OLED polarizing sheet attaching control method and system based on artificial intelligence

Abstract

The application discloses an OLED polaroid attaching control method and system based on artificial intelligence, which comprises the following steps: performing three-dimensional topography measurement on an OLED panel to be attached to obtain panel surface height distribution data, and extracting multi-order warping modal coefficients through orthogonal polynomial modal decomposition; combining the modal coefficients with attaching process parameters to construct a prediction feature vector, inputting the prediction feature vector into a pre-trained alignment offset prediction model to obtain a pushing direction translation component, a transverse translation component and an in-plane rotation component. The model uses the sum of a prediction error term and a physical symmetry constraint term as a training loss function. The alignment target coordinates of the polaroid are inversely compensated according to the prediction result, and roll-press attaching is performed according to the compensated coordinates. The application can predict attaching offset for panel warping difference, and improve the alignment precision and control stability of polaroid attaching.

Description

Technical Field

[0001] This application relates to the field of polarizer attachment control technology, specifically to an OLED polarizer attachment control method and system based on artificial intelligence. Background Technology

[0002] OLED display panels typically require the attachment of polarizers to their surfaces to reduce ambient light reflection and improve display performance. Existing automated production lines generally use a vision alignment system to identify alignment marks on the panel and polarizer. After initial positioning, a bonding roller is used to gradually press the polarizer coated with optically transparent adhesive onto the panel surface.

[0003] During the manufacturing process of OLED panels, such as thin film deposition, annealing, and encapsulation, the panel is affected by factors such as differences in the thermal expansion coefficient of the film layers, residual stress, and fluctuations in material properties, resulting in warping of varying degrees, such as bow-shaped, saddle-shaped, or asymmetrical twisting. During roll bonding, different areas of the panel experience local bending and displacement due to different warping states. The related deformations gradually accumulate along the direction of the bonding roller, which can easily cause the polarizer to shift longitudinally, laterally, and rotate in-plane relative to the panel, thereby affecting bonding accuracy and product yield.

[0004] Existing bonding equipment typically performs bonding directly based on visual alignment results or sets a fixed compensation amount based on historical deviations of products in the same batch. These methods struggle to reflect warpage differences between different panels and are ill-suited to offset fluctuations caused by variations in optically transparent adhesive viscosity, roller pressure, feed speed, and environmental temperature and humidity. Some data-driven prediction methods directly use high-dimensional surface topography data for modeling, making them susceptible to interference from local structural steps and measurement noise. Furthermore, with a limited number of samples on the production line, overfitting may occur, leading to insufficient stability in offset predictions for new panels. Therefore, existing technologies still need to further improve their ability to predict and compensate for OLED polarizer bonding offsets on a per-pane basis. Summary of the Invention

[0005] This application provides an OLED polarizer attachment control method and system based on artificial intelligence, which at least solves some of the technical problems existing in the related technologies described above.

[0006] According to a first aspect of the embodiments of this application, an artificial intelligence-based OLED polarizer attachment control method is provided, comprising: Three-dimensional topography measurements were performed on the surface of the OLED panel to be attached to obtain the height distribution data of the panel surface; Orthogonal polynomial mode decomposition was performed on the panel surface height distribution data to extract multi-order warping mode coefficients characterizing the panel warping morphology; The multi-order warping mode coefficients are combined with the attachment process parameters to construct a prediction feature vector. The prediction feature vector is then input into a pre-trained alignment offset prediction model to obtain the alignment offset prediction value of the panel. The alignment offset prediction value includes a translation component in the propulsion direction, a lateral translation component, and an in-plane rotation component. The alignment offset prediction model uses the sum of the prediction error term and the physical symmetry constraint term as the training loss function. The physical symmetry constraint term is calculated based on the sum of the model output after inverting the multi-order warping mode coefficients and the model output before inverting them. Determine the alignment target coordinates of the polarizer relative to the panel, shift the alignment target coordinates in the opposite direction of the alignment offset prediction value to obtain the compensated alignment target coordinates; position the polarizer according to the compensated alignment target coordinates and perform roll pressing attachment.

[0007] As an optional approach, the orthogonal polynomial mode decomposition employs Zernike polynomial basis functions; the in-plane coordinates of the panel are normalized to the unit circle, with the geometric center of the effective attachment area of ​​the panel as the origin and the half-length of the panel diagonal as the normalization radius; the panel surface height distribution data is fitted to a linear superposition of multiple Zernike polynomials on the normalized coordinates, and the modal coefficients are solved by the least squares method. After removing the piston term representing the overall height offset, the remaining coefficients constitute the multi-order warping modal coefficients.

[0008] As an optional approach, the prediction error term is the mean square value of the deviation between the predicted output of the alignment offset prediction model and the actual measured alignment deviation after panel attachment; the physical symmetry constraint term is calculated as follows: invert each component corresponding to the warp mode coefficient in the predicted feature vector, keep the components corresponding to the attachment process parameters unchanged, input the inverted feature vector into the alignment offset prediction model to obtain the flip prediction output, and add the flip prediction output to the original prediction output component by component and take the mean square value; the weight of the physical symmetry constraint term in the training loss function is determined by searching for the value with the optimal prediction accuracy on the validation set.

[0009] As an optional approach, the bonding process parameters include the dynamic viscosity of the current batch of optically transparent adhesive, the set pressure of the bonding roller, the set feed speed of the bonding roller, and the ambient temperature and relative humidity of the bonding station.

[0010] As an optional approach, when constructing the predicted feature vector, the features of each dimension after merging the multi-order warping mode coefficients and the attachment process parameters are standardized respectively; the standardization parameters are the mean and standard deviation of each feature in the training dataset.

[0011] As an optional approach, the alignment shift prediction model is a multilayer perceptron, which sequentially includes an input layer, multiple hidden layers, and an output layer, with each hidden layer connected in series. Each hidden layer applies a linear transformation to the output of the previous layer and then processes it through a nonlinear activation function. The input layer receives the predicted feature vector, and the output layer outputs the three components of the alignment shift prediction value.

[0012] As an optional approach, the multilayer perceptron includes three hidden layers: the first hidden layer contains 64 neurons, the second hidden layer contains 32 neurons, and the third hidden layer contains 16 neurons; the first and second hidden layers each have a dropout layer after the activation function, while the third hidden layer does not have a dropout layer; the output layer contains three neurons and uses linear output.

[0013] As an optional approach, the data used to train the alignment offset prediction model comes from production line records. Each record contains the multi-order warpage mode coefficients of the panel, the process parameters during attachment, and the actual alignment deviation measured by the automatic optical inspection process after attachment. The training uses the Adam optimizer, and the learning rate gradually decays from the initial value to the termination value according to the cosine annealing strategy. The training set and the validation set are divided proportionally, and the set of model parameters with the smallest prediction error on the validation set is saved.

[0014] As an optional approach, after roll bonding is completed, an automatic optical inspection process measures the actual alignment deviation between the polarizer and the panel. The actual alignment deviation, along with the multi-order warpage mode coefficients of the corresponding panel and the bonding process parameters, is stored in the database as new training samples. When the number of new training samples reaches a preset number, the alignment offset prediction model is incrementally trained using the current model parameters as the initial values ​​and a learning rate lower than that used in the initial training. The model is then replaced based on the accuracy of the validation set.

[0015] According to a second aspect of the embodiments of this application, an artificial intelligence-based OLED polarizer attachment control system is also provided, comprising: The three-dimensional topography measurement module is used to perform three-dimensional topography measurement on the surface of the OLED panel to be attached, and obtain the height distribution data of the panel surface. The mode decomposition module is used to perform orthogonal polynomial mode decomposition on the panel surface height distribution data to extract multi-order warping mode coefficients that characterize the warping morphology of the panel. The offset prediction module is used to combine the multi-order warp mode coefficients with the attachment process parameters to construct a prediction feature vector. The prediction feature vector is then input into a pre-trained alignment offset prediction model to obtain the alignment offset prediction value of the panel. The alignment offset prediction value includes a translation component in the propulsion direction, a lateral translation component, and an in-plane rotation component. The alignment offset prediction model uses the sum of the prediction error term and the physical symmetry constraint term as the training loss function. The physical symmetry constraint term is calculated based on the sum of the model output after inverting the multi-order warp mode coefficients and the model output before inverting them. The alignment compensation and attachment module is used to determine the alignment target coordinates of the polarizer relative to the panel, shift the alignment target coordinates in the opposite direction of the alignment offset prediction value to obtain the compensated alignment target coordinates, and position the polarizer according to the compensated alignment target coordinates and perform roll pressing attachment.

[0016] This application acquires three-dimensional topographic data of the OLED panel surface before polarizer attachment. Multi-order modal coefficients characterizing the overall warpage state of the panel are extracted through orthogonal polynomial mode decomposition. These coefficients, along with process parameters such as the dynamic viscosity of the optical transparent adhesive, rolling pressure, pushing speed, and ambient temperature and humidity, constitute the predictive features. This reduces the dimensionality of the original topographic data and weakens the interference of local structural steps and measurement noise on the prediction results. The alignment offset prediction model incorporates a physical symmetry constraint in its training loss, ensuring a corresponding reversal of the offset direction during warpage reversal. This improves the model's generalization ability even with limited production line samples. Based on the model's predicted pushing direction displacement, lateral displacement, and in-plane rotation, the alignment target coordinates of the polarizer are compensated in reverse. This enables panel-by-pane control based on the actual warpage state of different panels, reducing residual alignment deviations caused by fixed compensation. Post-attachment detection data can also be used for incremental training, allowing the model to adapt to slow changes in adhesive batches, equipment conditions, and environmental conditions. This improves polarizer attachment accuracy and production stability without altering the existing attachment machinery structure and basic process flow.

[0017] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Furthermore, no embodiment in this disclosure is required to achieve all the effects described above. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0019] Figure 1 This is a flowchart illustrating an AI-based OLED polarizer attachment control method provided in an embodiment of this disclosure.

[0020] Figure 2A flowchart for extracting multi-order warping mode coefficients provided in an embodiment of this disclosure.

[0021] Figure 3 This is a block diagram of the alignment offset prediction model provided in the embodiments of this disclosure.

[0022] Figure 4 This is a schematic block diagram of an AI-based OLED polarizer attachment control system provided in an embodiment of the present disclosure.

[0023] Figure 5 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0025] The AI-based OLED polarizer bonding control method disclosed in this embodiment is applicable to automated production lines that bond polarizers to the surface of OLED panels using a roll pressing method. The bonding surface of the polarizer is pre-coated with optically transparent adhesive (OCA). The bonding roller advances from one end to the other along the long side of the panel, gradually pressing the polarizer onto the front side of the panel. Before bonding, the panel is fixed on a vacuum adsorption platform, with a CCD vision alignment system above it. This system detects the panel alignment marks and adjusts the polarizer to the target position before initiating the roll pressing. After bonding, the panel enters a high-pressure degassing and automated optical inspection (AOI) process. The AOI measures the actual alignment deviation after bonding by identifying the relative displacement between the polarizer marks and the panel marks. For example, this method is designed for 6-inch and larger flexible OLED panels, where the production line alignment accuracy requirement is, for example, ±5 micrometers. The production line has a surface morphology detection station upstream of the bonding station, equipped with a three-dimensional measurement sensor. The three-dimensional morphology scan of the front side is completed before the panel reaches the bonding station.

[0026] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0027] Please see Figure 1 , Figure 1This is a flowchart of an OLED polarizer attachment control method based on artificial intelligence according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps: S1. Perform three-dimensional topography measurement on the surface of the OLED panel to be attached to obtain the height distribution data of the panel surface. S2, Perform orthogonal polynomial mode decomposition on the panel surface height distribution data to extract multi-order warping mode coefficients that characterize the panel warping morphology; S3, the multi-order warping mode coefficients are combined with the attachment process parameters to construct a prediction feature vector. The prediction feature vector is then input into a pre-trained alignment offset prediction model to obtain the alignment offset prediction value of the panel. The alignment offset prediction value includes a translation component in the propulsion direction, a lateral translation component, and an in-plane rotation component. The alignment offset prediction model uses the sum of the prediction error term and the physical symmetry constraint term as the training loss function. The physical symmetry constraint term is calculated based on the sum of the model output after inverting the multi-order warping mode coefficients and the model output before inversion. S4, determine the alignment target coordinates of the polarizer relative to the panel, shift the alignment target coordinates in the opposite direction of the alignment offset prediction value to obtain the compensated alignment target coordinates; position the polarizer according to the compensated alignment target coordinates and perform roll pressing attachment.

[0028] In some embodiments, for step S1, after the panel reaches the surface morphology detection station, the vacuum adsorption platform holds the bottom surface of the panel so that it adheres to the platform surface. A three-dimensional measurement sensor scans along the long side of the panel, with the scanning path covering the effective attachment area of ​​the panel; optionally, the sensor uses line laser triangulation, with an out-of-plane resolution of not less than 1 micrometer and an in-plane sampling interval configurable to 0.3 to 0.5 millimeters; after scanning, the panel surface height distribution data is obtained, expressed in a matrix... express, The coordinates for the roller pressing direction are: The horizontal coordinate is perpendicular to the direction of propulsion. The unit is micrometers, and the matrix dimension is These correspond to the number of sampling points in the two directions, respectively. As an example, for a panel with an area of ​​150 mm × 70 mm, when the sampling interval is 0.5 mm... , The matrix contains a total of 42,000 height values.

[0029] In some embodiments, for step S2, after the OLED panel undergoes multiple thin-film deposition and annealing processes in the previous steps, the difference in thermal expansion coefficients between the film layers and residual stress cause each panel to have varying degrees of warping: some are bow-shaped (the center of the panel bulges upward or downward), some are saddle-shaped (bulging and concave along two orthogonal directions respectively), and some are asymmetrically twisted. During the roll bonding process, the bonding roller advances from one end to the other. The panel warping causes small bending deformations in different directions in each area when the rollers come into contact. These local deformations gradually accumulate along the roll bonding stroke, eventually manifesting as the overall translation and rotation of the polarizer relative to the panel, i.e., alignment misalignment.

[0030] The localized steps formed on the surface of the internal TFT array, metal traces, touch electrodes, and other structural layers of the panel, typically only 0.5 to 3 micrometers high, contribute far less to the overall offset than the warpage component. (The last part, "dimension 1," appears to be a separate, unrelated sentence fragment and is omitted from the translation.) Directly inputting the original height matrix into the prediction model is problematic. Firstly, the data dimensionality is too high (e.g., 42,000 points), resulting in a model parameter count far exceeding the training sample size, which easily leads to overfitting. Secondly, high-frequency local steps constitute interference unrelated to offset, hindering the model from learning the mapping relationship between warping and offset. Therefore, this embodiment first performs modal decomposition on the panel surface height distribution data, extracting a few modal coefficients reflecting warping characteristics as input for subsequent predictions, while naturally filtering out high-frequency local details.

[0031] Specifically, please refer to Figure 2 , Figure 2 A flowchart illustrating the extraction of multi-order warping mode coefficients provided in an embodiment of this disclosure is shown, as follows: Figure 2 As shown in box 201, orthogonal polynomial mode decomposition is performed on the height distribution data of the panel surface.

[0032] According to embodiments of this disclosure, orthogonal polynomial mode decomposition employs Zernike polynomial basis functions. Zernike polynomials, defined on the unit circle, are a set of pairwise orthogonal polynomial basis functions widely used in optics to describe wavefront distortion. Arranged according to Noll numbers, each order of Zernike polynomial corresponds to a specific deformation mode: the two radial first-order components correspond to the panel along... and The overall tilt of the direction; the radial second order includes defocus term and two astigmatic terms, defocus corresponds to bow warping (the center of the panel is raised or recessed relative to the edge), and astigmatism corresponds to saddle warping (one convex and one concave along two orthogonal directions); the radial third order corresponds to coma (asymmetric bending) and cloverleaf deformation; the radial fourth order corresponds to spherical aberration and higher order symmetric deformation.

[0033] In practice, the in-plane coordinates are first normalized to the unit circle, with the geometric center of the effective attachment area of ​​the panel as the origin and the half-length of the panel diagonal as the coordinate origin. To normalize the radius, the coordinates of each sampling point are transformed as follows: ; , The coordinates are the center coordinates of the panel, in millimeters. The length is half the diagonal of the panel, in millimeters. (This must be met.) The sampling points are included in the fitting, while a small number of corner sampling points located outside the unit circle boundary are not included in the calculation. For example, when the panel diagonal length is 165 mm... The sampling points were located approximately 92 millimeters from the center at the four corners, which were outside the unit circle and were excluded. After these exclusions, approximately 37,000 valid sampling points remained.

[0034] Representing the panel height distribution on normalized coordinates as before Linear superposition of Zernike polynomials: ; Noll serial number Zernike basis functions; These are the modal coefficients, with units of and . Both are in micrometers. When 16 is selected, it covers all radial components from order 0 to 4, including warping modes such as piston, tilt, bow, saddle, coma, cloverleaf, and spherical aberration. Optionally, Adjust according to panel size and warping complexity. If the value is too small, important warp components will be missed, leading to a larger fitting residual. If the value is too large, it will introduce high-frequency components, increasing the input noise of subsequent models. For example, The method for determining this is as follows: On the validation set, from... Initially, increment by 1 to... For each The complete Zernike decomposition and model training process is executed, and the sum of the root mean square errors of the alignment offset predictions on the validation set is used as the evaluation metric. The value corresponding to the minimum of this metric is selected. Value. As a rule of thumb, the optimal size is a 6-inch panel (approximately 150mm diagonal). Typically, the value is between 14 and 17, with a 10-inch panel (approximately 250mm diagonal) being optimal. It is usually between 16 and 20.

[0035] In box 202, the multi-order warping modal coefficients characterizing the warping morphology of the panel are extracted using the least squares method. Specifically, the modal coefficients... The solution is obtained using the least squares method, with the objective function being the minimum sum of squares of the differences between the measured panel height values ​​and the Zernike superposition values ​​at all valid sampling points. A basis function matrix is ​​then constructed. : No. Line number Column elements are In the The matrix contains the values ​​at the normalized coordinates of each valid sampling point, with a matrix dimension equal to the number of valid sampling points × 100%. The coefficient vector can be obtained in one step using the pseudo-inverse matrix: ; This is a column vector of height values ​​for the effective sampling points, in micrometers; for example, hour The dimensions are 37000×16. Given a 16×16 symmetric positive definite matrix, invert it and multiply by . This yields 16 coefficients. The computational complexity of this matrix is ​​relatively low; solving for the coefficients of a single panel typically takes no more than 50 milliseconds. The first term obtained is the piston coefficient, representing the overall height offset of the panel (related to the absolute height of the panel placed on the platform), which is meaningless for describing the warp shape and is therefore removed. The 2nd to 16th terms, totaling 15 coefficients, are retained, forming a multi-order warp mode coefficient vector. Dimensions 15×1, with each component in micrometers.

[0036] As an example, among the 15 coefficients obtained after the above decomposition of a certain panel... (The defocus term, characterizing the degree of bow warping) is 3.2 micrometers. and (The values ​​for the scattered terms, which characterize saddle-shaped warping, are -1.1 μm and 0.8 μm, respectively. The absolute values ​​of the remaining higher-order terms are all less than 0.5 μm, indicating that the panel is mainly bow-shaped warping with a slight saddle-shaped component.

[0037] In some embodiments, for step S3, the multi-order warping mode coefficient vector The predicted feature vector is constructed by combining the current bonding process parameters. The bonding process parameters include five items: the dynamic viscosity of the current batch of OCA optically clear adhesive. (Unit: Pascal-second, provided by the supplier with batch test report), setting pressure of the bonding roller. (Unit: Newtons), setting the feed speed of the attachment roller (Unit: millimeters per second), ambient temperature at the attachment station (Unit: degrees Celsius), relative humidity (Unit: percentage); The above five parameters can be read in real time from the equipment control system and environmental sensors before each panel is attached, and the combined data yields the original feature vector. , dimension 20×1.

[0038] The numerical magnitudes of the 15 warpage mode coefficients are typically in the range of -5 to 5 micrometers, while the dimensions and magnitudes of the 5 attachment process parameters differ significantly. For example, the dynamic viscosity is on the order of 0.01 Pascals per second, the roller pressure is tens to hundreds of Newtons, and the temperature is 20 to 30 degrees Celsius. Directly inputting these 20 dimensions into the neural network results in features with large numerical magnitudes dominating the gradient calculation direction, while features with small magnitudes contribute little to the model parameter updates, hindering model convergence. Therefore, z-score standardization is performed on each feature dimension: the z-score is calculated on the training dataset... Mean of dimensional features and standard deviation Transform this dimension into: ; For the first dimensional original eigenvalues, and The unit and the feature of that dimension are the same, after transformation It is a dimensionless quantity; the standardized predicted feature vector is denoted as . The model has a dimension of 20×1. The 20 sets of means and standard deviations obtained during the training phase are stored as fixed parameters after the model is deployed. The same parameters are used to standardize each new panel.

[0039] For example, the third-dimensional original feature of a panel (Zernike third-order coefficients, i.e.) The directional tilt term has a value of 1.8 micrometers. The mean of this dimension in the training set is 0.3 micrometers, and the standard deviation is 1.2 micrometers. After standardization, this dimension takes the value... The 16th dimension, the original feature (OCA dynamic viscosity), is 0.015 Pa·s. The mean of this dimension in the training set is 0.013 Pa·s, and the standard deviation is 0.002 Pa·s. After standardization, it takes the following values: .

[0040] The alignment offset prediction model receives the generated 20-dimensional normalized prediction feature vector. As input, the warpage morphology of the panel determines the direction and magnitude of local bending deformation in each area during the rolling process. The cumulative effect of these deformations along the rolling stroke is manifested as the overall translational and rotational offset of the polarizer relative to the panel. The bonding process parameters affect the mechanical response characteristics of the panel and the OCA adhesive layer during the rolling contact process, thereby changing the magnitude of the offset. Therefore, using the warpage mode coefficient and bonding process parameters as inputs at the same time can capture the combined effect of the above two types of factors.

[0041] The model output is a predicted value of the alignment offset that the panel will produce during the roll bonding process, which includes three components: translational component in the pushing direction. (Unit: micrometer), lateral translation component (Unit: micrometer), in-plane rotational component (Unit: radians).

[0042] Since the translational and lateral translational components in the propulsion direction are typically on the order of several micrometers to tens of micrometers, the in-plane rotational component is typically... In radian-scale calculations, if physical quantities are directly used as learning targets during training, the large translational components will dominate the gradient direction of the loss function, while the rotational components will not be adequately learned. The output labels are also standardized: the standard deviations of the translational components in the propulsion direction, the lateral translational components, and the in-plane rotational components are calculated separately on the training dataset. , , Each component is divided by its corresponding standard deviation to obtain the dimensionless standardized output; during training, the standardized offset is used as the learning target, and during inference, the model output is multiplied by its corresponding standard deviation to restore the physical quantity; the standardized model output is denoted as... , dimension 3×1.

[0043] The training data is derived from historical production records of the production line. Each record contains: a vector of multi-order warpage mode coefficients of the panel after processing in step S1. (15-dimensional), bonding process parameters during bonding ( , , , , (A total of 5 items), and the actual alignment deviations measured in the AOI process after attachment (measured values ​​of translational components in the advancing direction, lateral translational components, and in-plane rotational components). Training will begin after collecting no less than 2000 valid records.

[0044] Optionally, the validity criteria for each record include: (1) the AOI process successfully identifies the polarizer mark and panel mark with a confidence level of not less than 0.95; (2) there are no equipment alarms or shutdowns during the attachment process; and (3) all three components of the actual alignment deviation are within the AOI measurement range (translation in the advancing direction ±200 micrometers, lateral translation ±150 micrometers, and in-plane rotation ±0.01 radians). The annotation data is automatically generated by the AOI system without manual verification; optionally, the training dataset covers production records for 6 consecutive months and includes at least 3 different OCA batches to ensure the integrity of the process parameter distribution.

[0045] In one embodiment, the alignment offset prediction model employs a multilayer perceptron (MLP). After Zernike mode decomposition, the input is only 20-dimensional, and there are no spatial local features or temporal dependencies that require convolutional operations for extraction. The mapping from warped mode coefficients to alignment offsets is physically continuous and smooth, meaning that small changes in the warped coefficients correspond to continuous changes in the offsets. A multilayer perceptron with three hidden layers is sufficient to fit this mapping, and the inference speed is fast.

[0046] Please see Figure 3 , Figure 3 The diagram illustrates the alignment shift prediction model structure provided in this embodiment. Specifically, the network, from the input layer to the output layer, includes an input layer that receives a 20-dimensional standardized prediction feature vector. The first hidden layer contains 64 neurons. After applying a linear transformation to the input, it passes through a ReLU activation function, followed by a dropout layer. The dropout probability can be configured from 0.1 to 0.2, for example, 0.15. The second hidden layer contains 32 neurons, undergoing a linear transformation, ReLU activation, and a dropout layer, with the same structure as the first hidden layer. The third hidden layer contains 16 neurons. After a linear transformation, it passes through a ReLU activation function, without a dropout layer. Since the number of neurons in this layer is already reduced to 16, the information redundancy is low. Introducing random dropout here might lead to the loss of effective features, thus impairing prediction accuracy. The output layer contains 3 neurons. It applies a linear transformation to the output of the third hidden layer, without using an activation function, directly outputting a 3D standardized prediction vector. .

[0047] The hidden layers are connected sequentially, without residual skip connections or branching structures. The number of neurons in the hidden layers decreases from 64 to 32 and then to 16, compressing the dimension of the feature representation layer by layer, so that the network can gradually extract information related to the 3-dimensional output from the 20-dimensional input. During the training phase, the dropout layer randomly sets the output of some neurons to zero with a certain probability, forcing the network not to over-rely on the response of certain specific neurons, thereby suppressing overfitting. During the inference phase, the dropout layer is turned off, and all neurons participate in the calculation normally.

[0048] In this embodiment, the total number of parameters of the MLP model is approximately [missing information]. The model (including bias terms) can be used for real-time inference on an industrial control computer (such as an ARM Cortex-A series or x86 embedded processor) without GPU acceleration. It is understood that the model parameters can be increased according to the actual situation, and this disclosure does not impose any restrictions.

[0049] In one embodiment, the conventional approach is to train the model using only the mean squared error between predicted and measured values. However, each product specification on the production line typically accumulates only a few thousand records. With a limited number of training samples, the model tends to memorize measurement noise and occasional disturbances in the training set rather than the physical laws governing warping and offset, leading to increased prediction bias on new panels and insufficient generalization ability. Therefore, this embodiment introduces a physical symmetry constraint term into the loss function, subjecting the model to constraints from both data fitting accuracy and physical laws during training.

[0050] The basis for the physical symmetry constraint lies in the linear superposition property of Zernike decomposition. It can be seen that all modal coefficients Inverting yields At that time, the height distribution of the reconstruction is That is, all height values ​​at each point on the surface have the opposite sign. Therefore, the multi-order warp mode coefficient vector is... Invert all components to obtain Physically equivalent to the height distribution of the panel surface being determined by... Become That is, the originally raised areas become concave, and the originally concave areas become raised, with the warping direction completely reversed. After the warping is reversed, the bending deformation direction of each area of ​​the panel during the rolling process is also reversed, the direction of the lateral force on the polarizer is reversed, and the direction of the alignment offset is reversed accordingly. Under the condition that the bonding process parameters remain unchanged, the model's output for the warping reversal input should be approximately equal to the negative value of the original output.

[0051] It should be noted that the premise for warpage reversal leading to offset direction reversal is that the bonding process parameters (roller pressure, feed speed, OCA viscosity, etc.) remain unchanged, and the mechanical response of the roller contact is symmetrical before and after the warpage direction reversal. Assuming the OCA adhesive layer is an isotropic material, the compressive deformation experienced by the raised area of ​​the panel during roller pressing is approximately equal in magnitude and opposite in direction to the tensile deformation experienced by the recessed area. Therefore, the cumulative offset direction reverses while the magnitude remains unchanged. This symmetry may deviate slightly when the adhesive layer exhibits nonlinear viscoelasticity. Therefore, this embodiment introduces the symmetry into the loss function in the form of soft constraints (rather than hard constraints), allowing the model to learn the deviation from the ideal symmetry under data-driven conditions.

[0052] Therefore, the training loss function is designed as the sum of the prediction error term and the physical symmetry constraint term: ; This represents the total loss.

[0053] Prediction error term Using the mean square error of the standardized output: ; This represents the number of samples in the current training batch. For the model to the first The predicted output for each sample is a 3-dimensional standardized dimensionless vector. The standardized measured offset label for this sample is also a 3-dimensional dimensionless vector. It represents the sum of the squares of the vector components; It is a dimensionless quantity.

[0054] In one embodiment, the physical symmetry constraint term Constructed as follows: For each sample in the training batch Take its 20-dimensional standardized prediction feature vector The first 15 dimensions represent the standardized multi-order warpage mode coefficients, and the last 5 dimensions represent the standardized attachment process parameters. Inverting each component of the first 15 dimensions while keeping the last 5 dimensions unchanged yields the flipped feature vector. The inversion operation here is performed in the standardized dimensionless space; z-score standardization is a linear transformation, and given that the mean of the warped mode coefficients in the training set is close to zero, inverting the standardized values ​​is approximately equivalent to directly inverting the mode coefficients in the original physical space. Input alignment offset prediction model to obtain flip prediction output The inverted prediction output is compared with the original prediction output. After summing the components, take the mean square value: ; This is the forward inference function for the alignment offset prediction model, outputting a 3×1 dimensionless standardized value; when the model fully satisfies physical symmetry... It is a zero vector. The value is 0; the greater the deviation, The larger the value, the more the network parameters will adjust in a direction that satisfies symmetry through gradient backpropagation.

[0055] During back propagation, The gradient simultaneously passes through and Two forward paths propagate back to shared network parameters. Specifically, for each sample in a batch, the model performs two forward inferences (once for the original features and once for the flipped features), with both inferences sharing the same set of network weights; during backpropagation, the gradients of the two paths are summed and used to update the weights. The number of inferences per training iteration is twice that of regular training, but due to the relatively small network size (approximately 2000 parameters), the inference overhead remains within an acceptable range.

[0056] in, The weights of the physical symmetry constraint terms control the balance between prediction accuracy and physical constraints. The larger the value, the stronger the constraint on physical symmetry, but this may come at the cost of sacrificing fitting accuracy on the training set. If the value is too small, the constraint effect will be insignificant. The value was determined by a grid search on the validation set, with a search range of 0.1 to 1.0. As an example, values ​​were taken as follows: Five models were trained with values ​​of 0.1, 0.3, 0.5, 0.7, and 1.0. The sum of the root mean square errors of the predictions of the three components—the translational component in the propulsion direction, the lateral translation component, and the in-plane rotation component—on the validation set was used as the evaluation metric. The model corresponding to the minimum value of this metric was selected. The calibration was performed using 2500 data points for a specific product specification. When the value is 0.5, the evaluation index is 7.4 (the root mean square errors of each component are 2.8, 2.5, and 2.1, respectively, and the units are standardized dimensionless values), which is lower than... 9.1 and Version 8.0 at the time.

[0057] By introducing a physical symmetry constraint, on the one hand, during training, each actual sample automatically generates a virtual flipped sample (warp inverted, expected offset inverted) without additional labeling, effectively doubling the amount of available training constraint information and expanding the model's effective learning signal under limited production line data. On the other hand, this constraint restricts the model's effective hypothesis space to a family of mappings with odd function characteristics about the warp coefficient, excluding physically unreasonable mapping patterns, such as two panels with completely opposite warp directions producing the same directional offset, thereby reducing the possibility of model memory noise. Removing this constraint (i.e., letting...) After that, with the same amount of training data, the root mean square error of the model's prediction on the validation set increased by about 40% to 60%.

[0058] Optionally, the model training process includes training using the Adam optimizer, with an initial learning rate, for example, set to... Adjusted using a cosine annealing strategy: The learning rate gradually decays from the initial value to the terminal value, following the decreasing segment of the cosine function. The terminal value is, for example, [missing value]. The decay process allows for rapid approximation of the low-value region of the loss function with a larger step size in the early stages of training, followed by a finer search within that region with a smaller step size in the later stages. The batch size can be configured from 16 to 64, for example, 32. A larger batch size increases the redundancy of gradient updates per step and is unsuitable for small datasets, while a smaller batch size leads to larger gradient estimation variance and training oscillations. The number of training epochs can be configured from 150 to 300, for example, 200. The training and validation sets are randomly divided in an 8:2 ratio. The sum of the root mean square errors of the three offset components predicted on the validation set is used as the model selection criterion, and the model parameters with the smallest error are saved.

[0059] For example, after collecting 2,500 valid production line records, a training set of 2,000 records and a validation set of 500 records were created. After 200 training rounds, on the validation set, the root mean square errors of the translational component in the propulsion direction and the lateral translation component, after being restored to physical quantities, were 2.6 micrometers and 2.3 micrometers, respectively, and the root mean square error of the in-plane rotation component was 0.004 degrees. All three components were less than half of the alignment accuracy requirement, thus meeting the deployment conditions.

[0060] In some embodiments, for step S4, after the panel arrives at the attachment station, the vision alignment system identifies the alignment marks on the panel and determines the alignment target coordinates of the polarizer relative to the panel. , and The unit is micrometers. The unit is radians. Meanwhile, the panel has completed the mode decomposition and feature construction in step S1 at the upstream topography inspection station, and the standardized predicted feature vector... Input the alignment shift prediction model trained in the previous steps. The model outputs 3D standardized values. Multiply each component by the standard deviation of the corresponding offset component in the training set to restore it to a physical quantity: , , These are the translational component (micrometers) in the propulsion direction, the lateral translational component (micrometers) and the in-plane rotational component (radians), respectively.

[0061] The alignment control system shifts the alignment target coordinates in the opposite direction to the predicted alignment offset, thus obtaining the compensated alignment target coordinates: ; , , All units are in micrometers; , , All units are in micrometers; , , All units are in radians. The mounting equipment's motion platform positions the polarizer according to the compensated alignment target coordinates, and then starts the rolling process. During the rolling process, the actual offset caused by panel warping is opposite to the direction of the pre-compensation amount; after the two cancel each other out, the alignment deviation at the end of the mounting process is close to zero.

[0062] As an example, a panel, after undergoing the aforementioned steps, yields a multi-order warpage mode coefficient vector. This vector is then combined with the bonding process parameters, standardized, and input into the model. The standardized prediction output is... Standard deviation of the translation component in the training focus direction micrometer, standard deviation of lateral translation component micrometer, standard deviation of in-plane rotation component Radius. The predicted alignment offset after restoration is: micrometer, micrometer, Radius. The propulsion direction component in the original alignment target coordinates given by the visual alignment system is 5000.000 micrometers, which is compensated to be... Micrometers; the lateral component is 3500.000 micrometers, after compensation it is... Micrometers; rotation component is 0 radians, after compensation is Curvature. The motion platform positions the polarizer according to the compensated coordinates and performs roll pressing. After the attachment is completed, the residual alignment deviation measured by AOI is reduced to within ±5 micrometers.

[0063] In one embodiment, after roll bonding is completed, the panel enters the AOI process, where the actual alignment deviation between the polarizer mark and the panel mark is measured, including the measured values ​​of the translation component in the advancing direction, the lateral translation component, and the in-plane rotation component. This actual alignment deviation, along with the multi-order warp mode coefficient vector of the panel and the bonding process parameters at that time, is written into the database to form a new training sample.

[0064] In one embodiment, during continuous production line operation, batch switching of OCA adhesive can cause changes in the adhesive layer viscosity, wear and tear on equipment components can gradually alter the roller contact conditions, and seasonal temperature and humidity fluctuations can also affect the mechanical properties of the adhesive layer. These factors may cause a slow change in the mapping relationship between warpage and offset. When the number of newly added training samples in the database reaches a preset number, for example, every 500 newly added valid records, the training and validation sets are re-divided using all the most recent data. Incremental training is performed using the parameters of the currently used model as initial values ​​and a learning rate lower than that used during initial training. In one example, the incremental training learning rate is 1 / 5 of the initial training rate. The model is trained for 50 to 100 epochs. After incremental training, the prediction error of the positional offset prediction is calculated on the validation set. If the accuracy is better than the current model, the model is replaced with the new model; otherwise, the current model is retained. Incremental training with a low learning rate and fewer epochs can track slow changes in the environment and device status while avoiding large fluctuations in the model caused by occasional outliers in the new data.

[0065] Therefore, this method utilizes existing panel surface morphology measurement data before bonding, compressing the high-dimensional surface morphology data into a few warpage mode coefficients with clear physical meaning through Zernike polynomial mode decomposition. These coefficients are then combined with bonding process parameters to construct a predictive feature vector, which is input into an alignment offset prediction model with physical symmetry constraints. This model predicts the alignment offset that will occur during the rolling process on a panel-by-pane basis and applies reverse compensation in advance during the alignment stage. Compared to a fixed compensation method based on batch statistical mean, this scheme can adapt to individual differences in warpage between panels, improving alignment accuracy from the batch statistical level to the panel-by-pane prediction level. Physical symmetry constraints enhance the model's generalization ability and reduce prediction errors when training data is limited. The entire process reuses data from existing testing equipment and is implemented at the software level without changing the bonding mechanical structure and process flow.

[0066] Please see Figure 4 , Figure 4 This is a structural block diagram of an OLED polarizer attachment control system based on artificial intelligence, provided in an embodiment of this application. Figure 4 As shown, the system includes: The three-dimensional topography measurement module 401 is used to perform three-dimensional topography measurement on the surface of the OLED panel to be attached, and obtain the height distribution data of the panel surface. The mode decomposition module 402 is used to perform orthogonal polynomial mode decomposition on the panel surface height distribution data and extract multi-order warping mode coefficients that characterize the warping morphology of the panel. The offset prediction module 403 is used to combine the multi-order warping mode coefficients with the attachment process parameters to construct a prediction feature vector, and input the prediction feature vector into a pre-trained alignment offset prediction model to obtain the alignment offset prediction value of the panel. The alignment offset prediction value includes a translation component in the propulsion direction, a lateral translation component, and an in-plane rotation component. The alignment offset prediction model uses the sum of the prediction error term and the physical symmetry constraint term as the training loss function. The physical symmetry constraint term is calculated based on the sum of the model output after inverting the multi-order warping mode coefficients and the model output before inversion. The alignment compensation and attachment module 404 is used to determine the alignment target coordinates of the polarizer relative to the panel, shift the alignment target coordinates in the opposite direction of the alignment offset prediction value to obtain the compensated alignment target coordinates, and position the polarizer according to the compensated alignment target coordinates and perform roll pressing attachment.

[0067] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0068] Based on the same inventive concept, this application also provides an electronic device, the method corresponding to which can be the method in the foregoing embodiments, and its problem-solving principle is similar to that method. For example... Figure 5 As shown, Figure 5 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods and / or technical solutions of the foregoing embodiments of the present application.

[0069] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a processor, it performs the functions defined in the methods of this application.

[0070] Another embodiment of this application provides a storage medium storing computer program instructions thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application.

[0071] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. An OLED polarizer attachment control method based on artificial intelligence, characterized in that, include: Three-dimensional topography measurements were performed on the surface of the OLED panel to be attached to obtain the height distribution data of the panel surface; Orthogonal polynomial mode decomposition was performed on the panel surface height distribution data to extract multi-order warping mode coefficients that characterize the panel warping morphology; The multi-order warping mode coefficients are combined with the attachment process parameters to construct a prediction feature vector. The prediction feature vector is then input into a pre-trained alignment offset prediction model to obtain the alignment offset prediction value of the panel. The alignment offset prediction value includes a translation component in the propulsion direction, a lateral translation component, and an in-plane rotation component. The alignment offset prediction model uses the sum of the prediction error term and the physical symmetry constraint term as the training loss function. The physical symmetry constraint term is calculated based on the sum of the model output after inverting the multi-order warping mode coefficients and the model output before inverting them. Determine the alignment target coordinates of the polarizer relative to the panel, shift the alignment target coordinates in the opposite direction of the alignment offset prediction value to obtain the compensated alignment target coordinates; position the polarizer according to the compensated alignment target coordinates and perform roll pressing attachment.

2. The method according to claim 1, characterized in that, The orthogonal polynomial mode decomposition uses Zernike polynomial basis functions; the in-plane coordinates of the panel are normalized to the unit circle, with the geometric center of the effective attachment area of ​​the panel as the origin and the half-length of the panel diagonal as the normalization radius; the panel surface height distribution data is fitted to a linear superposition of multiple Zernike polynomials on the normalized coordinates, and the modal coefficients are solved by the least squares method. After removing the piston term that represents the overall height offset, the remaining coefficients constitute the multi-order warping mode coefficients.

3. The method according to claim 1, characterized in that, The prediction error term is the mean square value of the deviation between the prediction output of the alignment offset prediction model and the actual measured alignment deviation after panel attachment. The physical symmetry constraint term is calculated as follows: the components corresponding to the warp mode coefficient in the prediction feature vector are inverted, while the components corresponding to the attachment process parameters remain unchanged. The inverted feature vector is input into the alignment offset prediction model to obtain the flip prediction output. The flip prediction output is added component by component to the prediction output before inversion, and the mean square value is taken. The weight of the physical symmetry constraint term in the training loss function is determined by searching for the value with the optimal prediction accuracy on the validation set.

4. The method according to claim 1, characterized in that, The bonding process parameters include the dynamic viscosity of the current batch of optically transparent adhesive, the set pressure of the bonding roller, the set feed speed of the bonding roller, and the ambient temperature and relative humidity of the bonding station.

5. The method according to claim 1, characterized in that, When constructing the predicted feature vector, the features of each dimension after merging the multi-order warping mode coefficients and the attachment process parameters are standardized respectively; the standardization parameters are the mean and standard deviation of each feature in the training dataset.

6. The method according to claim 1, characterized in that, The alignment shift prediction model is a multilayer perceptron, which sequentially includes an input layer, multiple hidden layers, and an output layer, with each hidden layer connected in series. Each hidden layer applies a linear transformation to the output of the previous layer and then processes it through a nonlinear activation function. The input layer receives the predicted feature vector, and the output layer outputs the three components of the alignment shift prediction value.

7. The method according to claim 6, characterized in that, The multilayer perceptron contains three hidden layers: the first hidden layer contains 64 neurons, the second hidden layer contains 32 neurons, and the third hidden layer contains 16 neurons. The first and second hidden layers have dropout layers after the activation function, while the third hidden layer does not have a dropout layer. The output layer contains three neurons and uses linear output.

8. The method according to claim 1, characterized in that, The data used to train the alignment offset prediction model comes from production line records. Each record contains the multi-order warpage mode coefficients of the panel, the process parameters during bonding, and the actual alignment deviation measured by the automatic optical inspection process after bonding. The training uses the Adam optimizer, and the learning rate gradually decays from the initial value to the termination value according to the cosine annealing strategy. The training set and the validation set are divided proportionally, and the set of model parameters with the smallest prediction error on the validation set is saved.

9. The method according to claim 1, characterized in that, After the roll bonding is completed, the automatic optical inspection process measures the actual alignment deviation between the polarizer and the panel. The actual alignment deviation, along with the multi-order warpage mode coefficients of the corresponding panel and the bonding process parameters, is stored in the database as new training samples. When the number of new training samples reaches a preset number, the alignment offset prediction model is incrementally trained with the current model parameters as the initial value and a learning rate lower than that used in the initial training. The model is then replaced based on the accuracy of the validation set.

10. An OLED polarizer attachment control system based on artificial intelligence, characterized in that, include: The three-dimensional topography measurement module is used to perform three-dimensional topography measurement on the surface of the OLED panel to be attached, and obtain the height distribution data of the panel surface. The mode decomposition module is used to perform orthogonal polynomial mode decomposition on the panel surface height distribution data to extract multi-order warping mode coefficients that characterize the warping morphology of the panel. The offset prediction module is used to combine the multi-order warp mode coefficients with the attachment process parameters to construct a prediction feature vector. The prediction feature vector is then input into a pre-trained alignment offset prediction model to obtain the alignment offset prediction value of the panel. The alignment offset prediction value includes a translation component in the propulsion direction, a lateral translation component, and an in-plane rotation component. The alignment offset prediction model uses the sum of the prediction error term and the physical symmetry constraint term as the training loss function. The physical symmetry constraint term is calculated based on the sum of the model output after inverting the multi-order warp mode coefficients and the model output before inverting them. The alignment compensation and attachment module is used to determine the alignment target coordinates of the polarizer relative to the panel, shift the alignment target coordinates in the opposite direction of the alignment offset prediction value to obtain the compensated alignment target coordinates, and position the polarizer according to the compensated alignment target coordinates and perform roll pressing attachment.

Images