following manner: Convert the coordinate system on the surface of the cell to a polar coordinate system with the center as the origin, calculate the radial distance and angle of each defect point, perform spatial smoothing on the defect point density using a Gaussian kernel function, control the gradient sharpness of the heat map by adjusting the kernel bandwidth parameter, and finally map the density value to the HSV color space, where red represents the high-density defect area and green represents the low-density area. Description of the Drawings

[0019] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0020] Figure 1 It is a block diagram of the detection device module of TOPCon of the present application; Figure 2 It is a flowchart of the detection method of the present application; Figure 3 It is a device diagram of the present application.

[0021] 1. Outer cover; 2. Light source; 3. Camera; 4. Lens. Detailed Embodiments

[0022] In the following description, for the purpose of illustration rather than limitation, specific details such as specific system structures and technologies are presented to thoroughly understand the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present application.

[0023] It should be understood that when used in the specification and appended claims of the present application, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0024] Please refer to Figures 1-3 , for example, traditional TOPCon detection equipment has the following bottlenecks: ‌Lack of light source diversity‌: Only using a single-color light source, such as infrared or visible light, cannot capture multiple types of defects such as passivation layer defects, hidden cracks, and grid line abnormalities at the same time. It is necessary to replace the light source for detection multiple times, with low efficiency. Frequent replacement of the light source will generate a large amount of unnecessary time, affecting the detection efficiency; Algorithm limitations: Traditional image processing algorithms are sensitive to light, and the misdetection rate can reach 15%-20% when process parameters fluctuate; Static determination threshold: Defect determination relies on empirical thresholds and cannot adapt to the dynamic changes of process parameters such as diffusion temperature and phosphorus concentration, resulting in an increase in the missed detection rate; Data island phenomenon: The detection data is not interacted with the MES system in real time, and process optimization lags behind.

[0025] Based on the above problems, the present application provides a TOPCon cell defect detection device, including: Optical imaging module: It includes a 25-million-pixel camera, a three-color integrated light source array and a multi-spectral filter group. The light source array is composed of red, green and blue LEDs and supports pulse synchronous triggering; Traditional algorithm processing module: Integrate the Halcon image processing library, configure the dynamic threshold segmentation unit of the VarThreshold operator, the adjustable kernel size erosion filtering unit and the light and dark separation unit, and use the ErosionRectangle1 operator for erosion filtering; AI defect detection module: Built-in an improved YOLOv8 neural network model, the model structure includes a Backbone network embedded with a CBAM attention mechanism, a BiFPN network bidirectional feature pyramid and a dynamic loss function optimizer; Dynamic threshold adjustment module: Connect to the MES system through the OPC UA interface, obtain diffusion temperature, phosphorus source concentration and annealing rate parameters in real time, and build a non-linear coupling calculation unit to output a dynamic correction coefficient for the defect determination threshold; Data integration module: Integrate the MySQL database and the MQTT protocol communication unit, and configure a defect distribution heat map generator, a process parameter trend analyzer and an SPC control chart generator; Calibration module: It includes a conveying device driven by a high-precision linear motor, an optoelectronic positioning sensor and a white balance calibrator.

[0026] The above solution can solve the following technical problems through the integration of corresponding modules: Optical imaging module: Using three-color light sources red, green and blue, it can cover the wavelength bands of 620-680nm, 520-560nm, and 450-480nm. The single imaging time can be controlled within 1S, and it can synchronously capture the abnormal reflection of the passivation layer under the red light source, the scattering characteristics of hidden cracks under the green light source, and the deformation of the grid lines under the blue light source, increasing the detection efficiency by 275%; Algorithm cooperation mechanism: This solution realizes the cascade of the traditional algorithm VarThreshold dynamic segmentation and the AI model improved YOLOv8. First, filter out areas <0.05mm through erosion filtering2 noise, and then enhance the features of subtle defects through the CBAM attention mechanism, increasing the detection rate of hidden cracks from 89% to 99.5%; ‌Dynamic threshold adjustment‌: Obtain the process parameters of MES in real time through the OPC UA interface, such as diffusion temperature or annealing rate, and dynamically correct the determination threshold using the non-linear coupling formula, reducing the missed detection rate caused by temperature fluctuations to 0.8%; ‌Closed-loop feedback‌: The MQTT protocol pushes the defect distribution heat map and SPC control chart to MES. When the same type of defect appears in 3 consecutive wafers of the same batch, the process parameter adjustment instruction is automatically triggered, and the optimization response time is shortened to 20 minutes.

[0027] This solution can solve over-etching and under-etching in traditional technical solutions: small kernel sizes cannot effectively remove large-area noise, and large kernel sizes are prone to misdeleting real defects such as microcracks, resulting in a noise suppression rate ≤ 60%; And it is ‌light-sensitive‌. When the image gray scale distribution is uneven, such as bright in the center and dark at the edges, a fixed kernel size will amplify local noise, increasing the false detection rate by 12% - 18%.

[0028] The kernel size adjustment formula of the adjustable kernel size corrosion filtering unit is as follows: ; where K is the corrosion kernel size, W and H are the width and height of the effective detection area of the image, and are the maximum and minimum gray values of the image.

[0029] It is worth mentioning that the above solution can achieve dynamic optimization of the kernel size, and can convert the physical size into the kernel size reference value. For example, for a standard cell of 156mm × 156mm, the calculation is 156×156 / 2000 = 12.168, 156×156 / 2000 = 12.168, indicating that larger images require larger corrosion kernels to cover the noise.

[0030] It can achieve quantization of the light uniformity. For example, in a low-contrast scenario, set = 100, = 30, and finally ln(100 / 30) ≈ 1.145, which can achieve reducing the kernel size to prevent over-etching. In a high-contrast scenario, such as = 225, = 5, and finally ln(225 / 6) ≈ 3.784, which can achieve expanding the kernel size to enhance noise suppression.

[0031] It is worth mentioning that the above formula can achieve dynamic optimization of the kernel village and physical size adaptation. For example, it can make the kernel size positively correlated with the effective area (W×H) of the image. Calculated based on a 156mm×156mm standard cell, a reference value is obtained. To adapt to large-size images, a larger kernel is needed to cover the noise. By quantifying the gray-scale contrast, for a low-contrast image (such as Imax = 100, Imin = 30), ln(100 / 31) = 1.145 is calculated, and the kernel size is automatically reduced to prevent over-etching; for a high-contrast image (Imax = 255, Imin = 5), ln(255 / 6) = 3.784 is calculated, and the kernel size can be expanded to enhance filtering.

[0032] The traditional YOLO model has the following problems in TOPCon defect detection. The surface texture of the cell, such as the grid pattern, is similar to real defects, such as hidden cracks, in the shallow feature map, resulting in the model misidentifying the texture as a defect; the width of hidden cracks is usually 5 - 20μm, which only accounts for 3 - 10 pixels in the input image, and the recall rate of the standard YOLOv8 for such targets is ≤85%. The formula for calculating the feature weight in the CBAM attention mechanism of the improved YOLOv8 model is: ; Among them, is the attention mechanism for the c-th channel, is the Sigmoid activation function; C is the number of feature channels, and H and W are the height and width of the feature map; is the feature value of the C-th channel at position .

[0033] In the above formula, calculates the global average feature intensity of channel c, representing the overall importance of the channel; max(Fc) extracts the feature value of the most significant region within the channel, highlighting local key information, such as hidden cracks. Through multiplication and Sigmoid activation, higher weights are assigned to high-response channels, such as hidden crack features, for example, 0.9, and low-response channels such as background textures are suppressed, with a weight ≈0.1. It is worth mentioning that the above scheme assigns weights to each channel, suppresses irrelevant channels (such as background textures), and enhances defect-related channels. For example, in the detection of abnormal grid lines, the weight of the blue channel (450 - 480nm) is increased to 0.92, while the weight of the red channel is decreased to 0.15; in the BiFPN feature fusion stage, higher spatial weights are assigned to areas with a high incidence of hidden cracks (such as the edges of the cell), increasing the pixel-level recall rate of 5μm-level cracks from 82% to 97%; compared with the standard YOLOv8, the number of parameters of the improved model only increases by 3.8% (from 43.7M to 45.4M).

[0034] Further preferably, the correction coefficient calculation formula of the dynamic threshold adjustment module is: ; Wherein, is the threshold correction coefficient; is the actual diffusion temperature, is the reference diffusion temperature setting value, phosphorus source concentration, is the reference concentration, is the annealing rate.

[0035] Adopting the above technical solution: The above solution can solve the problem that in the traditional TOPCon manufacturing process, fluctuations in diffusion temperature, phosphorus source concentration, and annealing rate will significantly affect the defect morphology. For example, when exceeding the reference temperature , the thickness of the passivation layer increases, the surface reflectivity changes, resulting in an increase in the misjudgment rate of optical detection. When it decreases, the density of the phosphosilicate glass (PSG) layer decreases, and hidden cracks are more likely to occur. It is necessary to lower the judgment threshold to improve the detection rate; When it is too high, the annealing is insufficient, the crack edge is blurred, and the confidence of the AI model decreases.

[0036] In the above formula, the temperature compensation term , the term normalizes the temperature deviation. For example, = 850 °C, = 845 °C, the contribution value is +0.1; The square root is used to weaken the influence of concentration changes. When decreases from 1.0 mol / L to 0.8 mol / L, the contribution value of this term decreases from 1.0 to 0.894; The term applies a penalty factor of 0.223 to a high rate (such as v_a = 15 m / min) to lower the threshold to prevent missed detection; It is worth mentioning that within the process window with fluctuating ±10 °C, changing ±30%, in the range of 8 - 15 m / min, the threshold is adaptively adjusted to keep the missed detection rate stable at 0.08% - 0.12%.

[0037] The BiFPN network adopts a weighted feature fusion strategy, deletes single input nodes, fuses multi-scale features through bidirectional paths, and adaptively assigns fusion weights based on feature importance.

[0038] In the traditional algorithm module, the ErosionRectangle1 operator uses a 3×3 rectangular kernel to filter noise with an area <0.1 mm² and retain the effective defect contour.

[0039] A detection method, which is applied to a TOPCon cell defect detection device as described in any one of the above, and is characterized by including: S1: Load the process parameter configuration file of the day, start the optical imaging module, adjust the intensity of the three-color light source through a white balance calibrator so that the gray value deviation of each channel ≤ 2%, verify the pulse synchronization of the area array camera and the conveying device, and control the positioning error within ±5μm; S2: Trigger the area array camera to collect RGB three-channel images through a photoelectric sensor, use the VarThreshold operator of Halcon for dynamic threshold segmentation to generate an initial defect area, dynamically adjust the erosion kernel size according to the size and gray distribution of the effective area of the image, perform rectangular kernel erosion filtering, and eliminate noise areas with an area < 0.05mm²; S3: Input the preprocessed image into the improved YOLOv8 model, enhance the hidden crack features through the CBAM attention mechanism, fuse multi-scale features based on the BiFPN network, output the defect category, coordinates and confidence level, and the recall rate of hidden cracks ≥ 99.5%; S4: Obtain the diffusion temperature, phosphorus source concentration and annealing rate in real time through the MES system, calculate the dynamic threshold correction coefficient according to the non-linear coupling relationship of the current process parameters, and adjust the hidden crack determination condition to confidence level ≥ 0.95 × correction coefficient; S5: Bind and store the defect data with the WaferID, generate a defect distribution heat map and a process parameter trend map, and push a process optimization instruction to the MES system when the same type of defect appears in 3 consecutive wafers of the same batch, update the SPC control chart and trigger an equipment maintenance alarm.

[0040] The specific method for dynamically adjusting the erosion kernel size in S2 is: Based on the product value of the width and height of the current effective detection area of the image in millimeters divided by 2000 to obtain the reference kernel size, take the natural logarithm of the ratio of the maximum gray value to the minimum gray value of the image, multiply the reference kernel size by the natural logarithm and then round up to obtain the actual pixel size of the erosion kernel used, where the minimum gray value is compensated by adding 1 to avoid division by zero error.

[0041] The calculation method of the dynamic threshold correction coefficient in S4 is: Divide the difference between the actual diffusion temperature and the reference diffusion temperature by 50 to obtain the temperature influence factor, multiply by the square root of the ratio of the actual phosphorus source concentration to the reference concentration, and then multiply by the exponential function of the negative annealing rate divided by 10 with the natural constant as the base, and take the result of the calculation plus 1 as the overall correction coefficient.

[0042] The generation method of the defect distribution heat map in S5 is: Convert the coordinate system on the surface of the cell to a polar coordinate system with the center as the origin, calculate the radial distance and angle of each defect point, perform spatial smoothing on the defect point density using a Gaussian kernel function, control the gradient sharpness of the heat map by adjusting the kernel bandwidth parameter, and finally map the density value to the HSV color space, where red represents the high-density defect area and green represents the low-density area.

[0043] Example 1: Implementation of TOPCon detection device Hardware configuration: The optical imaging system uses a 25-million-pixel area array camera combined with a three-color integrating light source, which can achieve a detection CT ≤ 0.8 s / cell and match the speed of the detector. The lens equipped with the camera can be selected according to actual needs to ensure good imaging effects. The light source module integrates a specially designed three-color integrating light source, which can provide uniform and stable illumination to meet the detection requirements. The mechanical transmission device uses a high-precision drive system, such as an SMC linear motor drive, with a speed accuracy of up to ±0.01 m / s, and cooperates with a vacuum chuck to achieve precise positioning of the silicon wafer, ensuring the stability of the detection process.

[0044] Example 2: Algorithm processing flow S1: Traditional algorithm module processing: Through Halcon morphological processing, first use the VarThreshold operator for positioning, which can quickly and accurately determine the area where defects may exist. Then use the ErosionRectangle1 operator for erosion filtering, and use a 3×3 rectangular kernel to filter out noise with an area <0.1 mm 2 to retain the effective defect contour. Then use the Tuple_Select operator for light and dark separation to further screen the candidate defect area and effectively reduce the missed detection rate.

[0045] S2: AI defect detection module inference: Input the preprocessed image into the improved YOLOv8 model. This model uses a Bifpn bidirectional pyramid network and introduces a CBAM attention mechanism, which greatly improves the learning ability for complex structural features such as hidden cracks. The model is deployed on a high-performance industrial control computer to ensure the inference speed and efficiency. Output the category and accurate coordinates of the defect.

[0046] S3: Dynamic threshold adjustment module: According to the process parameters fed back by MES, the dynamic threshold adjustment module automatically optimizes the defect determination threshold. For example, obtain the current batch sintering temperature in MES through the OPC UA interface. When the temperature deviation > 3°C, automatically adjust the determination threshold for defects such as hidden cracks to adapt to process fluctuations and ensure the accuracy of detection.

[0047] Example 3 Collaborative production line operation: Defect data and WaferID are bound through the MQTT protocol and stored in the MySQL database, facilitating subsequent data management and query. The data integration module establishes a real-time data feedback mechanism. This module can receive data from various sensors in real-time and transmit this data to the central server for processing through a high-speed communication interface. The MES system generates an SPC control chart (Cpk ≥ 1.33) based on this data. When specific situations such as concentric circle defects occur in 5 consecutive wafers, adjustment instructions for process parameters such as laser etching power are triggered. At the same time, the subsequent process parameter adjustment and optimization process is guided according to the detailed information and trend diagrams in the generated quality report.

[0048] Example 4 The detection method implementation starts the device and loads the process parameter configuration file for the day to ensure that the device operates according to the established process requirements. Execute the self-check program, calibrate the camera white balance with an error < 2%, verify the light source pulse synchronization to ensure imaging quality and detection accuracy. Start the production line conveyor belt. When the photoelectric sensor detects that the silicon wafer is in place, image acquisition is triggered.

[0049] It is worth mentioning that this patent introduces the attention mechanism structure of CBAM. First, the concept of the channel dimension is adopted, and a maximization pool is constructed to enhance the information expression ability of the image at different scales. At the same time, to further improve the generalization ability of the model, an average pool strategy is also introduced to ensure the effective extraction and fusion of multi-scale context information. Then, using the Concat operation, the pooled feature maps are concatenated and convolutional processing is performed on them to generate spatial attention weights. Finally, the output of the neural network is processed through the Sigmoid activation function for non-linear mapping. In this way, the model output can be effectively converted from a larger value to a smaller value, resulting in a normalized spatial attention map. The spatial attention mechanism weights the features of each spatial position in the photovoltaic cell defects, thereby highlighting the image regions of interest and reducing the influence of interference regions.

[0050] The BiFPN adopted in this patent optimizes the feature fusion path, deletes the nodes with only a single input to reduce redundancy, and adopts multiple repeated bidirectional paths to enable deep feature fusion and information transfer; compared with the traditional FPN structure, BiFPN introduces a bidirectional information flow mechanism between adjacent levels of the feature pyramid. This bidirectional interaction mechanism enables the information at different levels of the feature pyramid to be fully fused in two directions, thereby enhancing the richness and diversity of feature expression. In the evolution of the feature pyramid architecture, due to the adoption of only the top-down unidirectional propagation mechanism, the traditional FPN structure has the problem of insufficient fusion of high-level semantic information and low-level detailed features. To address this limitation, PANet realizes bidirectional cross-level feature interaction by constructing a reverse feature propagation path, significantly improving the semantic expression ability of feature maps at different scales. NAS-FPN, on the other hand, uses automated neural network search technology to construct an optimal feature fusion topology, but its complex search mechanism leads to a relatively low model convergence efficiency.

[0051] This patent uses Halcon for morphological analysis. First, based on features such as the gray value of the image, the Var_Threshold operator is used to accurately locate the target area and delineate the approximate range where defects may exist. Subsequently, through the ErosionRectangle1 operator, according to the pre-set rectangular structure element parameters, an erosion filtering operation is performed on the located image to remove noise interference and fine miscellaneous points in the image, making the image features more clearly prominent. Then, the Tuple_Select operator is used to separate light and dark according to the difference standard of pixel values, and the areas with different brightness features are accurately divided. According to this complete process, candidate defect areas can be quickly and comprehensively screened out, greatly improving the detection efficiency, effectively reducing the missed detection rate, and ensuring the accuracy and reliability of product quality detection.

[0052] In the above embodiments, the device components involved are all conventional device components unless otherwise specified, and the connection methods and control methods involved are all conventional connection methods and control methods unless otherwise specified.

[0053] The above has described the present invention in detail in combination with the embodiments. However, those skilled in the art can understand that without departing from the purpose of the present invention, various specific parameters in the above embodiments can be changed to form multiple specific embodiments, which are all within the common change range of the present invention and will not be elaborated here one by one.

Claims

1. A TOPCon solar cell defect detection device, characterized in that, include: Optical imaging module: includes a 25-megapixel camera, a three-color integral light source array, and a multi-spectral filter set. The light source array consists of red, green, and blue LEDs and supports pulse synchronization triggering. ‌Traditional algorithm processing module‌: Integrates the Halcon image processing library, configures the VarThreshold operator dynamic threshold segmentation unit, the adjustable kernel size erosion filter unit and the light and dark separation unit, and uses the ErosionRectangle1 operator for erosion filtering; ‌AI Defect Detection Module‌: Built-in improved YOLOv8 neural network model, the model structure includes Backbone network embedded with CBAM attention mechanism, BiFPN network bidirectional feature pyramid and dynamic loss function optimizer; Dynamic Threshold Adjustment Module: Connects to the MES system via the OPC UA interface to obtain diffusion temperature, phosphorus source concentration and annealing rate parameters in real time. It has a built-in nonlinear coupling calculation unit and outputs a dynamic correction coefficient for the defect judgment threshold. ‌Data Integration Module‌: Integrates MySQL database and MQTT protocol communication unit, configures defect distribution heat map generator, process parameter trend analyzer and SPC control chart generator; ‌Calibration module‌: includes a high-precision linear motor driven conveyor, a photoelectric positioning sensor and a white balance calibrator.

2. The TOPCon cell defect detection device according to claim 1, characterized in that, The kernel size adjustment formula of the adjustable kernel size erosion filter unit is: ； Wherein, K is the size of the corrosion core, W and H are the width and height of the effective detection area of the image, and are the maximum gray value and the minimum gray value of the image.

3. The TOPCon cell defect detection device according to claim 1, wherein, The feature weight calculation formula in the CBAM attention mechanism of the improved YOLOv8 model is: ； in, is the c-th channel attention mechanism, is the Sigmoid activation function; C is the number of feature channels, H and W are the height and width of the feature maps; For the Cth channel at position The characteristic value at .

4. The TOPCon cell defect detection device according to claim 1, wherein, The correction coefficient calculation formula of the dynamic threshold adjustment module is: ； Among them, is the threshold correction coefficient; is the actual diffusion temperature, is the set value of the reference diffusion temperature, phosphorus source concentration, is the reference concentration, is the annealing rate.

5. The defect detection device for TOPCon solar cells according to claim 1, wherein, The BiFPN network adopts a weighted feature fusion strategy, deletes a single input node, fuses multi-scale features through a bidirectional path, and adaptively allocates fusion weights based on feature importance.

6. The TOPCon cell defect detection device according to claim 1, characterized in that, In the traditional algorithm module, the ErosionRectangle1 operator uses a 3×3 rectangular kernel to filter noise with an area <0.1mm² and retain the effective defect contour.

7. A detection method, applied to a TOPCon cell defect detection device according to any one of claims 1-6, characterized in that, include: ‌S1:‌ Load the process parameter configuration file for the day, start the optical imaging module, adjust the intensity of the three-color light source through the white balance calibrator so that the grayscale value deviation of each channel is ≤2%, verify the pulse synchronization between the area array camera and the conveyor, and control the positioning error within ±5μm; ‌S2:‌ The photoelectric sensor triggers the area array camera to collect RGB three-channel images, and the Halcon VarThreshold operator is used to perform dynamic threshold segmentation to generate the initial defect area. The corrosion kernel size is dynamically adjusted according to the image effective area size and grayscale distribution, and rectangular kernel corrosion filtering is performed to remove noise areas with an area of <0.05mm²; ‌S3:‌ Input the preprocessed image into the improved YOLOv8 model, enhance the hidden crack features through the CBAM attention mechanism, fuse multi-scale features based on the BiFPN network, output the defect category, coordinates and confidence, and the hidden crack recall rate is ≥99.5%; S4: Obtain the diffusion temperature, phosphorus source concentration, and annealing rate in real time through the MES system. Calculate the dynamic threshold correction coefficient according to the non-linear coupling relationship of the current process parameters, and adjust the hidden crack determination condition to confidence level ≥ 0.95 × correction coefficient; S5: Bind and store the defect data with the WaferID, generate a heat map of defect distribution and a trend graph of process parameters. When the same type of defect appears in 3 consecutive wafers of the same batch, push a process optimization instruction to the MES system, update the SPC control chart, and trigger an equipment maintenance alarm.

8. A detection method according to claim 7, wherein The specific method for dynamically adjusting the corrosion kernel size in S2 is as follows: Based on the product value of the width and height of the effective detection area of the current image in millimeters divided by 2000 to obtain the reference kernel size, take the natural logarithm of the ratio of the maximum gray value to the minimum gray value of the image, multiply the reference kernel size by the natural logarithm and then round to obtain the pixel size of the corrosion kernel actually used, where the minimum gray value is compensated by adding 1 to avoid division by zero error.

9. A detection method according to claim 7, characterized in that, The calculation method of the dynamic threshold correction coefficient in S4 is as follows: Divide the difference between the actual diffusion temperature and the reference diffusion temperature by 50 to obtain the temperature influence factor, multiply it by the square root of the ratio of the actual value of the phosphorus source concentration to the reference concentration, and then multiply it by the exponential function of the negative annealing rate divided by 10 with the natural constant as the base. Add 1 to the calculation result as the overall correction coefficient.

10. A detection method according to claim 7, characterized in that, The generation method of the heat map of defect distribution in S5 is as follows: Convert the coordinate system on the surface of the cell to a polar coordinate system with the center as the origin, calculate the radial distance and angle of each defect point, use a Gaussian kernel function to perform spatial smoothing on the defect point density, control the gradient sharpness of the heat map by adjusting the kernel function bandwidth parameter, and finally map the density value to the HSV color space, where red represents the high-density defect area and green represents the low-density area.

Images