OLED display screen surface defect detection method and system

Abstract

The invention discloses an OLED display screen surface defect detection method and system, and relates to the technical field of display device surface detection.The method comprises the steps that according to different working procedure environments of an OLED display screen, the corresponding relation between working procedure environment characteristics and optical characteristic parameters of surface defects is obtained, establishing a dynamic association data set of defect physical characteristics changing along with the process, and extracting process association regularization parameters; based on the dynamic association data set, multi-modal defect features in the current process environment are extracted, a defect feature representation model is constructed through feature fusion and dimension reduction processing, and the severity level is evaluated according to the influence degree of defects on display performance; and generating a final detection report according to a severity level evaluation result. The method has the advantages of self-adaptive detection of a multi-process environment and more comprehensive measurement of defect physical characteristics, and further solves the problems of low detection accuracy and inaccurate defect severity evaluation of OLED display screen surface defects in the prior art.

Description

Technical Field

[0001] This invention relates to the field of surface inspection technology for display devices, and more specifically, to a method and system for detecting surface defects in OLED displays. Background Technology

[0002] With the rapid development of consumer electronics such as smartphones, wearable devices, and televisions, OLED displays have become the mainstream choice in the high-end display field due to their advantages such as self-emissiveness, high contrast, wide viewing angle, and flexibility. The manufacturing process of OLED display panels is complex, involving multiple steps including material feeding, debubbling, inspection, cutting, cleaning, and bonding. Each step can introduce surface defects such as bubbles, foreign objects, punctures, and burrs. These defects not only directly affect the appearance quality of the product but can also lead to functional problems such as display abnormalities, light leakage, and touch failure, severely reducing user experience and increasing after-sales repair costs. Statistics show that surface defects account for more than 30% of the total defect rate, making surface defect detection a crucial step in ensuring the quality of OLED products and of great significance for improving production yield and enterprise competitiveness.

[0003] Currently, surface defect detection for OLED displays primarily employs two methods: manual visual inspection and automated optical inspection (AOI). Manual visual inspection involves operators visually identifying surface defects under specific lighting conditions. AOI technology, on the other hand, uses industrial cameras to capture images of the product surface and employs image processing algorithms to automatically identify surface anomalies. In actual production, surface defect detection is integrated throughout the entire manufacturing process, including foreign object and puncture detection during incoming material inspection, detection of residual bubbles after degassing, detection of burrs and cracks after cutting, and comprehensive defect detection during final inspection. Furthermore, some production lines have introduced gamma testing to adjust product brightness, and visual and optical characteristic inspections are conducted at the MTL and AJU stages, forming a multi-layered, multi-angle defect detection system aimed at minimizing the entry of defective products into the market.

[0004] However, existing methods for detecting surface defects in OLED displays have significant shortcomings. While manual visual inspection offers high flexibility, it is heavily influenced by the experience, visual fatigue, and subjective judgment of inspectors, making it difficult to standardize inspection criteria and resulting in low efficiency that cannot meet the demands of large-scale production. Traditional AOI systems, although improving inspection efficiency, have low accuracy in identifying certain special defects (such as microbubbles and minor scratches), especially when dealing with transparent, highly reflective OLED materials, where missed or false detections are common. More importantly, existing inspection technologies do not fully consider the dynamic changes in defect characteristics of OLED displays under different process environments, lacking a comprehensive characterization of defect physical properties and failing to accurately assess the actual impact of defects on display performance. Furthermore, slight variations in raw materials from different batches, process parameters, and fluctuations in the production environment can all cause changes in the optical properties of the display surface, requiring frequent adjustments and optimizations of the inspection model to maintain stable inspection performance. As OLED displays develop towards higher resolution, irregular shapes, and flexibility, the limitations of existing inspection technologies become increasingly apparent.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes a method and system for detecting surface defects in OLED displays. This method and system have the advantages of adaptive detection in multi-process environments and more comprehensive measurement of the physical properties of defects, thereby solving the problems of low detection accuracy and inaccurate assessment of defect severity in existing technologies for OLED displays.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] According to one aspect of the present invention, a method for detecting surface defects in an OLED display screen is provided, the method comprising:

[0009] S1. Based on the different process environments of OLED displays, obtain the correspondence between process environment characteristics and optical characteristic parameters of surface defects, establish a dynamic correlation dataset of defect physical characteristics as the process changes, and extract process correlation regularization parameters.

[0010] S2. Based on the dynamic association dataset, extract multimodal defect features under the current process environment, construct a defect feature representation model through feature fusion and dimensionality reduction, and assess the severity level according to the degree of impact of defects on display performance;

[0011] S3. Based on the severity level assessment results, generate the final inspection report and complete the surface defect inspection of the OLED display.

[0012] Furthermore, based on the different process environments of OLED displays, the correspondence between process environment characteristics and optical property parameters of surface defects is obtained. A dynamic correlation dataset of defect physical properties changing with the process is established, and process correlation regularization parameters are extracted, including:

[0013] S11. Based on the sensor network of each process in the OLED display production line, collect key physical parameter data of different processes, and obtain standardized process environment characteristics through data preprocessing.

[0014] S12. Use multispectral reflectance measurement technology to scan and detect surface defects of OLED display screen, obtain surface defect characteristics, and extract optical characteristic parameters of surface defects;

[0015] Among them, optical characteristic parameters include reflectivity, transmittance and surface morphology features;

[0016] S13. Establish the mapping relationship between process environment characteristics and optical property parameters of surface defects, form a dynamic correlation dataset of defect physical properties as the process changes, and calculate the correlation regularization parameters for different processes.

[0017] Furthermore, based on the sensor network of each process in the OLED display production line, key physical parameter data of different processes are collected, and standardized process environment characteristics are obtained through data preprocessing, including:

[0018] S111. In the debubbling process of the display screen, pressure and temperature sensors are used to collect bubble pressure and temperature values.

[0019] S112. In the ultrasonic cleaning process of the display screen, the frequency value is obtained by an ultrasonic generator and the output power value is measured by a power meter.

[0020] S113. In the ultraviolet curing process of the display screen, the light intensity value of the UV light source is measured based on a light intensity meter, and the irradiation time value is recorded.

[0021] S114. Denoise and outlier processing are performed on the collected process parameter data, and a normalization method is used to map all parameter values ​​to a preset range to form standardized process environment characteristics.

[0022] Furthermore, multispectral reflectance measurement technology is used to scan and detect surface defects in the OLED display, obtaining surface defect features and extracting optical characteristic parameters of the surface defects, including:

[0023] S121. Use an automatic optical inspection system to scan the surface of the display screen to obtain high-resolution surface image data;

[0024] S122. Analyze surface image data using image processing algorithms, extract the average gray value of the defect area as a reflectivity index, the gray value difference between the defect and normal areas as a transmittance index, and the gradient value of the defect edge as a surface morphology index.

[0025] S123. Combining the extracted reflectivity, transmittance, and surface morphology indices, a comprehensive optical parameter vector characterizing the defect properties is constructed.

[0026] Furthermore, a mapping relationship is established between process environment characteristics and optical property parameters of surface defects, forming a dynamic correlation dataset of defect physical properties changing with the process, and correlation regularization parameters for different processes are calculated, including:

[0027] S131. Pair the process environment characteristics with the optical property parameters of surface defects according to the process sequence and sampling batch to construct an initial dataset containing process-defect pairing relationships;

[0028] S132. Analyze the initial dataset using multivariate statistical regression and machine learning algorithms to obtain the correlation law between the physical characteristics of defects and the changes in process parameters, and establish a correlation law mapping model.

[0029] S133. Based on the correlation law mapping model, calculate the influence coefficient of each process on defect formation to obtain the process correlation regularization parameters for different processes.

[0030] Furthermore, based on a dynamically correlated dataset, multimodal defect features are extracted under the current process environment. A defect feature representation model is constructed through feature fusion and dimensionality reduction. The severity level is then assessed based on the degree of impact of the defects on display performance, including:

[0031] S21. Query the range of defect physical characteristic parameters corresponding to the current process environment through the dynamic association dataset, and calculate the multimodal defect characteristics of different processes respectively.

[0032] Among them, the multimodal defect characteristics include the residual bubble characteristics of the degassing process, the surface cleanliness characteristics of the ultrasonic cleaning process, and the curing degree characteristics of the ultraviolet curing process.

[0033] S22. Using an improved OLED defect feature decomposition algorithm, tensor fusion is performed on the multimodal defect features of different processes, and the main features of the defects are extracted through dimensionality reduction to generate a comprehensive feature vector of the defects.

[0034] S23. Establish a mapping function from the comprehensive feature vector to the degree of defect impact. Calculate the comprehensive score and determine the severity level based on the degree of defect impact on display uniformity and color accuracy.

[0035] Furthermore, an improved OLED defect feature decomposition algorithm is used to perform tensor fusion on multimodal defect features from different processes, and the main features of the defects are extracted through dimensionality reduction processing to generate a comprehensive feature vector of the defects, including:

[0036] S221. Construct a multi-order defect feature tensor based on the multimodal defect features of different processes;

[0037] The multi-order defect feature tensor includes the number of defect samples, feature dimensions, and number of process detection modes;

[0038] S222. The improved OLED defect feature decomposition algorithm is used to decompose the multi-order defect feature tensor into the product of the defect core tensor and several modal factor matrices, and process correlation regularization parameters are introduced to minimize the reconstruction error.

[0039] S223. Add inter-process defect correlation constraints to determine the optimal defect core tensor and modal factor matrix, and generate a comprehensive feature vector that reflects the defect evolution characteristics of different processes.

[0040] Furthermore, the expression for the improved OLED defect feature decomposition algorithm is as follows:

[0041]

[0042] In the formula, D is the defect feature tensor; C is the defect core tensor; A, B, and E are the modal factor matrices of the number of defect samples, feature dimension, and number of process inspection modalities, respectively; W is the defect feature weight tensor. For element-wise multiplication; λ1, λ2, and λ3 are the process association regularization parameters for different processes; tr() is the trace operation of the matrix; ALA T is the matrix expression for the similarity constraint of defective samples; ||·||F is the F-norm; ||·||1 is the L1 norm; ×1, ×2, and ×3 are the pattern product operations performed along the 1st, 2nd, and 3rd dimensions of the tensor, respectively.

[0043] Furthermore, a mapping function from the comprehensive feature vector to the degree of defect impact is established. Based on the degree of impact of defects on display uniformity and color accuracy, a comprehensive score is calculated and the severity level is determined, including:

[0044] S231. Based on historical defect data and expert evaluation results, construct a random forest regression model; the input of the random forest regression model is a comprehensive feature vector, and the output is a score of the degree of influence of defects on display uniformity and color accuracy.

[0045] S232. Using mean squared error as the evaluation index, the random forest regression model is trained using cross-validation to obtain the optimized mapping function from the feature vector to the degree of defect impact.

[0046] S233. Using the trained mapping function, evaluate the comprehensive feature vector of the currently detected defect to obtain an impact score, and divide the score into different severity levels according to the preset severity threshold.

[0047] According to another aspect of the present invention, an OLED display surface defect detection system is also provided, the OLED display surface defect detection system comprising:

[0048] The defect correlation analysis module is used to obtain the correspondence between process environment characteristics and optical property parameters of surface defects based on different process environments of OLED displays, establish a dynamic correlation dataset of defect physical properties as the process changes, and extract process correlation regularization parameters.

[0049] The multimodal defect assessment module is used to extract multimodal defect features in the current process environment based on a dynamically correlated dataset, construct a defect feature representation model through feature fusion and dimensionality reduction, and assess the severity level according to the degree of impact of the defect on display performance.

[0050] The inspection report generation module is used to generate a final inspection report based on the severity level assessment results, thus completing the surface defect inspection of the OLED display.

[0051] The beneficial effects of this invention are as follows:

[0052] (1) By establishing a dynamic correlation dataset of the physical characteristics of defects changing with the process, the present invention can accurately capture the dynamic change law of defect characteristics of OLED display screens under different process environments such as debubbling, ultrasonic cleaning, and ultraviolet curing. This effectively solves the problem of neglecting the influence of process environment on defect detection in the prior art, significantly improves the accuracy and adaptability of detection, and enables the detection system to automatically adjust the detection parameters according to the changes in the actual production process, avoiding frequent manual calibration and model retraining.

[0053] (2) Through multispectral reflectance measurement technology and improved OLED defect feature decomposition algorithm, this invention achieves comprehensive characterization of defect physical properties. It can not only detect traditional visual features, but also accurately measure key physical parameters such as reflectance, transmittance and surface morphology of defects. It overcomes the detection limitations caused by existing technologies that rely solely on image visual features, especially the problems of missed detection and false detection when dealing with transparent and highly reflective OLED materials, and provides a reliable data foundation for the accurate evaluation of the impact of defects on display performance.

[0054] (3) By using multimodal defect feature fusion and a severity level assessment mechanism based on display performance impact, this invention establishes a complete mapping relationship from defect physical characteristics to display performance impact. It can accurately assess the specific impact of defects on display uniformity and color accuracy, providing a quantitative decision basis for production quality control, effectively improving the overall quality level of OLED displays, and meeting the technical requirements of high-resolution, irregular, and flexible OLED displays for precision testing. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic flowchart of a surface defect detection method for an OLED display screen according to an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of a surface defect detection system for an OLED display screen according to an embodiment of the present invention. Detailed Implementation

[0058] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0059] According to an embodiment of the present invention, a method and system for detecting surface defects in an OLED display screen are provided.

[0060] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a method for detecting surface defects in an OLED display screen is provided, the method comprising:

[0061] S1. Based on the different process environments of OLED displays, obtain the correspondence between process environment characteristics and optical characteristic parameters of surface defects, establish a dynamic correlation dataset of defect physical characteristics as the process changes, and extract process correlation regularization parameters.

[0062] S2. Based on the dynamic association dataset, extract multimodal defect features under the current process environment, construct a defect feature representation model through feature fusion and dimensionality reduction, and assess the severity level according to the degree of impact of defects on display performance;

[0063] S3. Based on the severity level assessment results, generate the final inspection report and complete the surface defect inspection of the OLED display.

[0064] In one embodiment, based on different process environments of the OLED display, the correspondence between process environment characteristics and optical property parameters of surface defects is obtained, a dynamic correlation dataset of defect physical properties changing with process is established, and process correlation regularization parameters are extracted, including:

[0065] S11. Based on the sensor network of each process in the OLED display production line, collect key physical parameter data of different processes, and obtain standardized process environment characteristics through data preprocessing.

[0066] S12. Use multispectral reflectance measurement technology to scan and detect surface defects of OLED display screen, obtain surface defect characteristics, and extract optical characteristic parameters of surface defects;

[0067] Among them, optical characteristic parameters include reflectivity, transmittance and surface morphology features;

[0068] S13. Establish the mapping relationship between process environment characteristics and optical property parameters of surface defects, form a dynamic correlation dataset of defect physical properties as the process changes, and calculate the correlation regularization parameters for different processes.

[0069] In one embodiment, based on the sensor network of each process in the OLED display production line, key physical parameter data of different processes are collected, and standardized process environment characteristics are obtained through data preprocessing, including:

[0070] S111. In the debubbling process of the display screen, pressure and temperature sensors are used to collect bubble pressure and temperature values.

[0071] S112. In the ultrasonic cleaning process of the display screen, the frequency value is obtained by an ultrasonic generator and the output power value is measured by a power meter.

[0072] S113. In the ultraviolet curing process of the display screen, the light intensity value of the UV light source is measured based on a light intensity meter, and the irradiation time value is recorded.

[0073] S114. Denoise and outlier processing are performed on the collected process parameter data, and a normalization method is used to map all parameter values ​​to a preset range to form standardized process environment characteristics.

[0074] Specifically, in the sensor network data acquisition process of the OLED display production line, firstly, high-precision pressure and temperature sensors are deployed in the degassing process, and the sensor sampling frequency is set to monitor the changes in bubble pressure and ambient temperature fluctuations in real time during the degassing process, ensuring that transient pressure changes during degassing can be captured. Then, in the ultrasonic cleaning process, real-time frequency data is acquired through a frequency detection module integrated into the ultrasonic generator, while a digital power meter is used to continuously monitor the ultrasonic output power, and a sampling interval is set to obtain data on the stability of power output during the cleaning process. Next, in the ultraviolet curing process, a calibrated light intensity meter is used to perform multi-point measurements on the UV light source, covering the entire curing area, while recording the irradiation time and light intensity attenuation curve for each batch. Finally, in the data preprocessing stage, a sliding window filtering algorithm is used to remove high-frequency noise, the 3σ criterion is used to identify and remove outliers, and finally, the Min-Max normalization method is used to uniformly map all parameter values ​​to the [0,1] interval, forming a standardized process environment feature vector with unified dimensions.

[0075] In one embodiment, multispectral reflectance measurement technology is used to scan and detect surface defects in an OLED display, obtain surface defect features, and extract optical characteristic parameters of the surface defects, including:

[0076] S121. The surface of the display screen is scanned using an automated optical inspection (AOI) system to obtain high-resolution surface image data;

[0077] S122. Analyze surface image data using image processing algorithms, extract the average gray value of the defect area as a reflectivity index, the gray value difference between the defect and normal areas as a transmittance index, and the gradient value of the defect edge as a surface morphology index.

[0078] S123. Combining the extracted reflectivity, transmittance, and surface morphology indices, a comprehensive optical parameter vector characterizing the defect properties is constructed.

[0079] Specifically, in the implementation of multispectral reflectance measurement technology, firstly, an AOI system equipped with a high-resolution CCD camera is used to scan the surface of the OLED display line by line, and the scanning resolution is set to ensure that surface defects at the micron level can be detected. The AOI system adopts a multi-angle illumination method, including LED light sources at four different angles: 0°, 15°, 30° and 45°. An image is acquired at each angle to form a multi-angle image sequence.

[0080] Specifically, in the process of analyzing surface image data using image processing algorithms, the extraction of reflectance index first involves Gaussian filtering preprocessing of the original grayscale image acquired by the AOI system to eliminate noise. Then, the Otsu thresholding method is used to automatically determine the segmentation threshold of the defect region. After binarizing the image, connected components are extracted as candidate defect regions. The arithmetic mean of the grayscale values ​​of all pixels within each defect region is calculated to obtain the average grayscale value of that defect region. In the calculation of transmittance index, a rectangular region three times the area of ​​the defect region is extended outward from the center of each detected defect region as a normal reference region. Possible defects in the reference region are then eliminated. After identifying other defective pixels, the average grayscale value of pixels within the reference area is calculated as the baseline value for the normal area. The transmittance index is obtained by dividing the absolute value of the difference between the average grayscale value of the defective area and the baseline value of the normal area by the baseline value of the normal area. For surface morphology index extraction, the Sobel edge detection operator is used to calculate the gradient of the defective area, calculating the grayscale change rate in both the horizontal and vertical directions. The gradient magnitude of each pixel is obtained through square root operations, and the distribution of gradient magnitudes across all pixels on the defective edge contour is statistically analyzed. The mean and standard deviation of the gradient magnitudes are calculated, and the surface morphology index is a weighted combination of the gradient mean and standard deviation. Finally, the three types of indices are combined to form a 12-dimensional comprehensive optical parameter vector, providing a complete data foundation for subsequent feature analysis.

[0081] In one embodiment, a mapping relationship is established between process environment characteristics and optical property parameters of surface defects, forming a dynamic correlation dataset of defect physical properties changing with the process, and correlation regularization parameters for different processes are calculated, including:

[0082] S131. Pair the process environment characteristics with the optical property parameters of surface defects according to the process sequence and sampling batch to construct an initial dataset containing process-defect pairing relationships;

[0083] S132. Analyze the initial dataset using multivariate statistical regression and machine learning algorithms to obtain the correlation law between the physical characteristics of defects and the changes in process parameters, and establish a correlation law mapping model.

[0084] S133. Based on the correlation law mapping model, calculate the influence coefficient of each process on defect formation to obtain the process correlation regularization parameters for different processes.

[0085] Specifically, in the process of establishing the dynamic association dataset, the process environment feature data and defect optical characteristic parameters are first timestamped according to production batches and time series to ensure that each defect sample can accurately correspond to the process environment state at the time of its formation. The data matching process adopts a sliding time window mechanism to match data within a preset time tolerance range.

[0086] Specifically, in establishing the correlation mapping model, the initial dataset is first subjected to data quality assessment and feature engineering. Missing value imputation, outlier detection, and data standardization ensure data quality. In the multivariate statistical regression analysis stage, stepwise regression is used to identify process parameters that significantly affect defect formation. The statistical significance of each parameter is evaluated using F-tests and t-tests, and insignificant variables with p-values ​​greater than 0.05 are removed. For the degassing process, a multivariate linear regression model is established for the bubble residue rate and degassing pressure, temperature, and time. The regression equation is Y1 = β0 + β1 × P + β2 × T + β3 × t + ε, where Y1 is the bubble residue rate, P is the degassing pressure, T is the temperature, and t is the degassing time. The regression coefficients β0, β1, β2, and β3 are estimated using the least squares method. Similarly, multivariate regression models are established for the ultrasonic cleaning and ultraviolet curing processes to construct a preliminary linear mapping relationship between process parameters and defect characteristics.

[0087] To capture more complex nonlinear correlations, this invention further applies the random forest algorithm for modeling and analysis. In this embodiment, a nonlinear regression model of process parameters and defect characteristics is constructed using random forest. Bootstrap sampling generates 100 training subsets, each training a decision tree. The maximum depth of the decision tree is set to 20, and the number of features considered when splitting at each node is 1 / 3 of the total number of features. Feature importance analysis is used to determine the contribution of each process parameter to defect formation. It is found that temperature in the degassing process, frequency in the ultrasonic cleaning process, and light intensity in the ultraviolet curing process are the three most critical factors affecting defect formation, with importance scores of 0.32, 0.28, and 0.25, respectively. Since the random forest model significantly outperforms the multivariate regression model on the test set, the prediction results of the multivariate regression model and the random forest model are finally weighted and averaged with a weight of 3:7 to form a comprehensive correlation mapping model. This model retains the interpretability of multivariate regression while fully utilizing the advantages of random forest in capturing nonlinear relationships, providing a reliable theoretical basis for the dynamic prediction of defect physical characteristics as process parameters change.

[0088] Specifically, in calculating the process correlation regularization parameters for different processes, the influence of the degassing process, ultrasonic cleaning process, and UV curing process on defect formation is analyzed using a correlation law mapping model, yielding three parameters: λ1, λ2, and λ3. λ1 ​​corresponds to the correlation regularization parameter for the degassing process, determined by statistically analyzing the influence weights of degassing pressure and temperature changes on residual bubble defects; λ2 corresponds to the correlation regularization parameter for the ultrasonic cleaning process, calculated based on the influence of ultrasonic frequency and power on surface cleanliness; and λ3 corresponds to the correlation regularization parameter for the UV curing process, obtained by evaluating the contribution of UV light intensity and irradiation time to curing degree defects. The values ​​of these three parameters typically range from 0.1 to 0.9 and satisfy the normalization constraint λ1 + λ2 + λ3 = 1, ensuring a reasonable allocation of weights for each process.

[0089] In one embodiment, based on a dynamically correlated dataset, multimodal defect features are extracted under the current process environment. A defect feature representation model is constructed through feature fusion and dimensionality reduction. The severity level is then assessed based on the degree of impact of the defects on display performance, including:

[0090] S21. Query the range of defect physical characteristic parameters corresponding to the current process environment through the dynamic association dataset, and calculate the multimodal defect characteristics of different processes respectively.

[0091] Among them, the multimodal defect characteristics include the residual bubble characteristics of the degassing process, the surface cleanliness characteristics of the ultrasonic cleaning process, and the curing degree characteristics of the ultraviolet curing process.

[0092] Specifically, in the multimodal defect feature extraction process, the first step is to query and match the current process environment parameters in the dynamic associated dataset. A nearest neighbor search algorithm is used to find the historical data record most similar to the current environment parameters, thus determining the corresponding range of physical characteristic parameters of the defects. For the degassing process, by analyzing data from pressure and temperature sensors, combined with the area, depth, and distribution density of the bubble regions detected by the AOI system, a bubble residue feature vector is calculated. The number of bubbles is obtained through connected component analysis, the average size is calculated using the arithmetic mean of all bubble areas, the depth distribution is obtained through laser ranging to determine the height difference between the bubble center and the surface, and the spatial clustering is quantified using nearest neighbor distance analysis. This vector contains 12 dimensions of feature parameters, including the number of bubbles, average size, depth distribution, and spatial clustering. For the ultrasonic cleaning process, surface cleanliness features are extracted based on ultrasonic frequency and power data, combined with surface contaminant detection results. The number of residual particles was counted using a particle counting algorithm, particle size distribution was grouped and counted using equivalent circle diameter, surface roughness was calculated using the standard deviation of height fluctuations, and cleanliness uniformity was quantified using the coefficient of variation of cleanliness in different areas. This included eight dimensions of characteristics: residual particle number, particle size distribution, surface roughness, and cleanliness uniformity. In the UV curing process, the degree of curing was calculated using light intensity and irradiation time parameters, combined with measurements of the optical properties of the cured area. Curing depth was calculated by integrating the degree of curing at different depths using infrared spectroscopy. Curing uniformity was measured using the coefficient of variation from multiple points within the cured area. Surface hardness variation was obtained by comparing the values ​​before and after nanoindentation testing. Optical transmittance was calculated by the percentage change in transmittance measured using a spectrophotometer. These ten dimensions of characteristic parameters—including curing depth, curing uniformity, surface hardness variation, and optical transmittance—ultimately formed a multimodal defect feature matrix for the three processes.

[0093] S22. Using an improved OLED defect feature decomposition algorithm, tensor fusion is performed on the multimodal defect features of different processes, and the main features of the defects are extracted through dimensionality reduction to generate a comprehensive feature vector of the defects.

[0094] S23. Establish a mapping function from the comprehensive feature vector to the degree of defect impact. Calculate the comprehensive score and determine the severity level based on the degree of defect impact on display uniformity and color accuracy.

[0095] In one embodiment, an improved OLED defect feature decomposition algorithm is used to perform tensor fusion on multimodal defect features from different processes, and the main features of the defects are extracted through dimensionality reduction to generate a comprehensive feature vector of the defects, including:

[0096] S221. Construct a multi-order defect feature tensor based on the multimodal defect features of different processes;

[0097] The multi-order defect feature tensor includes the number of defect samples, feature dimensions, and the number of process detection modes. The number of process detection modes corresponds to the number of degassing processes, ultrasonic cleaning processes, and ultraviolet curing processes mentioned in the preceding steps.

[0098] It should be noted that the term "modality" in this application refers to different data acquisition forms or representation methods, which is different from the modal concept in dynamics that represents inherent vibration characteristics. In this invention, the modal factor matrix refers to a low-rank representation matrix that corresponds to the three dimensions of defect sample number, feature dimension, and process inspection mode during tensor decomposition. Specifically, modal factor matrix A corresponds to the defect sample number dimension, with each row representing the representation of a defect sample in the dimensionality-reduced feature space; modal factor matrix B corresponds to the feature dimension, reflecting the contribution of each feature to the dimensionality-reduced space; and modal factor matrix E corresponds to the process inspection mode, characterizing the correlation between different process inspection modes.

[0099] S222. The improved OLED defect feature decomposition algorithm is used to decompose the multi-order defect feature tensor into the product of the defect core tensor and several modal factor matrices, and process correlation regularization parameters are introduced to minimize the reconstruction error.

[0100] Specifically, in this invention, the modal factor matrices A, B, and E are obtained through iterative solutions using the alternating least squares method. Specifically, in each iteration, B and E are first fixed, and A is solved by minimizing the objective function; then A and E are fixed, and B is solved; finally, A and B are fixed, and E is solved.

[0101] The modality factor matrix A has a dimension of [number of samples × R1], where R1 is the dimensionality reduction coefficient, typically set to 10%-20% of the original number of samples. Each row of matrix A corresponds to the representation of a defective sample in the reduced-dimensional space. A defective sample similarity constraint tr(ALA) is added during the solution process. T This ensures that similar defect samples remain close in the feature space.

[0102] The modality factor matrix B has a dimension of [total number of features × R²], where R² is the dimensionality reduction coefficient of the feature dimension, which is usually set to 30%-50% of the original feature dimension. The B matrix reflects the importance of each feature in the dimensionality reduction space. In the solution process, the feature sparsity constraint ||B||1 is added to select the most discriminative features.

[0103] The modality factor matrix E has the dimension of [number of process modes × R3], where R3 is the dimensionality reduction coefficient of the modality dimension, which is usually equal to or slightly smaller than the original number of modes. The E matrix represents the degree of correlation between different process detection modes. In the solution process, it is optimized by combining process correlation regularization parameters such as λ1, λ2, and λ3 to reflect the evolution law of defect characteristics between different processes.

[0104] S223. Add inter-process defect correlation constraints to determine the optimal defect core tensor and modal factor matrix, and generate a comprehensive feature vector that reflects the defect evolution characteristics of different processes.

[0105] Specifically, in the implementation of the improved OLED defect feature decomposition algorithm, a third-order tensor D is first constructed by combining the 12-dimensional bubble residue features of the degassing process, the 8-dimensional surface cleanliness features of the ultrasonic cleaning process, and the 10-dimensional curing degree features of the ultraviolet curing process. The tensor dimension is [number of samples × 30 × 3], where 30 is the sum of all feature dimensions and 3 is the number of process modes. The defect core tensor C has the dimension [R1 × R2 × R3], where R1, R2, and R3 are the core dimensions of each mode, and the optimal value is determined by a grid search method. During the tensor decomposition process, the modal factor matrices A, B, and E are iteratively solved using the alternating least squares method. In each iteration, the other two matrices are fixed, and the optimal solution for the current matrix is ​​obtained. The defect feature weight tensor W is determined by analyzing the contribution of each feature to the final detection result, and the feature importance weights are calculated using information gain and mutual information theory. The defect similarity Laplacian matrix L is constructed by calculating the Euclidean distance and Gaussian kernel function between defect samples, reflecting the similarity relationship between different defect samples. During the algorithm iteration process, the convergence threshold was set to 1e-6, and the maximum number of iterations was 1000. Finally, the comprehensive feature vector after dimensionality reduction was obtained, and the dimension was compressed to 20%-30% of the original feature space, effectively preserving the main feature information of the defect.

[0106] Specifically, the expression for the improved OLED defect feature decomposition algorithm is as follows:

[0107]

[0108] In the formula, D is the defect feature tensor, representing multimodal defect feature data collected from multiple processes; C is the defect core tensor, representing the core structural information of the defect features; A, B, and E are modal factor matrices representing the number of defect samples, feature dimensions, and number of process detection modalities, respectively; W is the defect feature weight tensor, used to adjust the importance of different defect features. λ1, λ2, and λ3 are the process association regularization parameters for different processes, corresponding to the process association regularization parameters for the degassing, ultrasonic cleaning, and UV curing processes calculated in step S13, used to adjust the weights of each constraint term; tr() is the trace operation of the matrix; ALA T is the matrix expression for the similarity constraint of defective samples; ||·||F is the F-norm; ||·||1 is the L1 norm; ×1, ×2, and ×3 are the pattern product operations performed along the 1st, 2nd, and 3rd dimensions of the tensor, respectively.

[0109] To better understand the improved OLED defect feature decomposition algorithm in this invention, the defect sample similarity constraint term tr(ALA) in the algorithm is now explained. T This will be explained in detail. In this expression, tr() is a function of ALA. T The trace operation in this matrix calculation refers to the sum of the elements on the main diagonal of a square matrix; A represents the modality factor matrix of the defect samples, with a size of n×r, where n is the number of defect samples and r is the feature dimension after dimensionality reduction; L is the Laplacian matrix of defect similarity, representing the similarity relationship between different defect samples. When the similarity between two defect samples is high, the corresponding L value is larger; A T This is the transpose of matrix A; the present invention minimizes tr(ALA) T This ensures that similar defect samples remain close in the feature space, which is crucial for tracking the evolution of defects across different processes. For example, in the above embodiment, for the inspection process of a batch of OLED displays, the defect similarity Laplace matrix L constructed by the system reflects the similarity relationship among 50 defect samples. By introducing this constraint, the algorithm successfully captures the evolution of similar bubble defects from the degassing process to the ultrasonic cleaning process, thereby significantly improving the accuracy of defect identification and reducing the false negative rate.

[0110] Specifically, the improved OLED defect feature decomposition algorithm in this invention minimizes the reconstruction error term ||DC×1A×2B×3E||F by minimizing the above-mentioned optimization objective function. 2 The algorithm incorporates three regularization terms to effectively extract and characterize defect features, including defect feature weights and defect similarity constraints, which improves the algorithm's ability to extract surface defect features of OLED displays.

[0111] In one embodiment, a mapping function from the comprehensive feature vector to the degree of defect impact is established. Based on the degree of impact of the defect on display uniformity and color accuracy, a comprehensive score is calculated and the severity level is determined, including:

[0112] S231. Based on historical defect data and expert evaluation results, construct a random forest regression model; the input of the random forest regression model is a comprehensive feature vector, and the output is a score of the degree of influence of defects on display uniformity and color accuracy.

[0113] S232. Using mean squared error as the evaluation index, the random forest regression model is trained using cross-validation to obtain the optimized mapping function from the feature vector to the degree of defect impact.

[0114] S233. Using the trained mapping function, evaluate the comprehensive feature vector of the currently detected defect to obtain an impact score, and divide the score into different severity levels according to the preset severity threshold.

[0115] Specifically, in the implementation of defect severity assessment, a training dataset containing 5000 historical defect samples is first collected. Each sample includes a comprehensive feature vector and a corresponding expert evaluation result. The expert evaluation result includes an impact score on display uniformity (0-10 points) and an impact score on color accuracy (0-10 points). The random forest regression model is constructed using multiple decision trees, and the bootstrap sampling method is used to enhance the model's generalization ability. Cross-validation is used to evaluate performance during model training, and the optimal model configuration is obtained through parameter tuning. In practical applications, the trained mapping function is applied to the currently detected defect comprehensive feature vector to obtain the impact scores on display uniformity and color accuracy. Then, the comprehensive score is calculated as a weighted average of the two scores, with weights of 0.6 and 0.4, respectively. Based on the comprehensive score, the defect severity level is divided into four levels: minor defects (0-2.5 points), general defects (2.5-5 points), severe defects (5-7.5 points), and fatal defects (7.5-10 points), providing a quantitative decision-making basis for subsequent quality control and process improvement of display surface defects.

[0116] To facilitate understanding of the above technical solution of the present invention, the following is a detailed explanation using defect detection in an OLED TV display production line as an example:

[0117] This OLED TV display production line comprises three key processes: debubbling, ultrasonic cleaning, and UV curing. First, during the defect correlation analysis phase, environmental parameters are collected through a sensor network deployed in each process: in the debubbling process, pressure sensors measure bubble pressure at 0.8-1.2 MPa, and temperature sensors record temperatures at 45-55°C; in the ultrasonic cleaning process, the ultrasonic generator frequency is set to 40 kHz, and the power meter measures an output power of 800 W; in the UV curing process, a light intensity meter measures the UV light source intensity at 15 mW / cm². 2 The irradiation time was 30 seconds. Simultaneously, multispectral reflectance measurement technology was used to scan the display surface, detecting a reflectance of 0.65, a transmittance of 0.82, and a surface morphology gradient value of 12.3 for bubble defects. Multivariate statistical regression analysis revealed a negative correlation between degassing pressure and bubble residue rate (correlation coefficient -0.76). A dynamic correlation dataset between the process environment and defect characteristics was established, and the correlation regularization parameter λ1 = 0.3 for the degassing process was calculated.

[0118] In the multimodal defect assessment phase, the system queries the dynamic correlation dataset to extract multimodal defect features under the current process environment: the degassing process detected 15 bubbles with an average size of 0.8 mm and a depth distribution of 0.2-0.5 mm; the ultrasonic cleaning process measured 8 residual particles with a surface roughness Ra = 0.12 μm; and the UV curing process detected a curing depth of 2.1 mm and a curing uniformity coefficient of variation of 0.08. A third-order tensor D∈R^(50×30×3) is constructed using an improved OLED defect feature decomposition algorithm. Tensor decomposition yields the defect core tensor C∈R^(10×8×2) and the corresponding modal factor matrix, generating a 128-dimensional comprehensive feature vector. Subsequently, the comprehensive feature vector is input into a trained random forest regression model, calculating a display uniformity impact score of 6.2 and a color accuracy impact score of 5.8, resulting in a comprehensive score of 6.04. Based on a preset threshold, this score is classified as a severe defect.

[0119] Ultimately, the system generated a test report showing that a total of 23 defects were detected in this batch of OLED displays, including 8 minor defects, 10 general defects, and 5 serious defects. It was recommended that the serious defect areas be reworked and that the degassing process pressure parameters be adjusted to 1.0 MPa to reduce the residual bubble rate. The entire testing process took 3.2 minutes, with a testing accuracy of 94.6%, which is 12.3 percentage points higher than that of traditional AOI systems.

[0120] like Figure 2 As shown, according to another embodiment of the present invention, an OLED display surface defect detection system is also provided, the OLED display surface defect detection system comprising:

[0121] The defect correlation analysis module 1 is used to obtain the correspondence between the process environment characteristics and the optical property parameters of surface defects according to the different process environments of OLED displays, establish a dynamic correlation dataset of defect physical properties as the process changes, and extract process correlation regularization parameters.

[0122] Multimodal defect assessment module 2 is used to extract multimodal defect features under the current process environment based on a dynamic correlation dataset, construct a defect feature representation model through feature fusion and dimensionality reduction, and assess the severity level according to the degree of impact of defects on display performance;

[0123] The inspection report generation module 3 is used to generate a final inspection report based on the severity level assessment results, thus completing the surface defect inspection of the OLED display.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting surface defects in an OLED display screen, characterized in that, include: S1. Based on the different process environments of OLED displays, obtain the correspondence between process environment characteristics and optical characteristic parameters of surface defects, establish a dynamic correlation dataset of defect physical characteristics as the process changes, and extract process correlation regularization parameters. S2. Based on the dynamic association dataset, extract multimodal defect features under the current process environment, construct a defect feature representation model through feature fusion and dimensionality reduction, and assess the severity level according to the degree of impact of defects on display performance; S3. Based on the severity level assessment results, generate the final inspection report and complete the surface defect inspection of the OLED display.

2. The method for detecting surface defects in an OLED display screen according to claim 1, characterized in that, The step of obtaining the correspondence between process environment characteristics and optical property parameters of surface defects based on different process environments of OLED displays, establishing a dynamic correlation dataset of defect physical properties changing with process, and extracting process correlation regularization parameters includes: S11. Based on the sensor network of each process in the OLED display production line, collect key physical parameter data of different processes, and obtain standardized process environment characteristics through data preprocessing. S12. Use multispectral reflectance measurement technology to scan and detect surface defects of OLED display screen, obtain surface defect characteristics, and extract optical characteristic parameters of surface defects; The optical characteristic parameters include reflectivity, transmittance, and surface morphology features; S13. Establish the mapping relationship between process environment characteristics and optical property parameters of surface defects, form a dynamic correlation dataset of defect physical properties as the process changes, and calculate the correlation regularization parameters for different processes.

3. The method for detecting surface defects in an OLED display screen according to claim 2, characterized in that, The sensor network based on each process in the OLED display production line collects key physical parameter data for different processes, and obtains standardized process environment characteristics through data preprocessing, including: S111. In the debubbling process of the display screen, pressure and temperature sensors are used to collect bubble pressure and temperature values. S112. In the ultrasonic cleaning process of the display screen, the frequency value is obtained by an ultrasonic generator and the output power value is measured by a power meter. S113. In the ultraviolet curing process of the display screen, the light intensity value of the UV light source is measured based on a light intensity meter, and the irradiation time value is recorded. S114. Denoise and outlier processing are performed on the collected process parameter data, and a normalization method is used to map all parameter values ​​to a preset range to form standardized process environment characteristics.

4. The method for detecting surface defects in an OLED display screen according to claim 2, characterized in that, The method of using multispectral reflectance measurement technology to scan and detect surface defects in an OLED display, obtaining surface defect features, and extracting optical characteristic parameters of the surface defects includes: S121. Use an automatic optical inspection system to scan the surface of the display screen to obtain high-resolution surface image data; S122. Analyze surface image data using image processing algorithms, extract the average gray value of the defect area as a reflectivity index, the gray value difference between the defect and normal areas as a transmittance index, and the gradient value of the defect edge as a surface morphology index. S123. Combining the extracted reflectivity, transmittance, and surface morphology indices, a comprehensive optical parameter vector characterizing the defect properties is constructed.

5. The method for detecting surface defects in an OLED display screen according to claim 2, characterized in that, The process establishes a mapping relationship between process environment characteristics and optical property parameters of surface defects, forming a dynamic correlation dataset of defect physical properties changing with the process, and calculates correlation regularization parameters for different processes, including: S131. Pair the process environment characteristics with the optical property parameters of surface defects according to the process sequence and sampling batch to construct an initial dataset containing process-defect pairing relationships; S132. Analyze the initial dataset using multivariate statistical regression and machine learning algorithms to obtain the correlation law between the physical characteristics of defects and the changes in process parameters, and establish a correlation law mapping model. S133. Based on the correlation law mapping model, calculate the influence coefficient of each process on defect formation to obtain the process correlation regularization parameters for different processes.

6. The method for detecting surface defects in an OLED display screen according to claim 1, characterized in that, The method involves extracting multimodal defect features based on a dynamically correlated dataset under the current process environment, constructing a defect feature representation model through feature fusion and dimensionality reduction, and assessing the severity level based on the degree of impact of defects on display performance, including: S21. Query the range of defect physical characteristic parameters corresponding to the current process environment through the dynamic association dataset, and calculate the multimodal defect characteristics of different processes respectively. The multimodal defect features include the bubble residue features of the degassing process, the surface cleanliness features of the ultrasonic cleaning process, and the curing degree features of the ultraviolet curing process. S22. Using an improved OLED defect feature decomposition algorithm, tensor fusion is performed on the multimodal defect features of different processes, and the main features of the defects are extracted through dimensionality reduction to generate a comprehensive feature vector of the defects. S23. Establish a mapping function from the comprehensive feature vector to the degree of defect impact. Calculate the comprehensive score and determine the severity level based on the degree of defect impact on display uniformity and color accuracy.

7. The method for detecting surface defects in an OLED display screen according to claim 6, characterized in that, The improved OLED defect feature decomposition algorithm is used to perform tensor fusion on multimodal defect features from different processes, and the main features of the defects are extracted through dimensionality reduction to generate a comprehensive feature vector of the defects, including: S221. Construct a multi-order defect feature tensor based on the multimodal defect features of different processes; The multi-order defect feature tensor includes the number of defect samples, feature dimensions, and number of process detection modes; S222. The improved OLED defect feature decomposition algorithm is used to decompose the multi-order defect feature tensor into the product of the defect core tensor and several modal factor matrices, and process correlation regularization parameters are introduced to minimize the reconstruction error. S223. Add inter-process defect correlation constraints to determine the optimal defect core tensor and modal factor matrix, and generate a comprehensive feature vector that reflects the defect evolution characteristics of different processes.

8. The method for detecting surface defects in an OLED display screen according to claim 7, characterized in that, The expression for the improved OLED defect feature decomposition algorithm is as follows: In the formula, D is the defect feature tensor; C is the defect core tensor; A, B, and E are modal factor matrices representing the number of defect samples, feature dimensions, and number of process inspection modalities, respectively; W is the defect feature weight tensor. For element-wise multiplication; λ1, λ2, and λ3 are the process association regularization parameters for different processes; tr() is the trace operation of the matrix; ALA T is the matrix expression for the similarity constraint of defective samples; ||·||F is the F-norm; ||·||1 is the L1 norm; ×1, ×2, and ×3 are the pattern product operations performed along the 1st, 2nd, and 3rd dimensions of the tensor, respectively.

9. The method for detecting surface defects in an OLED display screen according to claim 6, characterized in that, The establishment of a mapping function from the comprehensive feature vector to the degree of defect impact, and the calculation of a comprehensive score and determination of the severity level based on the degree of defect impact on display uniformity and color accuracy, include: S231. Based on historical defect data and expert evaluation results, a random forest regression model is constructed; the input of the random forest regression model is a comprehensive feature vector, and the output is a score of the degree of influence of defects on display uniformity and color accuracy. S232. Using mean squared error as the evaluation index, the random forest regression model is trained using cross-validation to obtain the optimized mapping function from the feature vector to the degree of defect impact. S233. Using the trained mapping function, evaluate the comprehensive feature vector of the currently detected defect to obtain an impact score, and divide the score into different severity levels according to the preset severity threshold.

10. An OLED display surface defect detection system, used to implement the OLED display surface defect detection method according to any one of claims 1-9, characterized in that, The system includes: The defect correlation analysis module is used to obtain the correspondence between process environment characteristics and optical property parameters of surface defects based on different process environments of OLED displays, establish a dynamic correlation dataset of defect physical properties as the process changes, and extract process correlation regularization parameters. The multimodal defect assessment module is used to extract multimodal defect features in the current process environment based on a dynamically correlated dataset, construct a defect feature representation model through feature fusion and dimensionality reduction, and assess the severity level according to the degree of impact of the defect on display performance. The inspection report generation module is used to generate a final inspection report based on the severity level assessment results, thus completing the surface defect inspection of the OLED display.

Images