HJT double-half-piece detection device and detection method

Abstract

The invention relates to the technical field of photovoltaic manufacturing, in particular to an HJT double-half-piece detection device and method. The high-speed optical imaging module is composed of a double-station 20 million-pixel camera, a double-cavity three-color integral light source and a vacuum adsorption conveying belt, a light source system supports red, green and blue three-channel independent control and 10kHz stroboflash, and the vibration suppression error is smaller than + / -5 microns; the AI vision algorithm fusion module is used for integrating a Halcon morphological processing unit and an improved YOLOv8 model, separating double-half-piece edge features through a Difference operator, extracting an overlapping region through an Intersection operator, and correcting size measurement based on a dynamic compensation coefficient; the YOLOv8 model is improved, a coordinate attention mechanism is embedded in a backbone network, and a 160 * 160 detection layer is newly added; according to the system, the high-speed and intelligent requirements of an HJT production line can be met while the detection precision is ensured.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic manufacturing technology, and in particular to a detection device and a detection method for HJT double half-cells. Background Art

[0002] As the global energy structure transition accelerates, HJT solar cells, with their high conversion efficiency, low temperature coefficient, and bifacial power generation characteristics, have become a core development direction for the photovoltaic industry. In HJT double-half-cell manufacturing, due to the use of ultra-thin silicon wafers and a busbar-less design, surface defects such as cracks, chipping, and color spots have a significant impact on photovoltaic performance, placing higher demands on detection technology. Traditional inspection equipment faces the following technical bottlenecks: edge coupling interference: During parallel inspection of double-half-cells, the edge features of adjacent wafers interfere with each other, resulting in significant offset errors in traditional positioning algorithms; missed detection of minor defects: Existing models are insufficiently capable of identifying micro-defects such as cracks and chipping, resulting in a high missed detection rate; and dimensional measurement deviation: Fixed ROI measurement methods are affected by environmental factors, resulting in unstable diagonal measurement accuracy.

[0003] While existing technologies use multi-station sorting to improve accuracy, this results in a longer cycle time (CT) of 1.2 seconds per two wafers, making it unsuitable for HJT production lines with speeds exceeding 8,000 wafers per hour. Deep learning algorithms have been introduced, but this hasn't solved the problem of double-half-wafer feature coupling. While using high-resolution line scan cameras improves accuracy, it significantly increases equipment cost and maintenance complexity.

[0004] Based on the above problems, there is an urgent need to develop a detection system that integrates high-precision optical imaging, adaptive morphological processing and improved deep learning algorithms to meet the high-speed and intelligent requirements of HJT production lines while ensuring detection accuracy, and breaking through the quality control bottleneck of photovoltaic cell manufacturing. Summary of the Invention

[0005] In view of the fact that the existing technology uses a multi-station inspection scheme to detect HJT double half-wafers to improve accuracy, which significantly increases equipment cost and maintenance complexity, the present application provides a HIT double half-wafer inspection device and inspection method to solve the above problems.

[0006] A HJT double half-chip detection device, comprising: High-speed optical imaging module: Consists of a dual-station 20-megapixel camera, a dual-cavity three-color integral light source, and a vacuum adsorption conveyor. The camera resolution is set to 0.03mm / pixel, and the light source system supports independent control of the red, green, and blue channels and a 10kHz strobe frequency, with a vibration suppression error of <±5μm. AI vision algorithm fusion module: Integrates the Halcon morphological processing unit with the improved YOLOv8 model. It uses the Difference operator to separate edge features of two half-slices, the Intersection operator to extract overlapping areas, and corrects size measurements based on dynamic compensation coefficients. The improved YOLOv8 model embeds a coordinate attention mechanism in the backbone network and adds a 160×160 detection layer. Production line collaborative control module: The WaferID is bound to the dual core of RFID and visual features, and the laser repair power grading adjustment instruction is triggered based on the defect distribution heat map. The edge computing node generates the process capability index control chart in real time and pushes it to the MES system. Dynamic Environmental Compensation Module: Automatically adjusts ROI boundaries and defect judgment thresholds based on cutting offset feedback from the MES and real-time environmental parameters, and is compatible with the deformation differences between 182mm and 230mm battery specifications.

[0007] The above technical solution uses a dynamic compensation coefficient α to achieve dual-factor temperature and humidity correction. After compensation, the battery edge positioning error is reduced from 0.15mm in the traditional solution to <0.05mm, meeting the micro-crack detection accuracy requirements of the HJT production line.

[0008] Preferably, the calculation formula of the dynamic compensation coefficient is: , in, is the dynamic compensation coefficient, is the thermal expansion coefficient of the material, It is the difference between the real-time ambient temperature and the standard temperature of 25℃. is the temperature correction reference value, taking 100℃, The diagonal length calculated by the Halcon sub-pixel measurement tool, is the nominal value of the battery cell specification; is the humidity influencing factor, is the real-time ambient humidity, The standard humidity is 60%.

[0009] Adopt the above technical solution: This solution can solve the serious limitations of traditional compensation algorithms by introducing The fractional structure describes the saturation characteristics of the temperature effect, which can prevent over-correction and square root design. Able to achieve compatible CTE difference of 182mm / 230mm battery (2.3×10 -6 / ℃ vs. 3.1×10 -6 / ℃), which reduces the cross-specification detection deviation and enables the process linkage of cutting speed and temperature.

[0010] Further preferably, the weight calculation of the coordinate attention mechanism satisfies: , in, is the attention weight of channel c at position (x, y), and They are horizontal and vertical feature encoding respectively; and is the spatial attenuation coefficient, is the Sigmoid activation function‌.

[0011] Adopting the above technical solution: This solution can solve the problem of spatial information blurring caused by traditional convolutional neural networks and the problem that the existing attention mechanism cannot distinguish between interference such as battery surface reflection and dust and the overlapping grayscale distribution of real defects in RGB images.

[0012] Further preferably, the dynamic adjustment formula of the defect judgment threshold is: , in, is the defect judgment threshold after dynamic adjustment, is the baseline defect threshold, is the humidity influencing factor, Indicates the real-time ambient humidity. Indicates standard humidity, Indicates the cutting speed deviation value, Indicates standard cutting speed.

[0013] Adopting the above technical solution: The above solution can solve the problem that a fixed threshold value cannot distinguish between real defects and noise. This solution can compensate for the nonlinear effect of humidity changes on imaging contrast and adjust the threshold sensitivity according to the cutting speed deviation.

[0014] Further preferably, the Halcon morphological processing includes: using a 5×5 circular kernel closing operation to filter out pseudo defects with an area less than 0.01 mm²; dynamically correcting the diagonal measurement benchmark based on the temperature compensation coefficient α; fusing multi-threshold segmentation results through the Union2 operator, and performing priority re-inspection according to the defect risk level.

[0015] Further preferably, the improved YOLOv8 model includes: adding a 160×160 detection layer after the 18th layer of the backbone network, fusing shallow high-resolution features, and embedding a coordinate attention mechanism in the backbone network.

[0016] A detection method, applied to an HJT double half-wafer detection device as described in any one of the above, is characterized by comprising: S1: Use a dual-station camera to capture images of two half-cell batteries, combined with a three-color integral light source to suppress reflection interference; S2: Use Halcon's Difference operator to separate the edge features of adjacent slices and extract candidate defect areas; S3: Use the improved YOLOv8 model to classify and detect defects, and enhance edge defect response through the CA module; S4: Dynamically compensate for dimensional measurement deviations based on environmental parameters and combine SPC analysis to monitor process capability index in real time; S5: Bind detection data with RFID and visual features to trigger closed-loop process optimization instructions.

[0017] Further preferably, the S2 further includes: calling the Intersection operator to extract the overlapping portion of the candidate area and the standard template, identifying edge collapse and corner chipping defects; calculating the diagonal length through a sub-pixel measurement tool, and correcting the thermal expansion error in combination with the temperature compensation coefficient α.

[0018] Further preferably, the inference time of the improved YOLOv8 model described in step S3 is ≤0.1 second / frame, supporting the deployment of the TensorRT acceleration engine.

[0019] Further preferably, in step S5, when 5 pieces of the same type of defects are detected consecutively, the laser repair power and the robot gripping posture are automatically adjusted in stages. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 This is the system block diagram of the HJT double half-chip detection device for this application; Figure 2 This is a flow chart of the detection method for this application; Figure 3 This is the equipment diagram for this application.

[0022] In the picture: 1. Outer cover; 2. Light source; 3. Camera; 4. Lens. DETAILED DESCRIPTION

[0023] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may 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 obscuring the description of the present application with unnecessary detail.

[0024] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0025] See also Figure 1-Figure 3 For example, although the existing technology uses a multi-station sorting solution to improve accuracy, it results in a test cycle (CT) of 1.2 seconds per double wafer, which cannot meet the requirements of HJT production lines with a capacity of more than 8,000 wafers per hour. Although deep learning algorithms have been introduced, the problem of double half-wafer feature coupling has not been solved. To address the above issues, this application provides an HJT double half-wafer inspection device, including: High-speed optical imaging module: Consists of a dual-station 20-megapixel camera, a dual-cavity three-color integral light source, and a vacuum adsorption conveyor. The camera resolution is set to 0.03mm / pixel, and the light source system supports independent control of the red, green, and blue channels and a 10kHz strobe frequency, with a vibration suppression error of <±5μm. AI vision algorithm fusion module: Integrates the Halcon morphological processing unit with the improved YOLOv8 model. It uses the Difference operator to separate edge features of two half-slices, the Intersection operator to extract overlapping areas, and corrects size measurements based on dynamic compensation coefficients. The improved YOLOv8 model embeds a coordinate attention mechanism in the backbone network and adds a 160×160 detection layer. Production line collaborative control module: The WaferID is bound to the dual core of RFID and visual features, and the laser repair power grading adjustment instruction is triggered based on the defect distribution heat map. The edge computing node generates the process capability index control chart in real time and pushes it to the MES system. Dynamic Environmental Compensation Module: Automatically adjusts ROI boundaries and defect judgment thresholds based on cutting offset feedback from the MES and real-time environmental parameters, and is compatible with the deformation differences between 182mm and 230mm battery specifications.

[0026] It is worth mentioning that this solution introduces a dynamic compensation coefficient α to achieve dual-factor coupled correction of temperature and humidity. After compensation, the battery edge positioning error is reduced from 0.15mm in the traditional solution to <0.05mm, meeting the accuracy requirements of HJT production lines for microcrack detection.

[0027] This patent optimizes the Halcon morphological processing flow and adopts a multi-operator cascade strategy to improve the detection accuracy of double half-cells: the Difference operator is used to separate the edge features of the double half-cells, and the coupling interference of adjacent cells is eliminated based on the dual-station image difference calculation, which significantly reduces the positioning deviation caused by the coupling interference of adjacent cells, and dynamically generates an adaptive ROI boundary to be compatible with the deformation differences of batteries of different specifications; based on the coordinates of the positioning feature points, the Math.Sqrt operator is called to calculate the diagonal length, and the influence of thermal expansion is dynamically compensated in combination with real-time environmental parameters; the Intersection operator is used to accurately extract the overlapping part of the candidate area and the standard template, effectively identify damage defects such as edge collapse and missing corners, suppress the interference of tiny pseudo defects, and improve the reliability of damage detection; the Union2 operator is performed on the multi-threshold segmentation results, and priority re-inspection is implemented based on the defect risk level to optimize detection efficiency and accuracy.

[0028] The improved YOLOv8 network structure in this patent is mainly composed of three parts. The backbone network (Backbone), the neck (Neck) and the detection head (Head). In order to enhance the feature extraction capability, the C2F structure is adopted in the backbone network and the neck part, replacing the C3 structure in YOLOv5, obtaining more gradient data, reducing the computational complexity, and greatly improving the learning ability of the convolutional neural network; in the head part, the mainstream decoupling head structure is adopted, so that the target position and category information can be separated. In terms of the loss function calculation, the Task-Aligned Assigner is responsible for the effective distribution of positive and negative samples, and is combined with the Distribution Focal Loss to further improve the model performance. In the data enhancement part, drawing on the YOLOX method, Mosaic enhancement is turned off in the last 10 epochs, thereby improving the accuracy of the model.

[0029] This patent introduces the CA coordinate attention mechanism into the YOLOv8 backbone network. This design significantly improves the model's ability to capture the sensitivity of target positions by fusing dual information of channels and spatial positions, effectively enhancing the recognition accuracy of tiny defects in complex backgrounds. The CA module achieves feature enhancement through two key stages: coordinate information embedding and attention generation. In the coordinate information embedding stage, the input features are decomposed horizontally and vertically through decomposition convolution operations, and bidirectional pooling encoding is performed separately to generate channel feature vectors containing direction perception, accurately capturing the spatial long-range dependencies across channels. In the attention generation stage, the bidirectional encoding features are fused and nonlinear transformations are performed using 1×1 convolutions, combined with the Sigmoid function to generate an attention map with spatial position weights. The CA mechanism explicitly embeds direction perception and position information in the channel attention through a coordinate dimension feature decoupling strategy, enabling the network to adaptively focus on the spatial distribution characteristics of key areas. The module adopts a lightweight structural design and integrates channel and spatial information through a cross-dimensional interaction mechanism. Without significantly increasing the amount of computation, it achieves collaborative optimization of target positioning and feature expression, effectively improving the detection rate and positioning accuracy of minor defects in complex industrial scenarios, while suppressing background noise interference.

[0030] The calculation formula of the dynamic compensation coefficient is: , in, is the dynamic compensation coefficient, is the thermal expansion coefficient of the material, It is the difference between the real-time ambient temperature and the standard temperature of 25℃. is the temperature correction reference value, taking 100℃, The diagonal length calculated by the Halcon sub-pixel measurement tool, is the nominal value of the battery cell specification; is the humidity influencing factor, is the real-time ambient humidity, The standard humidity is 60%.

[0031] It is worth mentioning that this solution is an optimization of the traditional compensation model using the linear temperature correction formula. The correction formula of the traditional compensation model is: However, when ΔT>30℃, HJT battery will expand nonlinearly due to the material (CTE=2.3×10 -6 / °C) and humidity absorption effects lead to compensation errors exceeding 0.2mm. Furthermore, manually calibrated parameters cannot adapt to the CTE differences between batteries of different specifications (182mm / 230mm). To address this issue, this embodiment further optimizes the correction formula of the compensation model based on consideration of various factors.

[0032] In the above formula, the terms are The fractional structure describes the saturation characteristics of the temperature effect. When ΔT approaches an extreme value, such as =80℃, the growth rate of this term slows down, avoiding the over-compensation problem of the traditional linear model. =50℃, when =50℃ when the value is 2, and =100℃, the value is 3 and the actual compensation amount is limited.

[0033] The parameter β was optimized in the range of 0.8-1.2 through orthogonal experiments to cover the CTE fluctuation (±5%) of different battery batches. Experiments show that when β=1.0, the compensation error for 230mm batteries is minimized (±0.03mm), while β=0.9 is suitable for 182mm batteries.

[0034] The square root of the above formula The design balances the deformation differences between batteries of different sizes. For example, the nominal length ratio of 182mm and 230mm batteries is 0.791, but the actual deformation exhibits a nonlinear relationship with length (due to edge stress concentration). The square root method adjusts this ratio to 0.889, which better matches the measured deformation data (R²=0.96).

[0035] Humidity item In the example, λ is dynamically bound to the cutting speed V. When V increases from 1.0m / s to 1.5m / s, λ increases from 0.1 to 0.2, and the humidity compensation sensitivity increases by 100%. This design is due to the fact that the airflow on the battery surface during high-speed cutting accelerates the evaporation of water, requiring a stronger humidity compensation weight. Specific embodiments

[0036] Under the working conditions of ΔT=50℃, H=80%RH, the edge positioning error after compensation is ≤0.05mm (the traditional solution is 0.2mm); When switching across specifications (182mm 230mm), the system automatically identifies the actual value of L and adjusts the square root term. The parameter adaptation time is ≤ 10 seconds, and the changeover efficiency is improved by 40%.

[0037] Adopt the above technical solution: This solution can solve the serious limitations of traditional compensation algorithms by introducing The fractional structure describes the saturation characteristics of the temperature effect, which can prevent over-correction and square root design. Able to achieve compatible CTE difference of 182mm / 230mm battery (2.3×10 -6 / ℃ vs. 3.1×10 -6 / ℃), which reduces the cross-specification detection deviation and enables the process linkage of cutting speed and temperature.

[0038] Further preferably, the weight calculation of the coordinate attention mechanism satisfies: , in, is the attention weight of channel c at position (x, y), and They are horizontal and vertical feature encoding respectively; and is the spatial attenuation coefficient, is the Sigmoid activation function‌.

[0039] For example, traditional attention mechanisms (such as the SE module) have a low response weight (average 0.31) for microcracks (0.03 mm) and are susceptible to interference from reflective noise (noise activation value reaches 0.75). Furthermore, pooling operations lead to the loss of sub-pixel features, resulting in a missed detection rate exceeding 8%.

[0040] The exponential decay enhancement mechanism is introduced into the formula to learn parameters and Control the horizontal and vertical attention attenuation gradients. For example, for horizontal microcracks, the model can automatically learn = 0.1 (weak attenuation) and = 0.5 (strong attenuation), which increases the weight of the crack area to 0.92, while the weight of the background noise decreases to 0.0.09.

[0041] The exponential function in the above formula The introduction of causes the weights of pixels far from the target area to decrease exponentially, effectively suppressing edge reflections. Experiments show that this mechanism improves noise suppression by three times over traditional Gaussian filtering (PSNR increases from 28dB to 35dB).

[0042] This formula also introduces a Gaussian prior constraint, overlaying a Gaussian distribution with a standard deviation of σ = 5 on the 160×160 feature map and performing a Hadamard product with the coordinate attention weight. This embeds prior knowledge (cracks are typically located within 5mm of the battery edge) into the model, increasing the attention weight of the edge region by an additional 20%.

[0043] In specific experiments, in the strong light interference test, the CA mechanism increased the microcrack detection rate from 82% to 98.7%, and reduced the false detection rate from 5.2% to 0.3%; The model inference time only increased by 1.2ms (total time 15.3ms), meeting the production line's 60fps real-time requirement.

[0044] Adopting the above technical solution: This solution can solve the problem of spatial information blurring caused by traditional convolutional neural networks and the problem that the existing attention mechanism cannot distinguish between interference such as battery surface reflection and dust and the overlapping grayscale distribution of real defects in RGB images.

[0045] Further preferably, the dynamic adjustment formula of the defect judgment threshold is: , in, is the defect judgment threshold after dynamic adjustment, is the baseline defect threshold, is the humidity influencing factor, Indicates the real-time ambient humidity. Indicates standard humidity, Indicates the cutting speed deviation value, Indicates standard cutting speed.

[0046] The traditional technical solution uses a fixed threshold that cannot eliminate the following interferences: humidity changes lead to surface reflection differences and cutting speed fluctuations cause edge deformation, which requires adjustment of the crack detection sensitivity. The value of is an initial threshold value set based on historical data or statistical methods, for example, The principle is to determine the benchmark value of crack length or grayscale difference.

[0047] In the above formula, the humidity adjustment term is By relative humidity change rate Quantify environmental interference, its design features include nonlinear response, when When the temperature deviation is positive, the threshold θ is increased proportionally to reduce the false detection caused by enhanced reflection; when When , the threshold is lowered to enhance sensitivity.

[0048] The coefficient in the above formula The calibration of The physical meaning of is the threshold change rate caused by unit humidity deviation. For example, if γ=0.05, when H=60%RH ( =40%RH), the humidity adjustment item is: , That is, the threshold Improved by 2.5% to balance the interference of reflections. Experiments show that Need to be optimized through orthogonal experiments. =20%-80%RH range, =0,03-0.08, the error detection rate can be achieved 1.5%.

[0049] Speed ​​adjustment By quantifying velocity fluctuations ( ) and the tolerance threshold ( ) ratio, dynamically adjust the threshold: when >0 (actual speed is higher than the set value), the threshold Boost to suppress glitch noise; when When <0 (insufficient speed), the threshold is lowered to capture potential defects.

[0050] It is directly related to the process capability of the equipment. For example, if the speed fluctuation allowed by the equipment is ±0.2m / s ( =0.2m / s), when ΔV=0.1m / s, the adjustment item is: , That is, the threshold θ is increased by 50%, effectively suppressing the vibration noise caused by high-speed cutting.

[0051] It should be mentioned that the humidity and speed adjustment items are superimposed in the form of a product during the rise, reflecting the synergistic effect of multiple parameters. When a certain parameter fluctuates violently, the product form can avoid the adjustment amount being excessively amplified by a single factor, and different production lines can be adjusted independently. and Adapt to diverse needs. For example, high-precision production lines ( =0.1m / s) can be set to a smaller value To enhance response sensitivity.

[0052] It is worth mentioning that the above scheme can solve the problem that a fixed threshold cannot distinguish between real defects and noise. This scheme can compensate for the nonlinear effect of humidity changes on imaging contrast and adjust the threshold sensitivity according to the cutting speed deviation.

[0053] The Halcon morphological processing includes: using a 5×5 circular kernel closing operation to filter out pseudo defects with an area less than 0.01 mm²; dynamically correcting the diagonal measurement benchmark based on the temperature compensation coefficient α; fusing multi-threshold segmentation results through the Union2 operator, and implementing priority re-inspection according to the defect risk level.

[0054] The improved YOLOv8 model includes: adding a 160×160 detection layer after the 18th layer of the backbone network, fusing shallow high-resolution features, and embedding a coordinate attention mechanism in the backbone network.

[0055] In this patent, adding a small target detection layer refers to upsampling the 18th layer of the backbone network, thereby expanding the size of the feature map. Subsequently, the feature map obtained by upsampling the 19th layer is fused with the feature map of the 2nd layer in the backbone network. This design ultimately forms a detection layer of size 160×160 when the input image size is 640×640. Compared with highly downsampled layers, this layer has a lower downsampling multiple and can retain a large amount of small target feature information, significantly improving the network model's sensitivity and detection capabilities for small target defects. Through this design, the network model can more effectively identify and process small defect targets of size 4×4, such as broken grids, virtual prints, and smaller node defects. The added small target detection layer not only meets the diverse needs for defect size recognition, but also significantly improves the accuracy and reliability of the model in defect detection tasks.

[0056] A detection method, applied to an HJT double half-wafer detection device as described in any one of the above, is characterized by comprising: S1: Use a dual-station camera to capture images of two half-cell batteries, combined with a three-color integral light source to suppress reflection interference; S2: Use Halcon's Difference operator to separate the edge features of adjacent slices and extract candidate defect areas; S3: Use the improved YOLOv8 model to classify and detect defects, and enhance edge defect response through the CA module; S4: Dynamically compensate for dimensional measurement deviations based on environmental parameters and combine SPC analysis to monitor process capability index in real time; S5: Bind detection data with RFID and visual features to trigger closed-loop process optimization instructions.

[0057] Further preferably, the S2 further includes: calling the Intersection operator to extract the overlapping portion of the candidate area and the standard template, identifying edge collapse and corner chipping defects; calculating the diagonal length through a sub-pixel measurement tool, and correcting the thermal expansion error in combination with the temperature compensation coefficient α.

[0058] Further preferably, the inference time of the improved YOLOv8 model described in step S3 is ≤0.1 second / frame, supporting the deployment of the TensorRT acceleration engine.

[0059] Further preferably, in step S5, when 5 pieces of the same type of defects are detected consecutively, the laser repair power and the robot gripping posture are automatically adjusted in stages.

[0060] Example 1 Dual-station optical inspection system: This system utilizes a dual-station configuration, with each station equipped with a 20-megapixel area array camera with a resolution of 0.03mm / pixel. This ensures clear imaging of the edge features of double-half-cell batteries. The light source system utilizes a dual-cavity, three-color integrating light source, dynamically adjusting the light intensity ratio through a PLC to eliminate reflective interference from the velvet structure on the battery surface. A laser displacement sensor monitors vibration amplitude in real time, triggering a dynamic compensation mechanism.

[0061] Example 2 Image processing and AI algorithm implementation S1: Halcon morphological preprocessing: The Difference operator is used to separate the overlapping edges of the two half-slices, and dynamic threshold segmentation is used to extract candidate defect areas. A 5×5 circular kernel closing operation is used to filter out noise points with an area less than 0.01 mm², retaining valid defect outlines such as chipping and hidden cracks. Halcon's sub-pixel measurement tool is used to calculate diagonal length, and ambient temperature parameters are obtained through the MES interface to dynamically correct dimensional deviations caused by thermal expansion.

[0062] S2: Improved YOLOv8 model deployment: Embedded the coordinate attention mechanism (CA module) in the backbone network to enhance sensitivity to edge defects; added a 160×160 detection layer, integrated shallow high-resolution features, and improved the detection rate of microcracks (≤50μm).

[0063] S3: Dynamic Inference: Deploys the TensorRT acceleration engine, with a single-frame inference time of ≤0.1 seconds; automatically adjusts the detection box boundaries based on the cutting offset feedback from MES.

[0064] S4: Multi-source data fusion: Real-time acquisition of ambient temperature, humidity, and equipment vibration data through the OPC UA interface to dynamically optimize defect judgment thresholds.

[0065] Example 3 Production line collaboration and closed-loop control implementation: Dual binding of RFID and visual features ensures zero-error association between WaferID and inspection data; when five similar defects are detected consecutively, process optimization instructions are triggered: the laser repair power is adjusted according to the defect density level; the edge computing node generates SPC control charts in real time, with a sampling frequency of once every 50 pieces. When the process capability index Cpk is less than 1.33, an early warning is automatically pushed to the MES.

[0066] Example 4 The implementation steps of the full-process inspection method are as follows: loading the process parameter configuration file for the day, including parameters such as camera exposure time, light source intensity, and transmission speed; camera white balance calibration; light source synchronization verification, pulse error <1μs; and conveyor belt positioning accuracy test.

[0067] Unless otherwise specified, the device components involved in the above embodiments are all conventional device components, and the connection methods and control methods involved are all conventional connection methods and control methods unless otherwise specified.

[0068] The present invention has been described in detail above with reference to the embodiments. However, those skilled in the art will appreciate that, without departing from the spirit of the present invention, the specific parameters in the above embodiments may be modified to form multiple specific embodiments, which are all within the common variation range of the present invention and will not be described in detail here.

Claims

1. A HJT double half-chip detection device, characterized in that: include: High-speed optical imaging module: Consists of a dual-station 20-megapixel camera, a dual-cavity three-color integral light source, and a vacuum adsorption conveyor. The camera resolution is set to 0.03mm / pixel, and the light source system supports independent control of the red, green, and blue channels and a 10kHz strobe frequency, with a vibration suppression error of <±5μm. AI vision algorithm fusion module: Integrates the Halcon morphological processing unit with the improved YOLOv8 model. It uses the Difference operator to separate edge features of two half-slices, the Intersection operator to extract overlapping areas, and corrects size measurements based on dynamic compensation coefficients. The improved YOLOv8 model embeds a coordinate attention mechanism in the backbone network and adds a 160×160 detection layer. Production line collaborative control module: The WaferID is bound to the dual core of RFID and visual features, and the laser repair power grading adjustment instruction is triggered based on the defect distribution heat map. The edge computing node generates the process capability index control chart in real time and pushes it to the MES system. Dynamic Environmental Compensation Module: Automatically adjusts ROI boundaries and defect judgment thresholds based on cutting offset feedback from the MES and real-time environmental parameters, and is compatible with the deformation differences between 182mm and 230mm battery specifications.

2. The HJT double half-chip detection device according to claim 1, characterized in that: The calculation formula of the dynamic compensation coefficient is: , in, is the dynamic compensation coefficient, is the thermal expansion coefficient of the material, It is the difference between the real-time ambient temperature and the standard temperature of 25℃. is the temperature correction reference value, taking 100℃, The diagonal length calculated by the Halcon sub-pixel measurement tool, is the nominal value of the battery cell specification; is the humidity influencing factor, is the real-time ambient humidity, The standard humidity is 60%.

3. The HJT double half-chip detection device according to claim 1, characterized in that: The weight calculation of the coordinate attention mechanism satisfies: ， in, is the attention weight of channel c at position (x, y), and They are horizontal and vertical feature encoding respectively; and is the spatial attenuation coefficient, is the Sigmoid activation function‌.

4. The HJT double half-chip detection device according to claim 1, characterized in that: The dynamic adjustment formula of the defect judgment threshold is: , in, is the defect judgment threshold after dynamic adjustment, is the baseline defect threshold, is the humidity influencing factor, Indicates the real-time ambient humidity. Indicates standard humidity, Indicates the cutting speed deviation value, Indicates standard cutting speed.

5. The HJT double half-chip detection device according to claim 1, characterized in that: The Halcon morphological processing includes: using a 5×5 circular kernel closing operation to filter out pseudo defects with an area less than 0.01 mm²; dynamically correcting the diagonal measurement benchmark based on the temperature compensation coefficient α; fusing multi-threshold segmentation results through the Union2 operator, and implementing priority re-inspection according to the defect risk level.

6. The HJT double half-chip detection device according to claim 1, characterized in that: The improved YOLOv8 model includes: adding a 160×160 detection layer after the 18th layer of the backbone network, fusing shallow high-resolution features, and embedding a coordinate attention mechanism in the backbone network.

7. A detection method, applied to the HJT double half-chip detection device according to any one of claims 1 to 6, characterized in that: include: S1: Use a dual-station camera to capture images of two half-cell batteries, combined with a three-color integral light source to suppress reflection interference; S2: Use Halcon's Difference operator to separate the edge features of adjacent slices and extract candidate defect areas; S3: Use the improved YOLOv8 model to classify and detect defects, and enhance edge defect response through the CA module; S4: Dynamically compensate for dimensional measurement deviations based on environmental parameters and combine SPC analysis to monitor process capability index in real time; S5: Bind detection data with RFID and visual features to trigger closed-loop process optimization instructions.

8. The detection method according to claim 7, characterized in that Step S2 further includes: calling the Intersection operator to extract the overlapping part of the candidate area and the standard template, identifying edge collapse and corner chipping defects; calculating the diagonal length through a sub-pixel measurement tool, and correcting the thermal expansion error in combination with the temperature compensation coefficient α.

9. The detection method according to claim 6, characterized in that The inference time of the improved YOLOv8 model described in step S3 is ≤ 0.1 seconds per frame, and it supports the deployment of the TensorRT acceleration engine.

10. The detection method according to claim 6, characterized in that: In step S5, when five pieces of the same type of defects are detected consecutively, the laser repair power and the robot gripping posture are automatically adjusted in stages.

Images