Performance detection method and system for liquid crystal display module

Abstract

The invention relates to the technical field of display screen detection, in particular to a liquid crystal display screen module performance detection method and system, and the method comprises the steps: collecting an image sequence of pure color pictures displayed by a display screen module at a full brightness level, and obtaining a screen display image of each pure color picture according to the image sequence; extracting a feature vector of the screen display image, performing feature analysis on the screen display image according to the feature vector, obtaining an abnormal point of each display pure color picture, and recording a spatial position and a deviation direction of the abnormal point; if at least one display pure color picture has the abnormal point, performing secondary judgment on the abnormal point by combining the spatial continuity and time stability characteristics of the adjacent frame of the abnormal point, and obtaining the abnormal type of the abnormal point; and generating an anomaly detection report of the display screen module. According to the method, the accuracy and robustness of detection are improved, misjudgment and missing detection can be effectively reduced, and the quality control level of the liquid crystal display screen module is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of display screen detection, and particularly to a method and system for detecting the performance of a liquid crystal display screen module. Background Art

[0002] With the rapid development of liquid crystal display technology, as the core component of a display device, the performance and quality of a liquid crystal display screen module directly affect the visual experience of end users. During the production and manufacturing process of the display screen module, due to the influence of various factors such as production processes, material properties, and environmental factors, display defects such as bright spots, dark spots, color deviation, and uneven brightness are likely to occur. These defects not only reduce the display effect but also may affect the service life and market competitiveness of the product. Traditional methods for detecting the performance of display screen modules mainly rely on manual visual inspection, which has problems such as low efficiency, strong subjectivity, and easy omission of inspections, and it is difficult to meet the requirements of modern large-scale production.

[0003] In recent years, automated detection technologies based on machine vision have gradually become the mainstream direction for detecting the performance of display screen modules. By collecting an image sequence of the display screen module showing a pure color picture at different brightness levels and using image processing algorithms to extract feature vectors, automatic identification and detection of display defects can be achieved. However, existing detection methods still have problems such as insufficient accuracy and high false positive rates when dealing with complex display defects, especially in the judgment of the spatial continuity and temporal stability characteristics of abnormal points, which still need to be further optimized. Therefore, developing a method for detecting the performance of a liquid crystal display screen module that can accurately identify and classify display defects while taking into account detection efficiency and accuracy has important practical significance for improving product quality and reducing production costs. Summary of the Invention

[0004] The present invention overcomes the deficiencies of the prior art and provides a method and system for detecting the performance of a liquid crystal display screen module.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows: In a first aspect of the present invention, a method for detecting the performance of a liquid crystal display screen module is disclosed, including the following steps: Collect an image sequence of the display screen module showing a pure color picture at full brightness levels, and obtain the screen display images of each pure color picture according to the image sequence; Extract the feature vectors of the screen display images, perform feature analysis on the screen display images according to the feature vectors, obtain the abnormal points of each pure color picture shown, and record the spatial positions and deviation directions of the abnormal points; If there are no abnormal points in each pure color picture shown, determine that the display screen module is a qualified product in terms of display performance; If there is at least one abnormal point in the display of a solid - color screen, the abnormal point is re - judged by combining the spatial continuity and temporal stability characteristics of the adjacent frames of the abnormal point to obtain the abnormal type of the abnormal point; Generate an abnormal detection report for the display module according to the spatial position, deviation direction and abnormal type of the abnormal point.

[0006] Preferably, extract the feature vectors of the screen display image, perform feature analysis on the screen display image according to the feature vectors, obtain the abnormal points of each solid - color display screen, and record the spatial position and deviation direction of the abnormal points. Specifically: Divide the screen display image into local units according to 3×3 pixel regions, extract the feature vectors of each central pixel point, and construct a normalized feature matrix; Pre - train a Gaussian mixture model using the expectation - maximization algorithm based on a preset sample set to obtain the trained Gaussian mixture model; Set a dynamic clustering center number threshold, import the feature vectors of each central pixel point into the trained Gaussian mixture model respectively, and obtain the matching degree between the current feature vector and each Gaussian component in real - time; If the matching degree is less than the preset threshold, trigger the component splitting mechanism to generate sub - components to adapt to the evolution of the feature space until the matching degree is greater than the preset threshold; Calculate the posterior probability that the feature vector of each central pixel point belongs to each Gaussian component, take the component corresponding to the maximum posterior probability as the belonging cluster, and calculate the Mahalanobis distance between the feature vector and the distance center according to the mean vector and covariance matrix of the clustering center; If the Mahalanobis distance between the feature vector and the distance center is greater than the abnormal determination value dynamically adjusted by the component to which it belongs, mark it as an abnormal point of the current screen, and record its spatial position and deviation direction.

[0007] Preferably, if the matching degree is less than the preset threshold, trigger the component splitting mechanism to generate sub - components to adapt to the evolution of the feature space until the matching degree is greater than the preset threshold. Specifically: When the matching degree of the feature vectors continuously received by a certain component is lower than the preset threshold, analyze the eigenvalue distribution of the component covariance matrix. If the ratio of the maximum eigenvalue to the minimum eigenvalue exceeds the anisotropy threshold, it is determined as an abnormal phenomenon of feature space stretching; And calculate the distribution dispersion of the component coverage area in the HSV color space. When the distribution dispersion is lower than the color aggregation standard, it is determined as an over - fitting phenomenon; If it is determined that there is an abnormal phenomenon in the feature space stretching, an adaptive kernel function is introduced to determine the optimal splitting point, the mean vectors of the two sub-components are initialized, and the covariance matrix of the original component is orthogonally decomposed in the direction of the maximum variance to generate a sub-component covariance matrix with direction adaptability, ensuring that the split sub-components can effectively capture the anisotropic distribution of the feature space; If it is determined that there is an overfitting phenomenon, the historical data of the atomic component is reallocated to the newly generated sub-components according to the Mahalanobis distance. At the same time, a momentum factor is introduced to smooth the parameter update process, and the online expectation maximization algorithm is used to iteratively optimize the weights, means, and covariance parameters of the sub-components. The optimization is terminated when the change rate of the logarithmic likelihood function of the sub-components is lower than the convergence threshold.

[0008] Preferably, if there is at least one abnormal point in the display of a solid color screen, the abnormal point is re-determined by combining the spatial continuity and temporal stability characteristics of the adjacent frames of the abnormal point to obtain the abnormal type of the abnormal point. Specifically: For the coordinates of the abnormal point, a 3×3 pixel region block corresponding to its position in different solid color screens is extracted and stacked along the time dimension to form spatio-temporal cube data; A three-dimensional convolutional neural network is used to extract the spatial texture features of the pixel region block. At the same time, a gated recurrent unit is used to capture the sequence pattern of the luminance value changing with time in combination with the spatio-temporal cube data, and the spatial texture features and the sequence pattern of the luminance value changing with time are spliced to form a cross-modal fusion descriptor; Taking the abnormal point as the central node, a fully connected edge is established with the cross-modal fusion descriptor, and the edge weight distribution entropy value of each edge is calculated; When the edge weight distribution entropy value is lower than the preset entropy value, the abnormal point is determined as an isolated abnormal point; when the edge weight distribution entropy value is greater than the preset entropy value, the abnormal point is determined as a linkage abnormal point.

[0009] Preferably, taking the abnormal point as the central node, a fully connected edge is established with the cross-modal fusion descriptor, and the edge weight distribution entropy value of each edge is calculated. Specifically: The abnormal point is set as the central node of the fully connected graph, and the adaptive neighborhood aggregation mechanism of the graph attention network is used to project the feature space of each connected cross-modal descriptor and obtain the projection coordinates; According to the projection coordinates, the Euclidean distance between the central node and the projection node is calculated, and the Euclidean distance is weighted to obtain the dynamic association weight between the central node and the projection node. A dynamic association weight matrix is constructed according to the dynamic association weight; Low-rank decomposition optimization is applied to the dynamic association weight matrix to obtain an optimized dynamic association weight matrix; Use a variational graph autoencoder to perform latent variable sampling on the optimized dynamic correlation weight matrix, fuse the local gradient information extracted by a 3D convolutional network through a message passing mechanism, and construct a fully connected topological graph with spatio-temporal correlation; In the fully connected topological graph, traverse all edges, obtain the path length values of each edge, and perform a ratio process on the path length values of each edge with a preset value to obtain the edge weight distribution entropy values of each edge.

[0010] Preferably, perform low-rank decomposition optimization on the dynamic correlation weight matrix to obtain an optimized dynamic correlation weight matrix, specifically: Decompose the original dynamic correlation weight matrix into a low-rank component and a sparse noise component. For the amplitude distribution characteristics of the sparse noise component, introduce a Huffman weighting function to impose non-linear suppression on outliers; when the low-rank component of an element at a certain position is higher than the confidence interval, activate a weight decay mechanism based on an adaptive momentum factor for correction; generate a low-rank basis matrix according to the processed low-rank component and sparse noise component; Adopt an incremental non-negative matrix factorization algorithm to project the low-rank basis matrix into a non-negative manifold space. If there are negative components in the decomposed basis vectors, trigger an orthogonal correction layer based on KL divergence, and iteratively optimize the coupling relationship between the basis matrix and the coefficient matrix through the alternating least squares method; Construct a graph Laplacian operator of the dynamic correlation weight matrix. When the spectral energy distribution of the low-rank component does not match the graph structure prior, inject a multi-scale spectral regularization term generated by a graph convolutional network, and use an attention mechanism to dynamically adjust the retention ratio of low-frequency components and high-frequency components; When the cosine similarity of the basis matrices in adjacent iterations reaches the convergence threshold, trigger an early stopping mechanism to output the optimized dynamic correlation weight matrix.

[0011] The eigenvectors include the RGB three-channel brightness values, the average absolute value of the gradient differences from adjacent 8 pixels, and the saturation in the HSV color space.

[0012] The pure color screens include pure black, pure white, pure red, pure green, and pure blue.

[0013] The second aspect of the present invention discloses a performance detection system for a liquid crystal display module. The performance detection of the liquid crystal display module includes a memory and a processor. A performance detection method program for the liquid crystal display module is stored in the memory. When the performance detection method program for the liquid crystal display module is executed by the processor, the steps of any one of the performance detection methods for the liquid crystal display module are implemented.

[0014] The third aspect of the present invention discloses a computer-readable storage medium. When a program for the performance detection method of the liquid crystal display module is executed by a processor, the steps of the performance detection method of any of the above-mentioned liquid crystal display modules are realized.

[0015] The present invention solves the technical defects existing in the background art and has the following beneficial effects: collecting an image sequence of a pure color picture displayed at all brightness levels of the display module, and obtaining a screen display image of each pure color picture according to the image sequence; extracting the feature vectors of the screen display image, performing feature analysis on the screen display image according to the feature vectors, obtaining the abnormal points of each displayed pure color picture, and recording the spatial positions and deviation directions of the abnormal points; if there are no abnormal points in each displayed pure color picture, determining the display module as a qualified product in terms of display performance; if there is at least one displayed pure color picture with abnormal points, performing a secondary determination on the abnormal points by combining the spatial continuity and temporal stability characteristics of the adjacent frames of the abnormal points to obtain the abnormal types of the abnormal points; generating an abnormal detection report of the display module according to the spatial positions, deviation directions and abnormal types of the abnormal points. This not only improves the accuracy and robustness of the detection, but also effectively reduces misjudgment and missed detection, and significantly improves the quality control level of the liquid crystal display module. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0017] Figure 1 It is the overall method flow chart of a performance detection method for a liquid crystal display module; Figure 2 It is the system block diagram of a performance detection system for a liquid crystal display module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention will be further described in detail below with reference to the drawings and specific embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0019] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0020] Such asFigure 1 As shown in Figure 1 , the first aspect of the present invention discloses a method for detecting the performance of a liquid crystal display module, including the following steps: S102. Collect an image sequence of the display module displaying a solid-color screen at all brightness levels, and obtain the screen display images of each solid-color screen according to the image sequence; S104. Extract the feature vectors of the screen display images, perform feature analysis on the screen display images according to the feature vectors, obtain the abnormal points of each displayed solid-color screen, and record the spatial positions and deviation directions of the abnormal points; S106. If there are no abnormal points in each displayed solid-color screen, determine the display module as a qualified product for display performance; S108. If there is at least one displayed solid-color screen with abnormal points, perform a secondary determination on the abnormal points by combining the spatial continuity and temporal stability characteristics of the adjacent frames of the abnormal points to obtain the abnormal types of the abnormal points; S110. Generate an abnormal detection report of the display module according to the spatial positions, deviation directions, and abnormal types of the abnormal points.

[0021] By collecting and analyzing the solid-color screen images of the display screen at different brightness levels, and using technologies such as Gaussian mixture models and three-dimensional convolutional neural networks, the present invention can accurately identify and classify abnormal points, distinguish isolated abnormal points and linked abnormal points, and generate a detailed abnormal detection report. It not only improves the accuracy and robustness of detection, but also effectively reduces misjudgment and missed detection, and significantly improves the quality control level of liquid crystal display modules.

[0022] Preferably, extracting the feature vectors of the screen display images, performing feature analysis on the screen display images according to the feature vectors, obtaining the abnormal points of each displayed solid-color screen, and recording the spatial positions and deviation directions of the abnormal points specifically includes: Divide the screen display image into local units according to 3×3 pixel regions, extract the feature vectors of each central pixel point, and construct a normalized feature matrix; Pre-train a Gaussian mixture model using the expectation maximization algorithm based on a preset sample set to obtain a trained Gaussian mixture model; It should be noted that, first, initialize the parameters of the Gaussian mixture model, including the weights, mean vectors, and covariance matrices of each Gaussian component; second, perform iterative training using the preset sample set, alternately execute the expectation step and the maximization step through the expectation maximization algorithm, and gradually optimize the model parameters so that the model can better fit the sample data distribution; finally, after several iterations, when the model parameters converge or reach a predetermined number of iterations, obtain the trained Gaussian mixture model, which can be used for subsequent feature vector classification and abnormal point detection tasks.

[0023] Set the threshold of the number of dynamic cluster centers, import the feature vectors of each central pixel into the trained Gaussian mixture model, and obtain the matching degree between the current feature vector and each Gaussian component in real time; It should be noted that after setting the threshold of the number of dynamic cluster centers, the feature vectors of each central pixel point are imported into the trained Gaussian mixture model one by one, and the matching degree between the current feature vector and each Gaussian component is obtained by calculating the probability density function value of the feature vector and each Gaussian component in real time.

[0024] If the matching degree is less than a preset threshold, a component splitting mechanism is triggered to generate sub-components to adapt to the feature space evolution until the matching degree is greater than the preset threshold; Calculate the posterior probability that the eigenvector of each central pixel belongs to each Gaussian component, take the component corresponding to the maximum posterior probability as the belonging cluster, and calculate the Mahalanobis distance between the eigenvector and the distance center according to the mean vector and covariance matrix of the cluster center; If the Mahalanobis distance between the feature vector and the distance center is greater than the abnormality judgment value dynamically adjusted by the corresponding component, it is marked as an abnormal point of the current picture, and its spatial position and deviation direction are recorded.

[0025] It should be noted that high-precision detection of abnormal points on the display screen is achieved by combining the dynamic adaptive Gaussian mixture model with the Mahalanobis distance algorithm. First, a normalized feature matrix is ​​constructed to retain the microscopic optical characteristics and eliminate the interference of brightness fluctuations; second, a dynamic clustering splitting mechanism is introduced to automatically generate sub-components when the matching degree between the feature vector and the preset Gaussian component is insufficient, effectively solving the problem that the traditional GMM model is insufficient in recognizing new abnormal patterns; finally, the judgment threshold is dynamically adjusted in the non-uniform chromaticity space through the Mahalanobis distance calculation weighted by the covariance matrix. This method can distinguish the abnormal morphological differences caused by process defects (such as local polarizer offset) and material aging (such as ITO line oxidation) in real time, and the deviation direction recording function can trace the cause of the abnormality (such as lateral deviation is mostly caused by abnormal driving signals), and achieve sub-pixel level (0.1mm accuracy) abnormal positioning.

[0026] Preferably, if the matching degree is less than a preset threshold, a component splitting mechanism is triggered to generate sub-components to adapt to the feature space evolution until the matching degree is greater than the preset threshold, specifically: When the matching degree of the eigenvectors continuously received by a component is lower than the preset threshold, the eigenvalue distribution of the component covariance matrix is ​​analyzed. If the ratio of the maximum eigenvalue to the minimum eigenvalue exceeds the anisotropy threshold, it is determined to be an abnormal phenomenon of feature space stretching; And calculate the distribution dispersion of the component coverage area in the HSV color space. When the distribution dispersion is lower than the color aggregation standard, it is judged as an overfitting phenomenon; If it is determined that there is an abnormal phenomenon of feature space stretching, an adaptive kernel function is introduced to determine the optimal splitting point, and the mean vectors of the two sub-components are initialized. Then, the covariance matrix of the original component is orthogonally decomposed in the direction of the maximum variance to generate a sub-component covariance matrix with direction adaptability, ensuring that the split sub-components can effectively capture the anisotropic distribution of the feature space. If it is determined that there is an overfitting phenomenon, the historical data of the atomic component is reallocated to the newly generated sub-components according to the Mahalanobis distance. At the same time, a momentum factor is introduced to smooth the parameter update process, and the online expectation-maximization algorithm is used to iteratively optimize the weights, means, and covariance parameters of the sub-components. The optimization is terminated when the change rate of the log-likelihood function of the sub-components is lower than the convergence threshold.

[0027] It should be noted that through the intelligent component splitting mechanism, it can effectively adapt to the changes in the feature space. Whether it is the abnormal stretching of the feature space or the overfitting phenomenon, it can ensure that the model can more accurately capture and classify the complex feature space through the adaptive generation of sub-components and parameter optimization, thus significantly improving the accuracy and robustness of feature matching and anomaly detection.

[0028] Preferably, if there is at least one abnormal point in the pure color screen display, the abnormal point is rejudged by combining the spatial continuity and temporal stability characteristics of the adjacent frames of the abnormal point to obtain the abnormal type of the abnormal point. Specifically: For the coordinates of the abnormal point, a 3×3 pixel region block corresponding to its position in different pure color screens is extracted and stacked along the time dimension to form spatio-temporal cube data. A three-dimensional convolutional neural network is used to extract the spatial texture features of the pixel region block. At the same time, through a gated recurrent unit and combined with the spatio-temporal cube data, the sequence pattern of the brightness value changing with time is captured, and the spatial texture features and the sequence pattern of the brightness value changing with time are spliced to form a cross-modal fusion descriptor. It should be noted that first, a 15×15 pixel region block is intercepted centered on each pixel to be measured, and 20 consecutive frames of brightness data are stacked along the time axis to form a 128×128×20 spatio-temporal cube and input into a three-dimensional convolutional neural network. Its hierarchical structure includes 3 3D convolutional layers (kernel sizes are 5×5×3, 3×3×2, 3×3×1) and corresponding max-pooling layers, using the ReLU activation function and the Batch Normalization layer, and finally a 256-dimensional spatial texture feature vector is extracted. At the same time, the RGB three-channel brightness values of the same region block are input into a bidirectional gated recurrent unit (BiGRU) according to the time series, with 128 hidden units set, and the brightness gradient change features of the key frames are weighted through the temporal attention mechanism to output a 128-dimensional time dynamic encoding. Then, a cross-modal alignment module is used to perform adaptive weight allocation on the two feature vectors, and after mapping to the same dimension through a fully connected layer, a splicing operation is performed to form a 384-dimensional fusion descriptor.

[0029] Taking the anomaly point as the central node, establish fully connected edges with the cross-modal fusion descriptors, and calculate the entropy value of the edge weight distribution for each edge; When the entropy value of the edge weight distribution is lower than the preset entropy value, the anomaly point is determined as an isolated anomaly point; when the entropy value of the edge weight distribution is greater than the preset entropy value, the anomaly point is determined as a linked anomaly point.

[0030] It should be noted that an isolated anomaly point refers to an anomaly point that has no obvious association with other anomaly points in terms of time and space. Such anomaly points appear in a single frame or have no continuity in time, manifested as the entropy value of the edge weight distribution being lower than the preset entropy value, such as a single-frame bright or dark point. A linked anomaly point, on the other hand, refers to an anomaly point that has an obvious association with other anomaly points in terms of time and space. Such anomaly points show a certain continuity and pattern in multiple frames, manifested as the entropy value of the edge weight distribution being higher than the preset entropy value, such as multi-frame coherent color deviation or uneven brightness.

[0031] In summary, through the spatio-temporal cube data and cross-modal fusion descriptors, the types of anomaly points can be accurately identified and classified, thereby improving the accuracy and reliability of anomaly detection, which helps to more comprehensively understand and process the display defects in the liquid crystal display screen according to the detection report and optimize the production process of the liquid crystal display screen module.

[0032] Preferably, taking the anomaly point as the central node, establish fully connected edges with the cross-modal fusion descriptors, and calculate the entropy value of the edge weight distribution for each edge, specifically: Set the anomaly point as the central node of the fully connected graph, adopt the adaptive neighborhood aggregation mechanism of the graph attention network, perform feature space projection on each connected cross-modal descriptor, and obtain the projection coordinates; Calculate the Euclidean distance between the central node and the projection node according to the projection coordinates, perform weighted processing on the Euclidean distance to obtain the dynamic association weight between the central node and the projection node, and construct a dynamic association weight matrix according to the dynamic association weight; Apply low-rank decomposition optimization to the dynamic association weight matrix to obtain an optimized dynamic association weight matrix; Use a variational graph autoencoder to perform latent variable sampling on the optimized dynamic association weight matrix, fuse the local gradient information extracted by the three-dimensional convolutional network through the message passing mechanism, and construct a fully connected topological graph with spatio-temporal correlation; It should be noted that the optimized dynamic association weight matrix is ​​input into the encoder module of the variational graph autoencoder, and the global topological features are aggregated through a multi-layer graph attention network to generate a parameterized distribution of the latent space mean and variance, and the reparameterization technique is used for latent variable sampling; at the same time, a three-dimensional convolutional network is used to slide along the spatiotemporal dimension to extract local gradient amplitude features, and the receptive field is expanded by the dilated convolution to capture long-range dependencies; then the spatiotemporal fusion message passing layer is designed to perform bilinear interaction between the latent variables and the local gradient tensor: the convolutional features are embedded and aligned with the graph nodes in the spatial dimension, and the state transfer chain is established in the time dimension through the gated recurrent unit; then a dynamic topology generator is constructed, with the latent variables as the initialization seeds, and the heat diffusion equation is used to simulate the spatiotemporal propagation process of the correlation strength between nodes, and the feature vector after message aggregation is projected into the cosine similarity space to generate a fully connected edge weight matrix; finally, a spectral normalization constraint is imposed to ensure the semi-positive definiteness of the Laplacian matrix of the generated graph, and the motion-related edges and static structure edges are separated by a decoupled spatiotemporal mask matrix to form a fully connected topological graph with both local gradient sensitivity and global latent variable continuity.

[0033] In the fully connected topology graph, all edges are traversed to obtain the path length value of each edge, and the path length value of each edge is compared with the preset value to obtain the edge weight distribution entropy value of each edge.

[0034] It should be noted that this method achieves in-depth analysis of anomaly propagation characteristics across time and space dimensions by constructing a fully connected topological relationship network of anomalies. Specifically, the spatial coupling relationship between anomalies and surrounding features is dynamically captured through the adaptive neighborhood aggregation of the graph attention network, and the redundant information of the weight matrix under noise interference is eliminated through low-rank decomposition optimization; combined with the hidden variable sampling mechanism of the variational graph autoencoder, the probability distribution of the anomaly evolution path is reconstructed in the latent space; after the message passing mechanism fuses the local gradient information, the weight distribution entropy value of each edge in the fully connected graph can not only characterize the radiation intensity of the current anomaly point to the surrounding area, but also reflect the transmission law of the anomaly pattern in the historical time series.

[0035] In summary, through the graph attention network and variational graph autoencoder, the edge weight distribution entropy of the anomaly points can be accurately calculated, so as to effectively distinguish isolated anomalies from linked anomalies and improve the accuracy and robustness of anomaly detection.

[0036] Preferably, low-rank decomposition optimization is applied to the dynamic association weight matrix to obtain an optimized dynamic association weight matrix, specifically: The original dynamic correlation weight matrix is decomposed into a low-rank component and a sparse noise component. Aiming at the amplitude distribution characteristics of the sparse noise component, a Huffman weighting function is introduced to impose non-linear suppression on outliers; when the low-rank component of an element at a certain position is higher than the confidence interval, the weight decay mechanism based on the adaptive momentum factor is activated for correction; a low-rank basis matrix is generated according to the processed low-rank component and sparse noise component. It should be noted that the robust principal component analysis framework is used to decompose the original dynamic correlation weight matrix into a low-rank component (representing the global correlation pattern) and a sparse noise component (capturing local outliers). Aiming at the characteristic that the amplitudes of sparse noise elements follow a heavy-tailed distribution, a weighted mapping operator based on the Huber function is designed, and non-linear soft threshold suppression is imposed on outliers exceeding the statistical confidence interval (such as three times the median absolute deviation) through a continuously differentiable threshold piecewise function; at the same time, a dynamic sliding window is constructed to monitor the amplitude fluctuations of each element in the low-rank component. When the cumulative statistic of a specific element is detected to exceed the preset confidence interval (estimated by Monte Carlo sampling simulation), the weight decay mechanism based on the adaptive momentum factor (fusing the current gradient direction and the historical update inertia) is triggered, and a dynamic contraction force pointing to the origin of the parameter space is imposed during the gradient descent process for correction; finally, the regularized low-rank component and the denoised sparse component are convexly combined by the alternating direction method of multipliers, and the explicit expression of the low-rank basis matrix is solved by the iteratively reweighted least squares method to ensure that it maximally retains the topological characteristics of the corrected components while satisfying the row and column orthogonality constraints.

[0037] The incremental non-negative matrix factorization algorithm is adopted to project the low-rank basis matrix into the non-negative manifold space. If there are negative components in the decomposed basis vectors, the orthogonal correction layer based on the KL divergence is triggered, and the coupling relationship between the basis matrix and the coefficient matrix is iteratively optimized by the alternating least squares method. It should be noted that the low-rank basis matrix is dynamically projected into the non-negative manifold space through batch-by-batch sampling: in each iteration, the non-negative basis matrix and coefficient matrix are initialized, and the projection gradient method is used to enforce the non-negativity of their components; when negative components are detected in the decomposed basis vectors, the orthogonal correction layer based on KL divergence is activated, and a bi-objective function that combines non-negative constraints and orthogonality is constructed (where KL divergence measures the distribution difference between the decomposed matrix and the original data, and the orthogonality term penalizes the correlation between basis vectors through the Frobenius norm), and at the same time, a relaxation variable is introduced to transform the non-convex problem into a solvable form; then the alternating least squares method is used for coupled optimization - when the basis matrix is fixed, the closed-form solution of the coefficient matrix is solved by the quasi-Newton method, and when the coefficient matrix is fixed, the learning rate is adaptively adjusted by combining the Barzilai-Borwein step size to update the basis matrix, and a non-negative threshold truncation is applied after each update; in addition, a residual feedback mechanism is designed to re-inject the corrected negative component residuals into the next round of incremental learning, and the momentum acceleration strategy is used to balance the convergence rates of the KL divergence loss and the orthogonality constraint, and finally, the progressive alignment of the basis matrix and the coefficient matrix in the non-negative manifold space is achieved.

[0038] Construct the graph Laplacian of the dynamic correlation weight matrix. When the spectral energy distribution of the low-rank component does not match the graph structure prior, inject the multi-scale spectral regularization term generated by the graph convolutional network, and use the attention mechanism to dynamically adjust the retention ratio of the low-frequency component and the high-frequency component. It should be noted that the graph Laplacian is constructed based on the dynamic correlation weight matrix, and the spectral energy distribution of the low-rank component is analyzed through Fourier transform. When a spectral dimension shift is detected between it and the topological constraints of the graph structure prior (such as node degree distribution, community structure), the graph convolutional network is activated to extract multi-scale subgraph features, and a spectral regularization term that includes local smoothness and global structure consistency is generated; then a frequency-domain gating unit is constructed using the multi-head attention mechanism to dynamically calculate the significance weights of the low-frequency component (reflecting the graph structure stability) and the high-frequency component (characterizing node heterogeneity), and the contribution ratio of the two in the regularization term is adjusted through a learnable scaling factor; finally, the spectral regularization term is injected into the graph constraint module of the objective function, and an adaptive frequency threshold is used to perform band-pass filtering on the high-oscillation components that deviate from the prior distribution, and at the same time, the graph convolutional kernel parameters are updated through backpropagation to correct the spectral domain error.

[0039] When the cosine similarity of the basis matrix between adjacent iterations reaches the convergence threshold, trigger the early stopping mechanism to output the optimized dynamic correlation weight matrix.

[0040] In summary, through low-rank decomposition and various optimization strategies, noise can be effectively removed and key features can be retained, and an optimized dynamic correlation weight matrix can be generated, thereby improving the accuracy and robustness of anomaly detection.

[0041] Among them, the above-mentioned feature vector includes the brightness values of the RGB three channels, the average absolute value of the gradient difference from the adjacent 8 pixels, and the saturation in the HSV color space; the solid-color screen includes pure black, pure white, pure red, pure green, and pure blue.

[0042] In this embodiment, the performance detection method may further include the following steps: If there are at least one abnormal point in the displayed solid-color screen, obtain the corresponding abnormal detection report, and obtain the spatial position distribution characteristics and abnormal types of the abnormal points according to the corresponding abnormal detection report; Based on the spatial position distribution characteristics and abnormal types of the abnormal points, through a pre-trained multi-layer knowledge graph model, map the abnormal mode to the potential equipment failure nodes in the production process flow of the display screen module, and establish a dynamic association network between the abnormal type and the coating, etching, and lamination process parameters; Collect the sensor time-series data stream of the relevant production equipment in real time, use the bidirectional LSTM network to extract the time-series feature vector of the equipment operation parameters, combine the spatial vector coding of the deviation direction in the abnormal detection report, construct a joint embedding space of the equipment state-abnormal features, and calculate the correlation weight matrix of the equipment parameters and abnormal features through the attention mechanism; When the abnormal correlation degree of the equipment parameters in the joint embedding space exceeds the dynamic threshold, trigger the reverse traceability analysis of the equipment parameters, and use the causal inference model to decompose the abnormal parameter chain into a directed acyclic graph to identify the core equipment parameter deviation items that cause the abnormality; For the identified core deviation items, construct a parameter optimization strategy generator based on the reinforcement learning framework, simulate the impact of parameter adjustment on the abnormal features through the Markov decision process, and select a parameter optimization scheme that can minimize the probability of abnormal points; Real-time feedback the optimized equipment parameters to the production control terminal, and update the equipment-abnormal association relationship in the knowledge graph to form a closed-loop optimization system; If there are still associated abnormalities at the same equipment node within consecutive optimization cycles, start the degradation prediction model of the equipment components, and predict the remaining service life and generate preventive maintenance instructions in combination with the Wiener process.

[0043] It should be noted that the method of optimizing the state of production equipment through the anomaly detection report aims to improve the manufacturing quality and efficiency of liquid crystal display module. Specifically, first, the location and type of anomaly points are identified through the anomaly detection report, and then the pre-trained knowledge graph model is used to associate these anomaly patterns with potential equipment failure nodes in the production process. Next, the sensor data of the equipment is collected in real time, and the bidirectional LSTM network and attention mechanism are used to analyze the association between the equipment parameters and anomaly features. When the anomaly correlation degree is found to exceed the threshold, the causal inference model is used to identify the core equipment parameter deviation items that cause the anomaly, and an optimization strategy is generated through the reinforcement learning framework. Finally, the optimized parameters are fed back to the production control terminal in real time, and the equipment-anomaly association relationship is continuously updated through the knowledge graph to form a closed-loop optimization system. If there are still problems with the same equipment node after multiple optimizations, the degradation prediction model of the equipment components is started to predict their remaining service life and generate preventive maintenance instructions. It can achieve the accurate positioning and analysis of anomaly points in the production process of liquid crystal display module, optimize the production equipment parameters by intelligent means, thus significantly improving the product quality and production efficiency. At the same time, it can also predict the degradation of equipment components, take preventive maintenance measures in advance, and extend the equipment life.

[0044] As Figure 2 shown, in the second aspect of the present invention, a performance detection system 6 for a liquid crystal display module is disclosed. The performance detection of the liquid crystal display module includes a memory 41 and a processor 52. A performance detection method program for the liquid crystal display module is stored in the memory 41. When the performance detection method program for the liquid crystal display module is executed by the processor 52, the steps of the performance detection method for any one of the liquid crystal display modules are realized.

[0045] In the third aspect of the present invention, a computer-readable storage medium is disclosed. When the performance detection method program for the liquid crystal display module stored in the computer-readable storage medium is executed by a processor, the steps of the performance detection method for any one of the liquid crystal display modules are realized.

[0046] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed with each other can be through some interfaces. The indirect coupling or communication connection of the devices or units can be electrical, mechanical, or other forms.

[0047] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0048] In addition, in each embodiment of the present invention, all the functional units may be integrated into one processing unit, or each unit may be separately regarded as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware, or in the form of a combination of hardware and software functional units.

[0049] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including the above method embodiments; and the foregoing storage medium includes: removable storage devices, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks or optical disks and other various media that can store program codes.

[0050] Alternatively, if the above-mentioned integrated units of the present invention are implemented in the form of software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present invention essentially or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods of the various embodiments of the present invention. And the foregoing storage medium includes: removable storage devices, ROM, RAM, magnetic disks or optical disks and other various media that can store program codes.

[0051] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, and all of them should be covered by the protection scope of the present invention.

Claims

1. A performance detection method for a liquid crystal display module, characterized in that: The following steps are involved: Collecting an image sequence of a pure color screen displayed at full brightness level of the display screen module, and acquiring a screen display image of each pure color screen according to the image sequence; Extracting feature vectors of screen display images, performing feature analysis on the screen display images according to the feature vectors, obtaining abnormal points of each pure color screen, and recording the spatial positions and deviation directions of the abnormal points; If there are no abnormal points in each display pure color screen, the display screen module is judged as a qualified product in display performance; If there is at least one abnormal point in the pure color display screen, the abnormal point is re-determined based on the spatial continuity and temporal stability characteristics of the adjacent frames of the abnormal point to obtain the abnormal type of the abnormal point; An abnormality detection report of the display screen module is generated according to the spatial position, deviation direction and abnormality type of the abnormal point.

2. A method for detecting the performance of a liquid crystal display module according to claim 1, characterized in that: Extract the feature vector of the screen display image, perform feature analysis on the screen display image according to the feature vector, obtain the abnormal points of each pure color screen, and record the spatial position and deviation direction of the abnormal points, specifically: Divide the screen display image into local units according to 3×3 pixel areas, extract the feature vector of each central pixel point, and construct a normalized feature matrix; Pre-training the Gaussian mixture model using the expectation maximization algorithm based on a preset sample set to obtain the trained Gaussian mixture model; Set the threshold of the number of dynamic cluster centers, import the feature vectors of each central pixel into the trained Gaussian mixture model, and obtain the matching degree between the current feature vector and each Gaussian component in real time; If the matching degree is less than a preset threshold, a component splitting mechanism is triggered to generate sub-components to adapt to the feature space evolution until the matching degree is greater than the preset threshold; Calculate the posterior probability that the eigenvector of each central pixel belongs to each Gaussian component, take the component corresponding to the maximum posterior probability as the belonging cluster, and calculate the Mahalanobis distance between the eigenvector and the distance center according to the mean vector and covariance matrix of the cluster center; If the Mahalanobis distance between the feature vector and the distance center is greater than the abnormality judgment value dynamically adjusted by the corresponding component, it is marked as an abnormal point of the current picture, and its spatial position and deviation direction are recorded.

3. A method for detecting the performance of a liquid crystal display module according to claim 2, characterized in that: If the matching degree is less than the preset threshold, the component splitting mechanism is triggered to generate sub-components to adapt to the feature space evolution until the matching degree is greater than the preset threshold, specifically: When the matching degree of the eigenvectors continuously received by a certain component is lower than the preset threshold, the eigenvalue distribution of the component covariance matrix is ​​analyzed. If the ratio of the maximum eigenvalue to the minimum eigenvalue exceeds the anisotropy threshold, it is determined to be an abnormal phenomenon of feature space stretching; And calculate the distribution dispersion of the component coverage area in the HSV color space. When the distribution dispersion is lower than the color aggregation standard, it is judged as an overfitting phenomenon; If it is determined to be an abnormal phenomenon of feature space stretching, an adaptive kernel function is introduced to determine the optimal splitting point, and the mean vectors of the two sub-components are initialized. The covariance matrix of the original component is orthogonally decomposed in the direction of maximum variance to generate a sub-component covariance matrix with directional adaptability, ensuring that the split sub-components can effectively capture the anisotropic distribution of the feature space. If it is determined to be an overfitting phenomenon, the historical data of the atomic component will be reallocated to the newly generated sub-component according to the Mahalanobis distance. At the same time, the momentum factor smoothing parameter update process is introduced, and the weights, means and covariance parameters of the sub-components are iteratively optimized using the online expectation maximization algorithm. The optimization is terminated when the rate of change of the log-likelihood function of the sub-component is lower than the convergence threshold.

4. The performance detection method of a liquid crystal display module according to claim 1, characterized in that: If there is at least one abnormal point in the pure color display screen, the abnormal point is judged again based on the spatial continuity and temporal stability characteristics of the adjacent frames of the abnormal point to obtain the abnormal type of the abnormal point, which is specifically: For the coordinates of the abnormal point, extract the 3×3 pixel area blocks at the corresponding positions in different pure color images, and stack them along the time dimension to form space-time cube data; A three-dimensional convolutional neural network is used to extract spatial texture features of the pixel area block, and a gated recurrent unit is used to capture a sequence pattern of brightness values ​​changing over time in combination with the space-time cube data, and the spatial texture features and the sequence pattern of brightness values ​​changing over time are spliced ​​to form a cross-modal fusion descriptor; Taking the outlier point as the central node, establishing a fully connected edge with the cross-modal fusion descriptor, and calculating the edge weight distribution entropy value of each edge; When the edge weight distribution entropy value is lower than the preset entropy value, the outlier is determined as an isolated outlier; when the edge weight distribution entropy value is greater than the preset entropy value, the outlier is determined as a linkage outlier.

5. A method for detecting the performance of a liquid crystal display module according to claim 4, characterized in that: Taking the outlier point as the central node, a fully connected edge is established with the cross-modal fusion descriptor, and the edge weight distribution entropy value of each edge is calculated, specifically: The outliers are set as the central nodes of the fully connected graph, and the adaptive neighborhood aggregation mechanism of the graph attention network is used to project each connected cross-modal descriptor into feature space and obtain the projection coordinates. Calculate the Euclidean distance between the central node and the projection node according to the projection coordinates, perform weighted processing on the Euclidean distance to obtain a dynamic association weight between the central node and the projection node, and construct a dynamic association weight matrix according to the dynamic association weight; Applying low-rank decomposition optimization to the dynamic association weight matrix to obtain an optimized dynamic association weight matrix; The variational graph autoencoder is used to sample hidden variables of the optimized dynamic correlation weight matrix, and the local gradient information extracted by the three-dimensional convolutional network is fused through the message passing mechanism to construct a fully connected topological graph with spatiotemporal correlation. In the fully connected topology graph, all edges are traversed to obtain the path length value of each edge, and the path length value of each edge is compared with the preset value to obtain the edge weight distribution entropy value of each edge.

6. A method for detecting the performance of a liquid crystal display module according to claim 5, characterized in that: Apply low-rank decomposition optimization to the dynamic association weight matrix to obtain an optimized dynamic association weight matrix, which is specifically: The original dynamic correlation weight matrix is ​​decomposed into low-rank components and sparse noise components. According to the amplitude distribution characteristics of the sparse noise components, the Huffman weighting function is introduced to impose nonlinear suppression on outliers. When the low-rank component of an element at a certain position is higher than the confidence interval, the weight attenuation mechanism based on the adaptive momentum factor is activated for correction. A low-rank basis matrix is ​​generated based on the processed low-rank components and sparse noise components. An incremental non-negative matrix decomposition algorithm is used to project the low-rank basis matrix into a non-negative manifold space. If the decomposed basis vector has a negative component, an orthogonal correction layer based on KL divergence is triggered, and the coupling relationship between the basis matrix and the coefficient matrix is ​​iteratively optimized by the alternating least squares method. A graph Laplacian operator with a dynamic correlation weight matrix is ​​constructed. When the spectral energy distribution of the low-rank component does not match the graph structure prior, the multi-scale spectral regularization term generated by the graph convolutional network is injected, and the retention ratio of low-frequency components to high-frequency components is dynamically adjusted using the attention mechanism. When the cosine similarity of the basis matrices of adjacent iterations reaches the convergence threshold, the early stopping mechanism is triggered to output the optimized dynamic association weight matrix.

7. The method for detecting the performance of a liquid crystal display module according to claim 1, characterized in that: The feature vector includes the RGB three-channel brightness value, the average of the absolute value of the gradient difference with the adjacent 8 pixels, and the saturation of the HSV color space.

8. The method for detecting the performance of a liquid crystal display module according to claim 1, characterized in that: The pure color images include pure black, pure white, pure red, pure green and pure blue.

9. A performance detection system for a liquid crystal display module, characterized in that: The performance detection of the LCD display module includes a memory and a processor, wherein the memory stores a performance detection method program of the LCD display module. When the performance detection method program of the LCD display module is executed by the processor, the performance detection method steps of the LCD display module as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a performance detection method program for a liquid crystal display module, and when the program is executed by a processor, the performance detection method steps for a liquid crystal display module as claimed in any one of claims 1 to 8 are implemented.

Images