OLED polarizer lamination correction method and system based on deep learning

Abstract

The invention relates to the technical field of modern display, and discloses an OLED polarizer lamination correction method and system based on deep learning, and the method comprises the steps: obtaining an initial dot matrix offset data set; calculating the pixel offset distribution density of each region according to the data set, and generating a region offset distribution diagram in combination with a partitioning rule; if an over-threshold region exists, extracting a feature vector and dividing an aggregation region by using DBSCAN to generate a deviation clustering result and a classification label; calculating fitting path preliminary planning parameters based on a genetic optimization algorithm and generating an initial path scheme; collecting production data in real time and processing dynamic interference to form a dynamic interference data set; performing grouping error judgment on the data set and dynamically correcting the initial path scheme to generate a corrected path scheme; and generating a control instruction according to the path correction scheme and performing dynamic fine adjustment to determine a final fitting path track. According to the method, accurate correction of micron-sized deviation can be realized.

Description

Technical Field

[0001] This invention relates to the field of modern display technology, and in particular to a method and system for bonding and correcting OLED polarizers based on deep learning. Background Technology

[0002] In the field of modern display technology, the precision of screen manufacturing processes directly affects the final visual effect and user experience. Among these components, the polarizer plays a crucial role in the overall display performance. Even slight deviations during the bonding process can easily cause light refraction or image distortion, thus affecting the display quality.

[0003] Existing bonding calibration methods rely on traditional measurement techniques and fixed rules. While these methods can meet some routine production needs, they fall short in complex real-world environments. Because the characteristics of screen pixels are extremely subtle, positional deviations and angular changes are often only at the micrometer level. Furthermore, factors such as batch-to-batch material differences and equipment fluctuations constantly accumulate, making it difficult for existing methods to capture and adjust in a timely manner. This results in the bonding path being difficult to optimize precisely in dynamically changing environments, easily leading to undercalibration issues that ultimately affect bonding quality and display performance.

[0004] In summary, existing technologies have the problem of being unable to accurately correct micron-level deviations, resulting in insufficient correction. Summary of the Invention

[0005] This invention provides a deep learning-based OLED polarizer bonding and correction method and system to achieve precise correction of micron-level deviations.

[0006] In a first aspect, to address the aforementioned technical problems, this invention provides a deep learning-based OLED polarizer bonding and correction method, comprising:

[0007] Obtain the initial dot matrix offset dataset;

[0008] Based on the initial dot matrix offset dataset, the distribution density of pixel offset in each screen area is calculated, and combined with the preset screen partitioning rules, a region offset distribution map containing the offset distribution density is obtained.

[0009] If there are regions in the regional offset distribution map that exceed the preset deviation anomaly threshold, feature vectors are extracted, and the DBSCAN algorithm is used to divide the deviation clustering regions, thereby generating regional deviation clustering results and deviation classification labels.

[0010] Based on the genetic optimization algorithm, the preliminary planning parameters of the fitting path are calculated based on the regional deviation clustering results and the deviation classification labels, and an initial fitting path scheme is generated.

[0011] Real-time acquisition of production datasets and dynamic interference feature processing to obtain dynamic interference datasets;

[0012] The dynamic interference dataset is grouped and error is determined, and the initial bonding path scheme is dynamically corrected to generate a corrected bonding path scheme.

[0013] Based on the corrected bonding path scheme, dynamic correction and control commands are generated to determine the final bonding path trajectory.

[0014] Preferably, the step of calculating the distribution density of pixel offsets within each screen region based on the initial dot matrix offset dataset, and combining it with preset screen partitioning rules to obtain a region offset distribution map containing the offset distribution density, includes:

[0015] The pixel offsets in the initial dot matrix offset dataset are divided into multiple sub-regions according to a preset screen partitioning rule, and the deviation density value is calculated for each sub-region to obtain the deviation density value of each sub-region.

[0016] If the deviation density value is greater than or equal to the preset density threshold, the sub-region is dynamically adjusted to obtain the adjusted regional deviation data.

[0017] The regional deviation data is mapped to screen partitions in a graphical manner to generate a regional offset distribution map that reflects the distribution density of deviation in each region.

[0018] Preferably, if there are regions in the regional offset distribution map that exceed a preset deviation anomaly threshold, feature vector extraction is performed, and the DBSCAN algorithm is used to divide the deviation clustering regions, thereby generating regional deviation clustering results and deviation classification labels, including:

[0019] Based on the region, abnormal pixel points are identified to obtain an abnormal point set;

[0020] For the set of outliers, the density, distribution radius, and mean offset of the outliers are calculated, and corresponding deviation feature vectors are generated.

[0021] Based on the DBSCAN algorithm and combined with a preset clustering threshold, the deviation feature vector is grouped into regions to generate regional deviation clustering results and deviation classification labels.

[0022] Preferably, the step of calculating preliminary planning parameters for the fitting path and generating an initial fitting path scheme based on the regional deviation clustering results and the deviation classification labels using a genetic optimization algorithm includes:

[0023] Based on the regional deviation clustering results and the deviation classification labels, a comprehensive screening value is calculated, and regions with a comprehensive screening value greater than a preset screening threshold are classified as key clustering regions.

[0024] Based on the key clustering regions, the multiple parameters contained therein are analyzed, and corresponding judgment thresholds are set for each parameter.

[0025] If any parameter value in the key clustering region exceeds its corresponding judgment threshold, the deviation pattern vector of that region is extracted, and the deviation pattern vector is iteratively calculated using a genetic optimization algorithm to generate preliminary planning parameters including bonding pressure, bonding speed, and bonding angle.

[0026] Based on the preliminary planning parameters, the path is adjusted to generate an initial fitting path scheme.

[0027] Preferably, the real-time acquisition of the production dataset and the processing of dynamic interference features to obtain the dynamic interference dataset include:

[0028] Real-time acquisition of production datasets, followed by classification processing based on frequency distribution and amplitude changes, yields a classified set of production data.

[0029] Based on the production data set, the time and intensity features of the interference signal are extracted. If any feature is higher than the preset corresponding feature threshold, the production data set is denoised to obtain the processed interference signal data.

[0030] The processed interference signal data is fused and matched with the classified production data set to generate the final dynamic interference dataset.

[0031] Preferably, the step of grouping and judging the dynamic interference dataset, and dynamically correcting the initial bonding path scheme to generate a corrected bonding path scheme includes:

[0032] The dynamic disturbance dataset is grouped to obtain vibration frequency data group and temperature fluctuation data group;

[0033] If any parameter in the vibration frequency data set or the temperature fluctuation data set exceeds the preset dynamic error range, a correction coefficient is calculated, and the correction coefficients are combined into a correction coefficient data set.

[0034] Based on the correction coefficient data set, the path parameters of the initial bonding path scheme are dynamically adjusted, and matched and verified with the dynamic interference dataset to generate the corrected bonding path scheme.

[0035] Preferably, the step of dynamically correcting and generating control commands based on the corrected bonding path scheme to determine the final bonding path trajectory includes:

[0036] Obtain the environmental dataset;

[0037] Based on the environmental dataset, an environmental adaptability assessment is conducted to obtain the environmental adaptability assessment results;

[0038] If any parameter in the environmental adaptability assessment result exceeds its corresponding preset judgment range, the corrected fitting path scheme is verified, the path parameters are corrected, and the adjusted path data is generated.

[0039] Based on the path data, control signals are converted, command sequences are output, and bonding operations are performed.

[0040] Feedback data is periodically collected during the bonding operation, and the execution error of the bonding path is finely adjusted in real time based on the feedback data until the error is lower than the preset tolerance, and the final bonding path trajectory is determined.

[0041] Secondly, the present invention provides an OLED polarizer bonding and correction system based on deep learning, comprising:

[0042] The data acquisition module is used to acquire the initial dot matrix offset dataset;

[0043] The offset distribution map module is used to calculate the distribution density of pixel offsets in each screen area based on the initial dot matrix offset dataset, and combine it with preset screen partitioning rules to obtain an area offset distribution map containing the offset distribution density.

[0044] The deviation module is used to extract feature vectors and use the DBSCAN algorithm to divide the deviation clustering regions if there are regions in the regional deviation distribution map that exceed the preset deviation anomaly threshold, thereby generating regional deviation clustering results and deviation classification labels.

[0045] The initial path module is used to calculate the preliminary planning parameters of the fitting path based on the regional deviation clustering results and the deviation classification labels using a genetic optimization algorithm, and to generate an initial fitting path scheme.

[0046] The interference data module is used to collect production datasets in real time and perform dynamic interference feature processing to obtain dynamic interference datasets.

[0047] The correction path module is used to group and determine the error of the dynamic interference dataset, and to dynamically correct the initial bonding path scheme to generate a corrected bonding path scheme.

[0048] The bonding path trajectory module is used to perform dynamic correction and control command generation based on the corrected bonding path scheme, and determine the final bonding path trajectory.

[0049] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the deep learning-based OLED polarizer bonding and correction method described in any one of the above.

[0050] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the deep learning-based OLED polarizer bonding and correction method described above.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] (1) In the initial stage, this invention collects micron-level offset data of screen dot matrix, divides the area according to preset screen partitioning rules and calculates offset density, so that the deviation distribution of each area is presented intuitively, which facilitates the rapid location of the area with concentrated deviation, realizes the refinement and visualization of deviation identification, and provides accurate basis for subsequent path planning.

[0053] (2) When abnormal deviation areas appear in the offset distribution map, the present invention directly identifies and clusters the abnormal pixels in these areas. By calculating the density, distribution radius and offset mean, a feature vector is generated, which realizes the group management and classification labeling of abnormal areas. It can highlight high-risk areas and process them first, reducing the risk of overall fitting deviation.

[0054] (3) In the path planning stage, this invention compares the key parameters in the clustering results with the threshold item by item, extracts the feature vectors of high deviation areas, performs iterative optimization, outputs key path parameters such as fitting pressure, speed, and angle, and generates an initial path scheme. This method can adjust the planning process according to the actual deviation characteristics, making the path more in line with the actual fitting requirements of the screen, and improving the adaptability and accuracy of the path scheme.

[0055] (4) During the production process, this invention continuously collects environmental data such as vibration frequency and temperature fluctuations, calculates correction coefficients in real time for parameters that exceed the set range, and makes adjustments and fine-tuning during path execution until the bonding error is lower than the allowable tolerance. This process achieves coordinated adaptation between the path and the environment, ensuring the stability and consistency of bonding operations in complex environments. Attached Figure Description

[0056] Figure 1This is a schematic diagram of the OLED polarizer bonding and correction method based on deep learning provided in the first embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the structure of the OLED polarizer bonding and correction system based on deep learning provided in the second embodiment of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Reference Figure 1 The first embodiment of the present invention provides a deep learning-based OLED polarizer bonding and correction method, comprising the following steps:

[0060] S11, Obtain the initial dot matrix offset dataset;

[0061] S12, based on the initial dot matrix offset dataset, calculate the distribution density of pixel offset in each screen area, and combine it with the preset screen partitioning rules to obtain a region offset distribution map containing the offset distribution density.

[0062] S13, if there is a region in the regional offset distribution map that exceeds the preset deviation anomaly threshold, feature vectors are extracted and the DBSCAN algorithm is used to divide the deviation clustering region, thereby generating regional deviation clustering results and deviation classification labels.

[0063] S14. Based on the genetic optimization algorithm, calculate the preliminary planning parameters of the fitting path based on the regional deviation clustering results and the deviation classification labels, and generate an initial fitting path scheme.

[0064] S15: Real-time acquisition of production datasets and dynamic interference feature processing to obtain dynamic interference datasets;

[0065] S16, group and determine the error of the dynamic interference dataset, and dynamically correct the initial bonding path scheme to generate the corrected bonding path scheme.

[0066] S17. Based on the corrected bonding path scheme, perform dynamic correction and control command generation to determine the final bonding path trajectory.

[0067] In step S11, the initial dot matrix offset dataset is obtained;

[0068] It's worth noting that the flexible screen is first fixed to a vibration-proof operating platform to ensure its stability and prevent additional displacement during scanning. Then, the focal length and scanning range of the high-resolution optical microscope are adjusted to match the actual screen size, and pixel-level resolution parameters are set, such as 1920×1080 resolution and 50 micrometer pixel pitch, to capture micrometer-level lateral and longitudinal displacement changes. During scanning, the device acquires raw image data of the screen's pixel matrix line by line, with each line containing the actual coordinates of the pixels. Since ambient light, mechanical vibrations, or internal noise can affect the raw image, weighted smoothing is performed on the two-dimensional pixel grayscale matrix after initial image acquisition. Specifically, the grayscale values ​​of the target pixel's 3×3 or 5×5 neighboring pixels are taken, and different weights are assigned based on distance (e.g., 0.4 for the center pixel, 0.2 for adjacent pixels, and 0.05 for diagonal pixels). The grayscale values ​​of each pixel are then weighted, summed, and normalized to obtain the smoothed grayscale of the center pixel. This method effectively suppresses isolated high-frequency noise while preserving the structural features of pixel edges. After weighting, the smoothed grayscale matrix is ​​transformed into the first image data for subsequent horizontal and vertical offset calculations.

[0069] Next, the first image data is compared point by point. Specifically, the ideal coordinates of each pixel are compared with the actual coordinates obtained from the scan, and the horizontal and vertical offsets are calculated. The offset is calculated by subtracting the ideal coordinates from the actual coordinates. For example, if the ideal position of a pixel is (100, 100) and the actual position is (100.3, 99.7), the horizontal offset is 0.3 micrometers, and the vertical offset is -0.3 micrometers. After all pixels have had their offsets calculated, they are filtered according to a preset micrometer-level threshold, typically set at 0.5 micrometers. Any pixel with an offset exceeding this threshold is marked as an abnormal pixel and compiled into an abnormal pixel set.

[0070] Next, the angle of the offset direction of the abnormal pixels needs to be extracted. Specifically, the rotational deviation is represented by calculating the angle between the line connecting the ideal and actual positions of each abnormal pixel and the horizontal axis. In this process, the reference direction of the horizontal axis is first determined, and the horizontal and vertical offsets of each abnormal point are substituted into the arctangent function to obtain the offset direction angle. For example, when a pixel's horizontal offset is 0.3 micrometers and its vertical offset is 0.2 micrometers, the offset direction is approximately 33.7 degrees by calculating arctan(0.2 / 0.3). Then, the difference between this direction and the horizontal axis is used as the rotational deviation angle, and the results are recorded in an angle offset table. After the angle offsets of all abnormal pixels are calculated, they are compiled into a second dataset containing the angle offset values.

[0071] Finally, the three types of data—horizontal offset, vertical offset, and angular deviation—are integrated to generate the final initial pixel offset dataset. During the integration process, weights are assigned based on the degree of influence of different parameters on the display effect; for example, the weight for horizontal offset is 0.5, the weight for vertical offset is 0.3, and the weight for angular deviation is 0.2. Taking an abnormal pixel as an example, if its horizontal offset is 0.6 micrometers, its vertical offset is 0.4 micrometers, and its angular deviation is 5 degrees, a comprehensive offset index can be obtained through weighted averaging, and this index is stored in the offset dataset.

[0072] In step S12, based on the initial dot matrix offset dataset, the distribution density of pixel offsets within each screen region is calculated, and combined with preset screen partitioning rules, a region offset distribution map containing the offset distribution density is obtained, including:

[0073] The pixel offsets in the initial dot matrix offset dataset are divided into multiple sub-regions according to a preset screen partitioning rule, and the deviation density value is calculated for each sub-region to obtain the deviation density value of each sub-region.

[0074] If the deviation density value is greater than or equal to the preset density threshold, the sub-region is dynamically adjusted to obtain the adjusted regional deviation data.

[0075] The regional deviation data is mapped to screen partitions in a graphical manner to generate a regional offset distribution map that reflects the distribution density of deviation in each region.

[0076] It's worth noting that the pixel horizontal, vertical, and angular offset values ​​recorded in the initial dot matrix offset dataset are divided into regions according to a preset screen partitioning rule. This partitioning rule is precisely set based on the overall screen resolution and physical size before execution. For example, a flexible screen with a resolution of 1920×1080 and a pixel pitch of 50 micrometers, if its total display area is 100,000 square micrometers, can be evenly divided into 100 small regions, each corresponding to a 100×100 pixel area, with each region having a size of 1,000 square micrometers. After partitioning, the pixel offset data contained in each region is extracted separately, forming a region-level dataset.

[0077] Subsequently, for each region's pixel offset, the deviation density value is calculated. The calculation method involves first determining the total number of pixels N in the region, then counting the number M of pixels whose offset exceeds a preset micrometer-level precision threshold, and finally obtaining the deviation density percentage using the formula M / N × 100%. This preset micrometer-level precision threshold is determined by 50% of the screen's minimum resolvable displacement. For example, if the minimum resolvable displacement is determined to be 1 micrometer through optical ranging experiments, then the threshold is set to 0.5 micrometers. For instance, if a sub-region contains 50 pixels, and 10 of these pixels have a horizontal or vertical offset exceeding 0.5 micrometers, then the deviation density for that region is 20%. After the calculation, the deviation density value for each sub-region is obtained.

[0078] Furthermore, the density results need to be weighted based on the functional partition weights of the screen. The display characteristics of flexible screens determine that different areas have different impacts on visual effects. For example, the central area of ​​the screen is used for displaying the main image and is more sensitive to offset anomalies, so its weight should be higher than that of the edge areas. The weight allocation is preset when the partitioning rules are formulated, for example, a weight of 0.6 for the central area, 0.3 for the middle area, and 0.1 for the edge area. During weighting, the deviation density value of each area is multiplied by the corresponding weight factor to obtain the weighted comprehensive deviation index. For example, if the deviation density of a certain central sub-region is 25% and its weight is 0.6, then the weighted index is 25% × 0.6 = 15%. If this index exceeds the set weighted density threshold, the weight needs to be dynamically adjusted. This weighted density threshold is specifically set as 15% × (1 + log 10(R)), where R is the radius of curvature of the flexible screen. The dynamic adjustment method involves linearly increasing the weights based on the initial weights. For example, when the deviation density exceeds the threshold by 10 percentage points, the weight is increased by 0.2, from 0.6 to 0.8, to reflect the change in the risk level of the region. This adjusted weight is then rematched with the original deviation density, and the regional deviation data is updated accordingly.

[0079] Next, a regional offset distribution map is generated to visually represent the deviation in each region. When generating the distribution map, the weighted deviation index is mapped to a color hierarchy, with higher values ​​resulting in darker colors. For example, regions with a deviation density above 30% are marked in dark red, regions between 15% and 30% in orange, and regions below 15% in green. The image is generated using a two-dimensional heatmap, with each small region corresponding to its actual screen location on the map, ensuring the distribution map accurately reflects the spatial location of offset anomalies. This visualization method allows operators to quickly identify high-risk areas and use them in subsequent deviation clustering or path optimization stages.

[0080] To illustrate with a specific example: On a flexible screen with a resolution of 2560×1440, 720 sub-regions were obtained, each with a size of 0.05 square millimeters. Analysis of the initial dot matrix offset dataset obtained from previous scanning revealed that the average offset in the central region of the screen was 0.6 micrometers, with a deviation density of 28%, exceeding the threshold of 15%; while the deviation density in the edge regions was only 5%. In the weighted calculation, the central region, with a weight of 0.6, achieved a weighted index of 16.8%, exceeding the weighted threshold of 15%, and was therefore marked as a region of high interest. In the generated region offset distribution map, this region is displayed in dark red, indicating that the offset problem in this region should be given priority consideration during subsequent clustering and path adjustment.

[0081] In step S13, if there are regions in the regional offset distribution map that exceed a preset deviation anomaly threshold, feature vector extraction is performed, and the DBSCAN algorithm is used to divide the deviation clustering regions, thereby generating regional deviation clustering results and deviation classification labels, including:

[0082] Based on the region, abnormal pixel points are identified to obtain an abnormal point set;

[0083] For the set of outliers, the density, distribution radius, and mean offset of the outliers are calculated, and corresponding deviation feature vectors are generated.

[0084] Based on the DBSCAN algorithm and combined with a preset clustering threshold, the deviation feature vector is grouped into regions to generate regional deviation clustering results and deviation classification labels.

[0085] It's worth noting that the process begins by meticulously reading the horizontal and vertical offset values ​​of each pixel, along with the overall offset magnitude, from the regional offset distribution map. For each pixel, the program compares the offset magnitude with a preset deviation anomaly threshold. This threshold is specifically 0.5 × (1 + 0.1 × R), where R is the radius of curvature of the flexible screen. For example, if the deviation anomaly threshold is set to 0.8 micrometers, and a pixel has a horizontal offset of 0.6 micrometers and a vertical offset of 1.0 micrometers, the offset magnitude calculated using the Pythagorean theorem is 1.166 micrometers, exceeding the threshold and thus being marked as an anomaly. By traversing all pixels on the screen, a complete set of anomalies is obtained, and the coordinates of each anomaly are recorded.

[0086] Next, feature extraction is performed on the set of outliers. Feature extraction includes three dimensions: outlier density, distribution radius, and mean offset. Density is calculated by the ratio of the number of outliers within a specified radius to the number of pixels covered by that radius. This specified radius is set to twice the pixel pitch. For example, on a screen with a pixel pitch of 50 micrometers, if there are 30 outliers within a circular area of ​​100 micrometers and the total number of pixels is 200, the density is 15%. The distribution radius is calculated by finding the maximum and minimum values ​​of the x and y coordinates in the set of outliers, forming a bounded rectangle using these two extreme values, and then using the diagonal length of this rectangle as the distribution radius. For example, if the nearest outlier is at (500, 500) and the farthest outlier is at (520, 490), the distribution radius is... The mean offset is calculated by averaging the offset magnitudes of all outliers. For example, if the offset values ​​of five outliers in a certain region are 1.2, 0.9, 1.1, 1.3, and 0.8 micrometers, then the mean offset is 1.06 micrometers. These characteristic parameters are combined into a deviation feature vector to represent the overall characteristics of the anomalies in that region.

[0087] To ensure comparability in subsequent clustering processes, the generated bias feature vectors are standardized. The standardization method normalizes each parameter to the range of 0 to 1 based on its maximum value. For example, when the highest density is 20%, the maximum distribution radius is 50 micrometers, and the maximum mean offset is 1.5 micrometers, a vector [15%, 22 micrometers, 1.06 micrometers] will be transformed into [0.75, 0.44, 0.71]. This standardized data avoids interference from different parameter units on clustering.

[0088] After preparing the feature vectors, the DBSCAN algorithm is used to cluster the set of outliers. Specifically, two key parameters are set: the neighborhood radius ε and the minimum number of points MinPts. For example, setting ε to a 5-pixel spacing and MinPts to 10 means that if a point has at least 10 points within a 5-pixel radius around it, then that point belongs to the cluster core. DBSCAN recursively searches for neighboring points starting from the core point, marking dense regions as a cluster until all points are classified or labeled as noise.

[0089] After clustering, each generated cluster is further assigned a risk level label based on its offset feature vector. The risk level is determined by a combination of the offset mean and density. For example, clusters with an offset mean exceeding 1.5 micrometers and a density higher than 20% are defined as high-risk groups, clusters with an offset mean between 1.0 and 1.5 micrometers and a density between 10% and 20% are medium-risk groups, and the rest are low-risk groups. The labeling results are directly mapped back to the screen area and can be used to prioritize high-risk areas in subsequent path planning.

[0090] To illustrate with a concrete example: On a 1920×1080 resolution screen, initial analysis identified approximately 5000 outliers, primarily concentrated in the upper left corner. The DBSCAN algorithm, running with ε = 10 pixels and MinPts = 20, divided the outliers into three clusters. The largest cluster, containing 2200 outliers at a density of 25% and a mean offset of 1.4 micrometers, was designated as the high-risk group. The final clustering output included not only the boundary coordinates and center point locations of each cluster but also a risk label for each cluster.

[0091] In step S14, based on the genetic optimization algorithm, preliminary planning parameters for the fitting path are calculated using the regional deviation clustering results and the deviation classification labels, and an initial fitting path scheme is generated, including:

[0092] Based on the regional deviation clustering results and the deviation classification labels, a comprehensive screening value is calculated, and regions with a comprehensive screening value greater than a preset screening threshold are classified as key clustering regions.

[0093] Based on the key clustering regions, the multiple parameters contained therein are analyzed, and corresponding judgment thresholds are set for each parameter.

[0094] If any parameter value in the key clustering region exceeds its corresponding judgment threshold, the deviation pattern vector of that region is extracted, and the deviation pattern vector is iteratively calculated using a genetic optimization algorithm to generate preliminary planning parameters including bonding pressure, bonding speed, and bonding angle.

[0095] Based on the preliminary planning parameters, the path is adjusted to generate an initial fitting path scheme.

[0096] It is worth noting that a comprehensive screening value is calculated for clustered regions based on the regional deviation clustering results and deviation classification labels. The comprehensive screening value is obtained by weighting three parameters: deviation density, mean offset, and distribution radius, and then correcting for deviation classification labels. This value is used to assess the overall risk of the clustered regions. Specifically, each parameter is first normalized; for example, the density normalization value is 0.8, the mean offset normalization value is 0.6, and the distribution radius normalization value is 0.4. Then, the parameters are weighted and summed according to preset weights (density 0.4, mean 0.4, radius 0.2) to obtain a basic comprehensive value of 0.64. Subsequently, a classification label correction coefficient is introduced; for example, the correction coefficient is 1.2 for high-risk areas, 1.0 for medium-risk areas, and 0.8 for low-risk areas. Therefore, the final comprehensive screening value for high-risk areas is 0.64 × 1.2 = 0.768, for medium-risk areas it is 0.64 × 1.0 = 0.64, and for low-risk areas it is 0.64 × 0.8 = 0.512. If the final screening composite value is higher than the preset screening threshold of 0.6, it is determined to be a key cluster area.

[0097] During the analysis process, a judgment threshold was set for each parameter. In multiple batches of flexible screen bonding experiments, long-term data were collected on three indicators: deviation density, mean offset, and distribution radius. Their distribution range under normal and abnormal bonding states was statistically analyzed. The upper quantile (e.g., 95th percentile) under normal bonding states was compared with the lower quantile (e.g., 5th percentile) under abnormal states, and the boundary value outside the overlapping interval was used as the basis for threshold setting. For example, if the deviation density was generally below 18% under normal states and above 22% under abnormal states, then the threshold was set to 20%; if the mean offset was below 1.0 micrometer under normal states and above 1.4 micrometers under abnormal states, then the threshold was set to 1.2 micrometers; if the distribution radius was below 35 micrometers under normal states and above 45 micrometers under abnormal states, then the threshold was set to 40 micrometers. If any parameter in this region exceeded the corresponding threshold, and the deviation classification label was determined to be high-risk or medium-risk, then this region was marked as a key abnormal clustering region.

[0098] Next, it is necessary to extract the deviation pattern vector of the key anomalous clustering region. The deviation pattern vector is composed of the above multidimensional parameters in a fixed order, for example, stored in the form of [density, mean, radius, directional consistency]. Each parameter is normalized, converting different dimensions into dimensionless values ​​between 0 and 1, facilitating subsequent optimization calculations. Assuming a certain anomalous region has a density of 25%, a mean offset of 1.4 micrometers, a distribution radius of 50 micrometers, and a directional consistency of 0.8, the corresponding normalized vector can be represented as [0.83, 0.93, 1.0, 0.8]. This vector fully reflects the severity and spatial characteristics of the deviation in this region.

[0099] After obtaining the deviation pattern vector, the algorithm proceeds to the iterative calculation phase of the genetic optimization algorithm. This algorithm aims to optimize the bonding path parameters, which include bonding pressure, bonding speed, and bonding angle. First, an initial population is defined, generating multiple sets of random path parameter combinations, such as a pressure range of 15 to 25 Pa, a speed range of 0.4 to 0.7 mm / s, and an angle range of 10 to 20 degrees. In each iteration, all combinations within the population are evaluated using a fitness function. The fitness function measures the quality of the path parameters by calculating the degree of matching between them and the deviation pattern vector; for example, high-density, high-deviation regions are more suitable for lower speeds and higher pressures to ensure bonding accuracy. After evaluation, combinations with high fitness are selected for the next generation, and new parameter sets are generated through crossover and mutation operations, gradually approaching the optimal solution.

[0100] The number of iterations is set based on convergence; for example, it can be stopped after 100 iterations when the fitness improvement tends to stabilize. The final output parameters are the optimal fitting path planning parameters, such as pressure of 21 Pa, speed of 0.5 mm / s, and angle of 15 degrees. These parameters are integrated into a preliminary planning parameter set and combined with screen partition data to generate an initial fitting path scheme.

[0101] When generating a path plan, the path is prioritized based on the deviation distribution of each partition. Priority is determined by the screening composite value and risk level label of each partition. For example, partitions with a screening composite value higher than 0.7 and a high-risk level are given the highest priority, followed by medium-risk partitions, and low-risk partitions last. If multiple partitions are simultaneously classified as high-risk, their screening composite values ​​are compared, with the higher value taking precedence. If the composite values ​​are the same, the average offset is used as a secondary sorting criterion to ensure that areas with larger offsets are processed first. Assuming the screen is divided into 16 partitions, and both the bottom right and top left partitions are classified as high-risk, with a screening composite value of 0.82 for the bottom right partition and 0.79 for the top left partition, the planned path will first cover the bottom right partition, then the top left partition, and then gradually cover other areas with higher offsets along the diagonal. The path is recorded as a coordinate sequence, with each node containing its location coordinates and corresponding path parameters, ensuring point-by-point control during execution.

[0102] After generating the initial path, it needs to be smoothed to eliminate broken lines or abrupt jumps caused by partition boundaries or parameter changes. Specifically, cubic spline interpolation is used to fit key points in the path sequence, generating a continuous curve. Smoothing not only improves the path's executability but also reduces mechanical impact during equipment movement, enhancing fit stability.

[0103] For example, on a flexible screen with a resolution of 2560×1440, the initial clustering analysis identified the lower right corner as a high-risk area with a deviation density of 24% and a mean offset of 1.3 micrometers. The normalized deviation pattern vector is [0.8, 0.87, 0.9, 0.75]. After iterative genetic algorithm analysis, the output bonding parameters are a pressure of 20 Pa, a speed of 0.48 mm / s, and an angle of 14 degrees. Path planning, combined with the priority of this area, generates a path sequence that covers the high-deviation area diagonally starting from the lower right corner. Finally, spline interpolation is used to obtain the initial bonding path scheme.

[0104] In step S15, the production dataset is collected in real time and subjected to dynamic interference feature processing to obtain a dynamic interference dataset, including:

[0105] Real-time acquisition of production datasets, followed by classification processing based on frequency distribution and amplitude changes, yields a classified set of production data.

[0106] Based on the production data set, the time and intensity features of the interference signal are extracted. If any feature is higher than the preset corresponding feature threshold, the production data set is denoised to obtain the processed interference signal data.

[0107] The processed interference signal data is fused and matched with the classified production data set to generate the final dynamic interference dataset.

[0108] It's worth noting that accelerometers, thermal sensors, and laser displacement sensors are deployed to monitor vibration, temperature, and displacement, respectively. The accelerometer's sampling frequency is set to 1000 Hz, capturing environmental vibration information from 0 to 200 Hz and outputting the frequency distribution and amplitude value per second. The thermal sensor has an accuracy of 0.1 degrees Celsius, recording the temperature once per minute to reflect the temperature fluctuation range. The laser displacement sensor has a resolution of 0.01 mm and is used to capture the real-time displacement error of the bonding head or workpiece. After data acquisition, the system performs preliminary classification of various data types according to preset rules. For example, vibration frequencies are classified into three segments: low frequency (0-30 Hz), medium frequency (30-100 Hz), and high frequency (above 100 Hz); temperature fluctuations are classified into three categories based on amplitude: less than 0.5 degrees Celsius, medium 0.5-2 degrees Celsius, and greater than 2 degrees Celsius; and displacement errors are classified into three categories based on offset: less than 0.2 mm, medium 0.2-0.5 mm, and greater than 0.5 mm. The output of the classification process is a set of production data, in which each record contains a category label and a corresponding numerical range.

[0109] After classification, time and intensity features need to be extracted from the original vibration waveform signal. Time features mainly include the fluctuation period and peak occurrence time, while intensity features include the main frequency peak, amplitude peak, and energy distribution. Taking actual data as an example, a vibration signal acquired in a certain instance had a main frequency of 50 Hz, an amplitude of 0.9 mm, and a fluctuation duration of 0.25 seconds. This signal exceeded the preset interference threshold (threshold is 0.6 mm) and needed to undergo noise reduction processing. Noise reduction employed wavelet transform to decompose the signal into different frequency bands, applying threshold compression to the high-frequency noise portion, retaining only the effective components of the main frequency and low-frequency bands. After processing, the signal amplitude decreased from 0.9 mm to 0.75 mm, the noise peak was effectively eliminated, and the main frequency and overall energy characteristics were preserved. The noise reduction output is the processed interference signal data, in time series format, including the main frequency value, corrected amplitude, fluctuation duration, and corresponding timestamp.

[0110] The denoised interference signal needs to be correlated and matched with the classified environmental data from the first stage to construct a complete set of interference features. Specifically, this involves matching two types of data based on timestamps and sampling windows. For example, a 50 Hz vibration signal detected at 14:05:20 is merged with temperature fluctuations (1.8 degrees Celsius) and displacement errors (0.4 mm) within the same time window to generate a multidimensional interference record. This process also calculates the Pearson correlation coefficients between the variables. For instance, within the 14:05:20 time window, the vibration frequency sampling value is 50 Hz, corresponding to an amplitude of 0.9 mm, a temperature fluctuation of 1.8 degrees Celsius, and a displacement error of 0.4 mm. Recording 200 data points at a sampling frequency of 1000 Hz, the correlation coefficient between vibration and temperature fluctuations, calculated using the above formula, is approximately 0.82, and the correlation coefficient between vibration and displacement error is approximately 0.67. This result indicates a high positive correlation between vibration and temperature fluctuations, while a moderate positive correlation exists between vibration and displacement error.

[0111] The fusion results are processed and stored as a standardized dataset, with fields including timestamp, vibration frequency, amplitude, temperature fluctuation, displacement error, fluctuation duration, and correlation coefficient. For example, a record might be formatted as: "14:05:20|50 Hz|0.75 mm|1.8 degrees Celsius|0.4 mm|0.25 seconds|Vibration-temperature correlation coefficient 0.82". This dataset ensures that various parameters can be directly accessed during subsequent dynamic adjustments to the bonding path. If certain records in the dataset exceed preset safety ranges, such as vibration frequencies above 60 Hz, temperature fluctuations exceeding 2 degrees Celsius, or displacement deviations exceeding 0.5 mm, the system will mark that period as a key interference period. The resulting dynamic interference dataset accurately reflects the vibration, temperature, and displacement fluctuations at the production site and also integrates the characteristic information of the interference signals.

[0112] In step S16, the dynamic interference dataset is grouped and error is determined, and the initial bonding path scheme is dynamically corrected to generate a corrected bonding path scheme, including:

[0113] The dynamic disturbance dataset is grouped to obtain vibration frequency data group and temperature fluctuation data group;

[0114] If any parameter in the vibration frequency data set or the temperature fluctuation data set exceeds the preset dynamic error range, a correction coefficient is calculated, and the correction coefficients are combined into a correction coefficient data set.

[0115] Based on the correction coefficient data set, the path parameters of the initial bonding path scheme are dynamically adjusted, and matched and verified with the dynamic interference dataset to generate the corrected bonding path scheme.

[0116] It is worth noting that a detailed grouping operation was performed on the dynamic interference dataset. This dataset contains two core types of information: environmental vibration frequency and temperature fluctuation amplitude, collected in real time during the production process. To ensure the accuracy of subsequent judgments, the dynamic interference dataset was grouped in two levels according to data type and collection time: First, the vibration frequency data and temperature fluctuation data were independently organized into two data groups; second, each data group was further subdivided according to the collection time period, for example, dividing the data into several sub-data groups with a time window of 0.5 seconds. This grouping method ensures that abnormal features are not masked by the overall mean and can clearly locate the duration and trend of the interference.

[0117] After grouping, error judgment is performed on the vibration frequency data group and the temperature fluctuation data group. Specifically, the value of each sub-data group is compared one by one with a preset dynamic error threshold. The dynamic error threshold is determined based on the equipment's operating tolerance range and historical production statistics; for example, the vibration frequency threshold is set to 70 Hz, and the temperature fluctuation amplitude threshold is set to 1.2 degrees Celsius. When any parameter in a sub-data group exceeds the corresponding threshold, it is marked as an "abnormal group." For example, in a certain acquisition cycle, the average value of the vibration frequency data group reaches 85 Hz, higher than the 70 Hz threshold; simultaneously, the temperature fluctuation amplitude is 1.0 degrees Celsius, not exceeding the 1.2 degrees Celsius threshold. In this case, the system only marks the vibration frequency portion as abnormal and records the cause of the abnormality and the time period for subsequent correction calculations.

[0118] Next, correction coefficients are calculated for the groups marked as anomalous. The calculation employs a recursive prediction method combined with real-time bias. Specifically, the time-series features of the anomalous groups are extracted first, such as the rate of frequency change, duration, and peak value. Then, Kalman filtering prediction is performed, fusing historical state estimates with current measured values ​​to obtain the optimal estimate. The specific formula is as follows:

[0119]

[0120] in, This is the optimal estimate. z is the predicted value from the previous time step. k K is the current measured value. k This is the gain factor, used to control the fusion ratio between predicted and measured values. The gain factor K... k Determined according to the following proportional formula: The value of α ranges from 2 to 5 and can be dynamically adjusted based on the stability of the production conditions. For example, if the vibration and temperature fluctuations in the flexible screen bonding production environment are within normal ranges (vibration below 30 Hz and temperature fluctuation below 0.5 degrees Celsius), it indicates stable operating conditions and high reliability of historical predictions. In this case, α can be set to a lower value of 2. Conversely, if a sudden disturbance occurs during production, such as a rapid increase in vibration frequency to 80 Hz and temperature fluctuation exceeding 1.5 degrees Celsius, it indicates drastic changes in real-time data, requiring a rapid response. In this case, α should be increased to 5.

[0121] Subsequently, the optimal estimate obtained is used to calculate the correction coefficient. For example, if the target reference value is 70 Hz and the optimal estimate after fusion is 78 Hz, then the correction coefficient is 70 ÷ 78 ≈ 0.9, which means that the relevant path parameters need to be scaled to 90% of the original value. For temperature fluctuation anomalies, if the target threshold is 1.2 degrees Celsius and the optimal estimate is 1.3 degrees Celsius, then the correction coefficient is 1.2 ÷ 1.3 ≈ 0.92.

[0122] Subsequently, the generated correction coefficients are integrated into a "correction coefficient data set" to correct key parameters in the initial bonding path scheme. The correction process is as follows: for the bonding pressure, bonding angle, and bonding speed parameters, corresponding correction coefficients are applied weighted according to their sensitivity to different disturbances. For example, if the bonding pressure is mainly affected by temperature fluctuations, it is adjusted using a temperature correction coefficient of 0.95; if the bonding angle and speed are mainly affected by vibration, they are adjusted using a vibration correction coefficient of 0.92. For example, if the original bonding pressure is 0.5 Newtons, it becomes 0.475 Newtons after correction with a coefficient of 0.95; if the original bonding angle is 0.02 degrees, it becomes 0.0184 degrees after correction with a coefficient of 0.92. The corrected data forms a new path parameter data set, which is then matched and verified against a dynamic disturbance dataset.

[0123] The matching verification process aligns the corrected parameters with the dynamic disturbance features over time, calculating their correlation and residual error within the same time period. For example, by comparing the corrected angle change curve with the vibration frequency change curve, if the correlation coefficient drops below 0.2 and the residual deviation is less than the preset tolerance of 0.015 degrees, the correction is deemed effective. After successful verification, the final "corrected bonding path scheme" is output. This scheme includes complete parameters such as adjusted pressure, angle, and speed, as well as corresponding timestamps and disturbance feature records, for direct use in subsequent bonding execution stages.

[0124] In step S17, based on the corrected bonding path scheme, dynamic correction and control commands are generated to determine the final bonding path trajectory, including:

[0125] Obtain the environmental dataset;

[0126] Based on the environmental dataset, an environmental adaptability assessment is conducted to obtain the environmental adaptability assessment results;

[0127] If any parameter in the environmental adaptability assessment result exceeds its corresponding preset judgment range, the corrected fitting path scheme is verified, the path parameters are corrected, and the adjusted path data is generated.

[0128] Based on the path data, control signals are converted, command sequences are output, and bonding operations are performed.

[0129] Feedback data is periodically collected during the bonding operation, and the execution error of the bonding path is finely adjusted in real time based on the feedback data until the error is lower than the preset tolerance, and the final bonding path trajectory is determined.

[0130] It's worth noting that before the bonding operation, the system needs to synchronously collect environmental parameters from multiple types of sensors deployed on the production site, including indicators that significantly affect bonding stability, such as temperature, humidity, and air pressure. For example, a high-precision thermistor sensor is used to collect the current ambient temperature at a frequency of 1Hz, with a recording accuracy of 0.1 degrees Celsius; humidity is collected by a capacitive humidity sensor, covering a range of 20% to 90%; and air pressure is obtained through a barometer, with an accuracy of ±1 hPa. The acquired raw data must be compared one by one with preset environmental adaptability threshold ranges before entering subsequent judgments. For example, the temperature is required to be controlled between 25 and 27 degrees Celsius, humidity between 40% and 60%, and air pressure between 1010 and 1020 hPa. When any indicator exceeds this range, the system immediately marks the state as "environmental anomaly" and triggers a secondary correction process for the path parameters.

[0131] In the secondary path parameter calibration stage, the operational logic involves overlaying the real-time environmental assessment results with the existing calibrated path parameters to calculate a compensation scheme for the current abnormal environment. The compensation method is based on two types of parameter adjustments: pressure compensation and speed compensation. Assuming the current humidity detection value is 65%, exceeding the upper limit by 5%, historical data shows that every 5% increase in humidity leads to a 2% decrease in adhesive strength. To maintain bonding firmness, the system linearly increases the pressure parameter from 5.0 Newtons to 5.2 Newtons. Taking abnormal air pressure as another example, when the air pressure value is 1005 hPa, below the lower limit of 5 hPa, the equipment is more prone to positional drift during bonding. Therefore, the bonding speed needs to be reduced from 2.0 mm / s to 1.8 mm / s to extend the bonding time between the material and the substrate, offsetting the risk of displacement caused by air pressure. After calibration, the system generates new path data, covering the corrected coordinates and execution parameters of all key nodes, and is marked as a "dynamic calibration version" for subsequent control command generation.

[0132] The next step is the generation and mapping of control commands. The path data consists of multiple key points, each containing three types of information: spatial coordinates, bonding speed, and bonding pressure. The system converts these physical quantities into specific motor drive signals and execution command sequences by calling a pre-built command mapping table. For example, when the coordinates of a key point on the path are (X = 100mm, Y = 50mm), corresponding to a speed of 1.8 mm / s and a pressure of 5.2 N / s, the command mapping table will output a set of PWM waveform parameters and current control values ​​to drive the linear motor to the target position at the set speed, while simultaneously triggering the pressure control valve to maintain the expected pressure. Throughout the path, similar command sequences are arranged sequentially according to the path points, forming a complete execution queue.

[0133] Finally, during the bonding operation, the system needs to monitor the path execution error in real time through periodic sampling. Error monitoring relies on high-resolution displacement sensors and a real-time sampling system to compare the current position of the device with the desired trajectory position. If the deviation exceeds a preset tolerance (e.g., 0.05 mm), fine-tuning is immediately triggered. Fine-tuning is achieved by dynamically correcting the motor drive signal. For example, if a 0.1 mm offset is detected at a turning point, the system calculates the required correction amount in real time and adjusts the motor speed and step signal to compress the deviation to within 0.03 mm. This type of fine-tuning operation is repeated in each cycle until the entire path is completed, ensuring that the trajectory accuracy remains within a controllable range.

[0134] For example, in an actual production bonding process, the initial ambient temperature was 28.5 degrees Celsius, humidity was 65%, and air pressure was 1005 hPa, both exceeding the threshold values. After evaluation, the system determined that a dual compensation operation was required, increasing the pressure by 0.2 Newtons and decreasing the speed by 0.2 mm / s. During the bonding process, real-time monitoring detected a 0.08 mm deviation at the third critical inflection point, which was fine-tuned to reduce the error to 0.04 mm. The final output bonding path trajectory achieved over 98% overlap with the expected trajectory, meeting the design accuracy requirements.

[0135] Reference Figure 2 The second embodiment of the present invention provides an OLED polarizer bonding and correction system based on deep learning, comprising:

[0136] The data acquisition module is used to acquire the initial dot matrix offset dataset;

[0137] The offset distribution map module is used to calculate the distribution density of pixel offsets in each screen area based on the initial dot matrix offset dataset, and combine it with preset screen partitioning rules to obtain an area offset distribution map containing the offset distribution density.

[0138] The deviation module is used to extract feature vectors and use the DBSCAN algorithm to divide the deviation clustering regions if there are regions in the regional deviation distribution map that exceed the preset deviation anomaly threshold, thereby generating regional deviation clustering results and deviation classification labels.

[0139] The initial path module is used to calculate the preliminary planning parameters of the fitting path based on the regional deviation clustering results and the deviation classification labels using a genetic optimization algorithm, and to generate an initial fitting path scheme.

[0140] The interference data module is used to collect production datasets in real time and perform dynamic interference feature processing to obtain dynamic interference datasets.

[0141] The correction path module is used to group and determine the error of the dynamic interference dataset, and to dynamically correct the initial bonding path scheme to generate a corrected bonding path scheme.

[0142] The bonding path trajectory module is used to perform dynamic correction and control command generation based on the corrected bonding path scheme, and determine the final bonding path trajectory.

[0143] It should be noted that the OLED polarizer bonding and correction system based on deep learning provided in this embodiment of the invention is used to execute all the process steps of the OLED polarizer bonding and correction method based on deep learning in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0144] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a deep learning-based OLED polarizer bonding and correction program. When the processor executes the computer program, it implements the steps described in the various deep learning-based OLED polarizer bonding and correction method embodiments above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the initial path module.

[0145] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0146] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0147] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0148] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0149] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0150] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0151] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A deep learning-based method for bonding and correcting OLED polarizers, characterized in that, include: Obtain the initial dot matrix offset dataset; Based on the initial dot matrix offset dataset, the distribution density of pixel offset in each screen area is calculated, and combined with the preset screen partitioning rules, a region offset distribution map containing the offset distribution density is obtained. If there are regions in the regional offset distribution map that exceed the preset deviation anomaly threshold, feature vectors are extracted, and the DBSCAN algorithm is used to divide the deviation clustering regions, thereby generating regional deviation clustering results and deviation classification labels. Based on the genetic optimization algorithm, the preliminary planning parameters of the fitting path are calculated based on the regional deviation clustering results and the deviation classification labels, and an initial fitting path scheme is generated. Real-time acquisition of production datasets and dynamic interference feature processing to obtain dynamic interference datasets; The dynamic interference dataset is grouped and error is determined, and the initial bonding path scheme is dynamically corrected to generate a corrected bonding path scheme. Based on the corrected bonding path scheme, dynamic correction and control commands are generated to determine the final bonding path trajectory.

2. The OLED polarizer bonding and correction method based on deep learning according to claim 1, characterized in that, The step of calculating the pixel offset distribution density within each screen region based on the initial dot matrix offset dataset, and combining it with preset screen partitioning rules, to obtain a region offset distribution map containing the offset distribution density, includes: The pixel offsets in the initial dot matrix offset dataset are divided into multiple sub-regions according to a preset screen partitioning rule, and the deviation density value is calculated for each sub-region to obtain the deviation density value of each sub-region. If the deviation density value is greater than or equal to the preset density threshold, the sub-region is dynamically adjusted to obtain the adjusted regional deviation data. The regional deviation data is mapped to screen partitions in a graphical manner to generate a regional offset distribution map that reflects the distribution density of deviation in each region.

3. The OLED polarizer bonding and correction method based on deep learning according to claim 1, characterized in that, If the region offset distribution map contains regions exceeding a preset deviation anomaly threshold, feature vector extraction is performed, and the DBSCAN algorithm is used to divide the deviation clustering regions, thereby generating region deviation clustering results and deviation classification labels, including: Based on the region, abnormal pixel points are identified to obtain an abnormal point set; For the set of outliers, the density, distribution radius, and mean offset of the outliers are calculated, and corresponding deviation feature vectors are generated. Based on the DBSCAN algorithm and combined with a preset clustering threshold, the deviation feature vector is grouped into regions to generate regional deviation clustering results and deviation classification labels.

4. The OLED polarizer bonding and correction method based on deep learning according to claim 1, characterized in that, The method based on the genetic optimization algorithm calculates preliminary planning parameters for the fitting path based on the regional deviation clustering results and the deviation classification labels, and generates an initial fitting path scheme, including: Based on the regional deviation clustering results and the deviation classification labels, a comprehensive screening value is calculated, and regions with a comprehensive screening value greater than a preset screening threshold are classified as key clustering regions. Based on the key clustering regions, the multiple parameters contained therein are analyzed, and corresponding judgment thresholds are set for each parameter. If any parameter value in the key clustering region exceeds its corresponding judgment threshold, the deviation pattern vector of that region is extracted, and the deviation pattern vector is iteratively calculated using a genetic optimization algorithm to generate preliminary planning parameters including bonding pressure, bonding speed, and bonding angle. Based on the preliminary planning parameters, the path is adjusted to generate an initial fitting path scheme.

5. The OLED polarizer bonding and correction method based on deep learning according to claim 1, characterized in that, The real-time acquisition of the production dataset, followed by dynamic interference feature processing, yields a dynamic interference dataset, including: Real-time acquisition of production datasets, followed by classification processing based on frequency distribution and amplitude changes, yields a classified set of production data. Based on the production data set, the time and intensity features of the interference signal are extracted. If any feature is higher than the preset corresponding feature threshold, the production data set is denoised to obtain the processed interference signal data. The processed interference signal data is fused and matched with the classified production data set to generate the final dynamic interference dataset.

6. The OLED polarizer bonding and correction method based on deep learning according to claim 1, characterized in that, The process of grouping and judging the dynamic interference dataset, and dynamically correcting the initial bonding path scheme to generate a corrected bonding path scheme includes: The dynamic disturbance dataset is grouped to obtain vibration frequency data group and temperature fluctuation data group; If any parameter in the vibration frequency data set or the temperature fluctuation data set exceeds the preset dynamic error range, a correction coefficient is calculated, and the correction coefficients are combined into a correction coefficient data set. Based on the correction coefficient data set, the path parameters of the initial bonding path scheme are dynamically adjusted, and matched and verified with the dynamic interference dataset to generate the corrected bonding path scheme.

7. The OLED polarizer bonding and correction method based on deep learning according to claim 1, characterized in that, The step of dynamically correcting and generating control commands based on the corrected bonding path scheme to determine the final bonding path trajectory includes: Obtain the environmental dataset; Based on the environmental dataset, an environmental adaptability assessment is conducted to obtain the environmental adaptability assessment results; If any parameter in the environmental adaptability assessment result exceeds its corresponding preset judgment range, the corrected fitting path scheme is verified, the path parameters are corrected, and the adjusted path data is generated. Based on the path data, control signals are converted, command sequences are output, and bonding operations are performed. Feedback data is periodically collected during the bonding operation, and the execution error of the bonding path is finely adjusted in real time based on the feedback data until the error is lower than the preset tolerance, and the final bonding path trajectory is determined.

8. A deep learning-based OLED polarizer bonding and correction system, characterized in that, include: The data acquisition module is used to acquire the initial dot matrix offset dataset; The offset distribution map module is used to calculate the distribution density of pixel offsets in each screen area based on the initial dot matrix offset dataset, and combine it with preset screen partitioning rules to obtain an area offset distribution map containing the offset distribution density. The deviation module is used to extract feature vectors and use the DBSCAN algorithm to divide the deviation clustering regions if there are regions in the regional deviation distribution map that exceed the preset deviation anomaly threshold, thereby generating regional deviation clustering results and deviation classification labels. The initial path module is used to calculate the preliminary planning parameters of the fitting path based on the regional deviation clustering results and the deviation classification labels using a genetic optimization algorithm, and to generate an initial fitting path scheme. The interference data module is used to collect production datasets in real time and perform dynamic interference feature processing to obtain dynamic interference datasets. The correction path module is used to group and determine the error of the dynamic interference dataset, and to dynamically correct the initial bonding path scheme to generate a corrected bonding path scheme. The bonding path trajectory module is used to perform dynamic correction and control command generation based on the corrected bonding path scheme, and determine the final bonding path trajectory.

9. An electronic device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the deep learning-based OLED polarizer bonding and correction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the deep learning-based OLED polarizer bonding and correction method as described in any one of claims 1 to 7.

Images