A visual inspection and evaluation method and system for liquid crystal screen sealing quality

Abstract

The application discloses a kind of liquid crystal screen glue quality visual inspection and evaluation method, system, after obtaining encapsulation process timing multimodal image data, the multidimensional space-time state atlas of encapsulation area is generated based on encapsulation process timing multimodal image data;Multi-dimensional space-time state atlas is analyzed to timing dynamic characteristics, and the behavior mode feature of encapsulation solidification process is extracted;Encapsulation solidification process behavior mode feature is input into glue rheology-thermal curing coupling inversion model, and process parameter deviation interval and material state distribution diagram are output;Dynamic statistical reference behavior mode library is established;Finally, the current behavior mode feature is matched and deviation degree is calculated based on dynamic statistical reference behavior mode library, and process parameter deviation interval and material state distribution diagram are combined, and encapsulation quality comprehensive evaluation report and process optimization guidance suggestion are output.The application realizes depth dynamic diagnosis and root cause process optimization to encapsulation solidification process, and improves the accuracy and initiative of quality control.

Description

Technical Field

[0001] This invention relates to the field of LCD screen packaging quality inspection technology, and in particular to a visual inspection and evaluation method and system for LCD screen encapsulation quality. Background Technology

[0002] In the manufacturing process of LCD displays, the encapsulation process is a crucial step in ensuring the sealing of screen edges and preventing leakage and dust ingress. This process involves multiple stages, including dispensing, leveling, and curing. The morphological evolution and internal state of the adhesive during curing directly affect the reliability of the final encapsulation. Currently, industry quality inspection primarily relies on offline sampling after the process is completed, such as checking static indicators like the continuity and width uniformity of the adhesive lines using optical methods. With increasing demands for automation, online inspection technology based on machine vision is gradually being applied. By acquiring and analyzing images of the cured adhesive lines using a visible light camera, some surface defects can be automatically identified. To further improve the comprehensiveness and accuracy of inspection, related technical solutions are beginning to incorporate multi-source information fusion. For example, by simultaneously acquiring thermal imaging and 3D morphology data of the encapsulation structure and combining this data with a historical defect sample library for feature matching and threshold judgment, comprehensive detection and preliminary classification of various types of defects can be achieved. This, to a certain extent, shifts quality control from a single dimension to multiple dimensions, and from result sampling to online monitoring.

[0003] However, existing detection methods based on multi-source information still rely on comparing and judging the static characteristics of the completed process. For dynamic curing processes like sealing, which involve colloidal rheology, thermal conduction, and chemical cross-linking, current technologies struggle to capture and analyze the continuous evolution of the material's state within the process window. When a quality anomaly is detected, the system typically only indicates the presence of a defect, failing to trace the root cause of the anomaly's underlying process parameter deviations or provide proactive optimization guidance. This leaves existing quality control systems lacking depth and initiative in dealing with complex and dynamic sealing processes. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a visual inspection and evaluation method and system for the sealing quality of LCD screens. By analyzing the dynamic evolution law of the sealing curing process and tracing its process root cause, the invention achieves in-depth diagnosis of sealing quality and precise process optimization.

[0005] To achieve the aforementioned technical objectives, in a first aspect, the technical solution adopted by the present invention is: a visual inspection and evaluation method for the sealing quality of a liquid crystal display screen, comprising:

[0006] Acquire multimodal image data of the sealing process, which includes visible light image sequences, near-infrared spectral image sequences, and three-dimensional line laser scanning point cloud sequences that are synchronously acquired within the complete process window from dispensing to curing.

[0007] Based on the temporal multimodal image data of the sealing process, a multidimensional spatiotemporal state map of the sealing area is generated through spatiotemporal registration and data fusion algorithms. The multidimensional spatiotemporal state map contains the fused texture features, spectral features and three-dimensional morphological features at each time point.

[0008] The temporal dynamic features of the multidimensional spatiotemporal state map are analyzed to extract the behavioral pattern features of the sealing and curing process. The behavioral pattern features of the sealing and curing process are obtained by calculating the spectral intensity change field and the three-dimensional morphological displacement field between consecutive time frames.

[0009] The characteristics of the sealing and curing process behavior pattern are input into the colloidal rheology-thermal curing coupled inversion model. The process parameter deviation range and material state distribution map that lead to the observed behavior pattern are output through inverse solution. The colloidal rheology-thermal curing coupled inversion model is built on the basis of the forward physical simulation model.

[0010] A dynamic statistical reference behavior pattern library is established. The dynamic statistical reference behavior pattern library generates standard behavior pattern clusters and their allowable fluctuation ranges for each process stage by accumulating historically qualified sealing process time-series multimodal image data and performing cluster analysis.

[0011] Based on a dynamic statistical reference behavior pattern library, the behavior pattern characteristics of the current sealing curing process in the inspected sealing area are matched and the deviation is calculated. Combined with the deviation range of process parameters and the material state distribution map, a comprehensive assessment report on sealing quality and process optimization guidance suggestions are output.

[0012] In some embodiments, based on the time-series multimodal image data of the sealing process, a multidimensional spatiotemporal state map of the sealing region is generated through a spatiotemporal registration and data fusion algorithm, including:

[0013] The visible light image sequence, near-infrared spectral image sequence and three-dimensional line laser scanning point cloud sequence are time-stamp aligned and spatial coordinate system unified to generate a time-synchronized and spatially aligned registered multimodal data stream;

[0014] For each time frame in the registered multimodal data stream, perform the following fusion operation:

[0015] Extract the visible light image of the corresponding time frame from the visible light image sequence, calculate the local texture descriptor of the visible light image, and generate a texture feature map;

[0016] Near-infrared spectral images of corresponding time frames are extracted from the near-infrared spectral image sequence. Multiple preset characteristic spectral bands are selected, the reflection intensity of each characteristic spectral band is calculated, and a spectral feature map is generated.

[0017] Extract the 3D point cloud corresponding to the time frame from the 3D line laser scanning point cloud sequence, project the 3D point cloud onto the same imaging plane as the visible light image, calculate the elevation statistics of the region corresponding to each pixel after projection, and generate a 3D topographic feature map.

[0018] The texture feature map, spectral feature map and three-dimensional topography feature map are aligned pixel by pixel in space, and the texture feature value, reflection intensity value of multiple feature spectral bands and three-dimensional topography elevation statistics at each spatial pixel position are combined into a multi-dimensional fused feature vector.

[0019] Based on the multi-dimensional fusion feature vectors at all spatial pixel locations, construct the fusion feature vector field for the current time frame;

[0020] The fused feature vector fields of each time frame are arranged and indexed in chronological order to form a multidimensional spatiotemporal state map of the sealing area. The structure of the multidimensional spatiotemporal state map enables the querying and slice analysis of the multidimensional fused feature vectors by time dimension and spatial location.

[0021] In some embodiments, temporal dynamic feature analysis is performed on the multidimensional spatiotemporal state map to extract behavioral pattern features of the sealing and curing process, including:

[0022] From the multidimensional spatiotemporal state map, extract the fused feature vector field of consecutive time frames in chronological order;

[0023] For each spatial cell location in the fused feature vector field, perform the following calculation:

[0024] Calculate the difference in texture feature values ​​between adjacent time frames for the spatial pixel location to obtain the gradient of texture feature change;

[0025] Extract the reflection intensity values ​​of multiple characteristic spectral bands of the spatial pixel location in consecutive time frames to form a time series of each characteristic spectral band, calculate the first difference of the time series, and obtain the spectral intensity change sequence.

[0026] Calculate the displacement vector of the spatial pixel location in the statistical value of the three-dimensional topography elevation between adjacent time frames to obtain the displacement of the three-dimensional topography change.

[0027] Based on the texture feature change gradient, spectral intensity change sequence and three-dimensional morphology change displacement of all spatial pixel locations, a dynamic change field of the sealing area within the observation time window is constructed. The dynamic change field includes the texture change field, spectral intensity change field and three-dimensional morphology displacement field.

[0028] Pattern recognition analysis is performed on the dynamic changing field to extract the behavioral pattern features of the sealing and curing process, which characterize the physical process of sealing and curing. These behavioral pattern features include the distribution pattern of the retraction velocity at the edge of the colloid, the trajectory pattern of the internal bubble movement, the abnormal pattern of the local spectral curing rate, and the uneven shrinkage pattern of the three-dimensional morphology.

[0029] In some embodiments, the colloidal rheology-thermosetting coupled inversion model is constructed through the following steps:

[0030] A forward physical simulation model of colloidal rheology-thermal curing was established. The inputs of the forward physical simulation model of colloidal rheology-thermal curing include dispensing path parameters, adhesive quantity parameters, temperature field parameters, and time parameters. The output of the forward physical simulation model of colloidal rheology-thermal curing is the predicted multidimensional spatiotemporal state map of the sealing area.

[0031] The colloidal rheology-thermosetting forward physics simulation model is encapsulated as a differentiable computational graph, making the output of the colloidal rheology-thermosetting forward physics simulation model differentiable with respect to the input parameters of the colloidal rheology-thermosetting forward physics simulation model.

[0032] Using the behavioral pattern characteristics of the sealing and curing process as observation constraints, the difference between the predicted output of the colloidal rheology-thermal curing forward physical simulation model and the observation constraints is constructed as a loss function;

[0033] The gradient descent optimization algorithm is used to iteratively optimize the loss function. The gradient is backpropagated through a differentiable computation graph to adjust the input parameters of the colloidal rheology-thermosetting forward physical simulation model until the loss function converges.

[0034] The deviation between the input parameters of the optimized colloidal rheology-thermosetting forward physical simulation model and the standard process parameters is output as the process parameter deviation range.

[0035] Meanwhile, the internal material state distribution obtained by running the colloidal rheology-thermosetting forward physical simulation model under optimized input parameters is output as a material state distribution map.

[0036] In some embodiments, the behavioral pattern characteristics of the sealing and curing process are input into the colloidal rheology-thermal curing coupled inversion model, and the process parameter deviation range and material state distribution map that lead to the observed behavioral pattern are output through inverse solution, including:

[0037] The behavioral pattern characteristics of the sealing and curing process are used as the observation constraint input for the colloidal rheology-thermal curing coupled inversion model;

[0038] Within the colloidal rheology-thermal curing coupled inversion model, the observation constraints are quantified into target feature vectors corresponding to the output of the colloidal rheology-thermal curing forward physical simulation model.

[0039] The gradient descent optimization algorithm is executed to iteratively adjust the dispensing path parameters, adhesive quantity parameters, temperature field parameters, and time parameters, so as to minimize the loss function value between the predicted feature vector and the target feature vector output by the colloidal rheology-thermal curing forward physical simulation model.

[0040] When the loss function converges, record the optimized values ​​of the dispensing path parameters, glue quantity parameters, temperature field parameters, and time parameters at this point.

[0041] Calculate the difference between the optimized value and the preset standard process parameter value, and output the confidence interval of the difference as the process parameter deviation interval.

[0042] Meanwhile, internal state variables are extracted from the converged colloidal rheology-thermal curing forward physical simulation model. These internal state variables include the colloidal viscosity field, curing degree field, and stress field. The internal state variables are then mapped to spatial locations to generate a material state distribution map.

[0043] In some embodiments, a dynamic statistical reference behavior pattern library is established, including:

[0044] Collect multimodal image data of the encapsulation process of LCD screens that have been judged to be qualified through offline testing from historical production batches;

[0045] Based on the collected time-series multimodal image data of the sealing process, a corresponding set of historical qualified sealing curing process behavior patterns is generated;

[0046] The set of historical qualified sealing curing process behavior patterns is divided into time slices according to the process stages, which include the dispensing stage, the leveling stage and the curing stage.

[0047] For the historical qualified sealing and curing process behavior patterns in each process stage, a clustering algorithm is used to classify the patterns and generate multiple standard behavior pattern clusters for that process stage.

[0048] For each standard behavior pattern cluster, calculate the mean and standard deviation of all historical qualified sealant curing behavior pattern characteristics within the standard behavior pattern cluster in each feature dimension. Use the mean as the center of the standard behavior pattern cluster and the standard deviation of a preset multiple as the allowable fluctuation range of the standard behavior pattern cluster.

[0049] The standard behavior pattern clusters and their corresponding allowable fluctuation ranges are organized and stored according to the process stage to form a dynamic statistical reference behavior pattern library.

[0050] In some embodiments, based on a dynamic statistical reference behavior pattern library, the sealing curing process behavior pattern characteristics of the currently inspected sealing area are matched and deviations are calculated, including:

[0051] The sealing and curing process behavior patterns of the current inspected sealing area are divided into the same process stages as those in the dynamic statistical reference behavior pattern library according to the time dimension.

[0052] For each process stage, retrieve all standard behavior pattern clusters and their allowable fluctuation ranges corresponding to that process stage from the dynamic statistical reference behavior pattern library;

[0053] Calculate the feature vector of the sealing curing process behavior pattern of the current inspected sealing area within this process stage, and the feature space distance between it and the central feature vector of each retrieved standard behavior pattern cluster;

[0054] The cluster of standard behavioral patterns with the smallest feature space distance is selected as the best matching cluster;

[0055] Determine whether the characteristic vector of the sealing curing process behavior pattern of the current inspected sealing area falls within the allowable fluctuation range of the best matching cluster in this process stage;

[0056] If it falls within the allowable fluctuation range, the standardized deviation between the feature vector of the sealing curing process behavior pattern of the current inspected sealing area and the central feature vector of the best matching cluster is calculated as the matching deviation of this process stage.

[0057] If it does not fall within the allowable fluctuation range, it is judged as no match, and the cross-boundary distance from the feature vector of the sealing curing process behavior mode of the current inspected sealing area to the boundary of the allowable fluctuation range of the best matching cluster is calculated as the matching deviation of this process stage.

[0058] Combine the matching deviations of all process stages in chronological order to generate the overall behavior pattern deviation sequence of the currently inspected sealing area.

[0059] In some embodiments, the standardized deviation between the feature vector of the sealing curing process behavior pattern characteristics of the current inspected sealing area and the central feature vector of the best matching cluster is calculated as the matching deviation for this process stage, including:

[0060] Extract the feature vector of the sealing curing process behavior pattern of the current inspected sealing area within the process stage, and denote it as the current feature vector;

[0061] Extract the central feature vector of the best matching cluster, and denote it as the reference center vector;

[0062] Extract the standard deviation of each feature dimension within the allowable fluctuation range of the best matching cluster, and denot it as the standard deviation of each dimension;

[0063] Calculate the difference between the current feature vector and the reference center vector in each feature dimension to obtain the difference in each dimension;

[0064] Divide the difference between each feature dimension by the standard deviation of the corresponding feature dimension to obtain the standardized difference of that feature dimension.

[0065] The sum of squares of the standardized differences of all feature dimensions is taken, and then the square root of the sum of squares is taken to obtain the standardized deviation between the current feature vector and the reference center vector.

[0066] The calculated standardized deviation is used as the matching deviation output for the process stage.

[0067] In some embodiments, by combining the deviation range of process parameters with the material state distribution diagram, a comprehensive evaluation report on sealing quality and process optimization guidance suggestions are output, including:

[0068] The overall behavior pattern deviation sequence is input into the quality assessment rule engine, which predefines the deviation threshold and defect type mapping relationship for different process stages.

[0069] Based on whether the matching deviation of each process stage in the overall behavior pattern deviation sequence exceeds the deviation threshold of the corresponding stage, it is determined whether there is a defect and the process stage to which the defect belongs.

[0070] When a defect is identified, the suspected abnormal process parameters causing the defect are located by combining the deviation range of process parameters. Suspected abnormal process parameters include deviation of dispensing path parameters, deviation of adhesive quantity parameters, deviation of temperature field parameters, or deviation of time parameters.

[0071] At the same time, by combining the material state distribution map, abnormal state areas inside the material can be identified. These abnormal state areas include areas of insufficient curing, stress concentration areas, or areas rich in impurities.

[0072] Based on the process stage to which the defect belongs, the suspected abnormal process parameters, and the abnormal state area inside the material, a quality assessment report is generated that includes the defect type, defect severity level, defect spatial location, root cause process parameter inference, and description of the abnormal material state.

[0073] Based on the root cause process parameters in the quality assessment report, corresponding optimization suggestions are retrieved from the preset process parameter adjustment knowledge base. The optimization suggestions include suggestions on the direction and amount of adjustment for suspected abnormal process parameters.

[0074] The quality assessment report and optimization suggestions are combined to output a comprehensive sealing quality assessment report and process optimization guidance suggestions.

[0075] In a second aspect, the present invention also provides a visual inspection and evaluation system for the sealing quality of a liquid crystal display screen, applicable to the method described in the first aspect. The system includes a multimodal temporal image acquisition module, a multidimensional spatiotemporal state map generation module, a behavior pattern feature extraction module, a process inversion and state analysis module, a dynamic reference pattern library management module, and a quality assessment and decision output module. The multimodal temporal image acquisition module is used to acquire temporal multimodal image data of the sealing process. This temporal multimodal image data includes a visible light image sequence, a near-infrared spectral image sequence, and a three-dimensional line laser scanning point cloud sequence, all synchronously acquired within the complete process window from dispensing to curing. The multidimensional spatiotemporal state map generation module is used to generate a multidimensional spatiotemporal state map of the sealing area based on the temporal multimodal image data of the sealing process, using a spatiotemporal registration and data fusion algorithm. The multidimensional spatiotemporal state map includes fused texture features, spectral features, and three-dimensional morphological features at each time point. The behavior pattern feature extraction module is used to perform temporal dynamic feature analysis on the multidimensional spatiotemporal state map to extract the behavior of the sealing curing process. The sealing curing process behavior pattern is obtained by calculating the spectral intensity change field and three-dimensional morphological displacement field between consecutive time frames. The process inversion and state analysis module is used to input the sealing curing process behavior pattern characteristics into the colloidal rheology-thermal curing coupled inversion model. Through inverse solving, it outputs the process parameter deviation range and material state distribution map that lead to the observed behavior pattern. The colloidal rheology-thermal curing coupled inversion model is built on the forward physical simulation model. The dynamic reference pattern library management module is used to establish and maintain a dynamic statistical reference behavior pattern library. The dynamic statistical reference behavior pattern library generates standard behavior pattern clusters and their allowable fluctuation ranges for each process stage by accumulating historically qualified sealing process time-series multimodal image data and performing cluster analysis. The quality assessment and decision output module is used to match and calculate the deviation of the sealing curing process behavior pattern characteristics of the current inspected sealing area based on the dynamic statistical reference behavior pattern library. Combined with the process parameter deviation range and material state distribution map, it outputs a comprehensive sealing quality assessment report and process optimization guidance suggestions.

[0076] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: The present invention provides a visual inspection and evaluation method and system for the sealing quality of LCD screens. After acquiring multimodal image data of the sealing process, a multidimensional spatiotemporal state map of the sealing area is generated based on the multimodal image data of the sealing process; the multidimensional spatiotemporal state map is analyzed for temporal dynamic features to extract the behavioral pattern features of the sealing curing process; the behavioral pattern features of the sealing curing process are input into the colloid rheology-thermal curing coupled inversion model to output the process parameter deviation range and material state distribution map; a dynamic statistical reference behavioral pattern library is established; finally, based on the dynamic statistical reference behavioral pattern library, the current behavioral pattern features are matched and the deviation is calculated, and combined with the process parameter deviation range and material state distribution map, a comprehensive sealing quality evaluation report and process optimization guidance suggestions are output. The present invention realizes in-depth dynamic diagnosis and root cause process optimization of the sealing curing process, improving the accuracy and initiative of quality control. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 This is a schematic diagram of steps S101 to S106 of the method described in the specific implementation embodiment;

[0079] Figure 2 This is a schematic diagram of steps S201 to S204 of the method described in the specific implementation embodiment;

[0080] Figure 3 This is a schematic diagram of the structure of the visual inspection and evaluation system described in the specific implementation.

[0081] The reference numerals for the above figures are as follows:

[0082] 1. Visual inspection and evaluation system;

[0083] 11. Multimodal temporal image acquisition module;

[0084] 12. Multidimensional spatiotemporal state map generation module;

[0085] 13. Behavioral pattern feature extraction module;

[0086] 14. Process Inversion and State Analysis Module;

[0087] 15. Dynamic Reference Pattern Library Management Module;

[0088] 16. Quality assessment and decision output module. Detailed Implementation

[0089] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] Please see Figure 1 In a first aspect, this embodiment provides a visual inspection and evaluation method for the sealing quality of a liquid crystal display screen, including:

[0091] S101. Acquire multimodal image data of the sealing process. The multimodal image data of the sealing process includes visible light image sequences, near-infrared spectral image sequences, and three-dimensional line laser scanning point cloud sequences that are synchronously acquired within the complete process window from dispensing to curing.

[0092] S102. Based on the time-series multimodal image data of the sealing process, a multidimensional spatiotemporal state map of the sealing area is generated through spatiotemporal registration and data fusion algorithms. The multidimensional spatiotemporal state map includes the fused texture features, spectral features and three-dimensional morphological features at each time point.

[0093] S103. Perform time-series dynamic feature analysis on the multidimensional spatiotemporal state map, and extract the behavior pattern features of the sealing and curing process. The behavior pattern features of the sealing and curing process are obtained by calculating the spectral intensity change field and the three-dimensional morphological displacement field between consecutive time frames.

[0094] S104. Input the characteristics of the sealing and curing process behavior mode into the colloidal rheology-thermal curing coupled inversion model, and output the process parameter deviation range and material state distribution map that lead to the observed behavior mode through inverse solution. The colloidal rheology-thermal curing coupled inversion model is built on the basis of the forward physical simulation model.

[0095] S105. Establish a dynamic statistical reference behavior pattern library. The dynamic statistical reference behavior pattern library generates standard behavior pattern clusters and their allowable fluctuation ranges for each process stage by accumulating historically qualified sealing process time-series multimodal image data and performing cluster analysis.

[0096] S106. Based on the dynamic statistical reference behavior pattern library, match and calculate the behavior pattern characteristics of the sealing curing process in the current inspected sealing area. Combine the deviation range of process parameters and the material state distribution map to output a comprehensive evaluation report on sealing quality and guidance suggestions for process optimization.

[0097] In step S101, the multimodal image data of the sealing process is a set of time-series data synchronously acquired by multiple sensors arranged in a coordinated manner within the complete process time window from the start of dispensing to the end of curing of the LCD screen sealing. Visible light image sequences are captured by an industrial camera to record the temporal changes in the surface macroscopic morphology, color, and texture of the sealing area. Near-infrared spectral image sequences are acquired through a spectral imaging device, reflecting the evolution of the chemical composition and molecular bonding state of the colloid over time during curing. Three-dimensional line laser scanning point cloud sequences are generated by a line laser scanner to accurately measure the temporal changes in the three-dimensional morphology (such as height and contour) of the sealing area surface. Synchronous acquisition ensures the inherent temporal and spatial correlation of different modal data, laying the foundation for subsequent fusion analysis. This step, by acquiring multi-dimensional time-series image data covering the entire sealing process, constructs a data foundation for in-depth process analysis.

[0098] In step S102, the spatiotemporal registration and data fusion algorithm is used to solve the problem of alignment and integration of multi-source heterogeneous data in time and space. Spatiotemporal registration synchronizes image sequences from different sensors on the time axis and performs unified mapping on the spatial coordinate system, enabling data from different modalities to represent the state of the same physical location at the same time. Based on this, the data fusion algorithm integrates the registered visible light texture information, near-infrared spectral information, and three-dimensional morphological information at each time point to generate a unified multi-dimensional spatiotemporal state map. Each spatiotemporal unit of this map contains fused texture features, spectral features, and three-dimensional morphological features, thus forming a four-dimensional data volume that can comprehensively and continuously describe the state evolution of the sealing area. This step generates a high-dimensional state model that can completely characterize the dynamic process of sealing by fusing multimodal time-series data.

[0099] In step S103, temporal dynamic feature analysis extracts information reflecting the dynamic evolution of the process from static feature snapshots. By calculating the changes in spectral features between consecutive time frames, a spectral intensity variation field can be obtained, revealing the spatiotemporal distribution of chemical processes such as the colloid curing reaction rate and uniformity. By calculating the displacement of three-dimensional morphological features between consecutive time frames, a three-dimensional morphological displacement field can be obtained, reflecting the spatiotemporal patterns of physical deformations such as colloid shrinkage and flow. The encapsulation curing process behavior pattern characteristics are a high-level feature set that characterizes specific curing physical mechanisms, obtained after further analysis and refinement of these dynamic variation fields. This step elevates the original observation data into behavioral pattern characteristics reflecting the intrinsic physicochemical processes of curing by analyzing the dynamic variation fields.

[0100] In step S104, the colloidal rheology-thermocuring coupled inversion model is a reverse reasoning tool based on physical laws. Its foundation is a forward physical simulation model capable of simulating the flow, deformation, and curing process of colloids under given process parameters (such as dispensing volume and temperature profile). The inversion model utilizes the computability of this forward physical simulation model, taking the encapsulation and curing process behavior patterns extracted in step S103 as the actually observed "results" or "effects." It uses an optimization algorithm to inversely solve for the most likely combination of input process parameters that would lead to the observed result and its uncertainty range, i.e., the process parameter deviation range. Simultaneously, the inversion process also outputs the spatial distribution of the colloid's internal states (such as viscosity and degree of cure) corresponding to the optimized parameters, i.e., a material state distribution map. This step, through inverse solving driven by a physical model, establishes a quantitative correlation between observed abnormal behavior patterns and potential process root causes and internal material states.

[0101] In step S105, the establishment of the dynamic statistical reference behavior pattern library relies on the digital accumulation and knowledge extraction of historical successful production experience. By collecting a large amount of multimodal image data of the sealing process corresponding to historical qualified products, and processing it through the aforementioned steps to obtain its behavior pattern characteristics, a feature set of qualified samples is formed. Cluster analysis is used to perform unsupervised learning on this feature set, grouping samples with similar dynamic behavior patterns into one category, forming standard behavior pattern clusters for each process stage (such as leveling, preliminary curing, and complete curing). The allowable fluctuation range of each cluster is determined by statistically analyzing the distribution of sample features within that cluster, thus defining the acceptable normal variation boundary of the standard pattern. This step automatically summarizes the normal dynamic behavior standards of the sealing process from historical data using machine learning methods.

[0102] In step S106, the behavioral pattern characteristics of the currently inspected sealing area are matched with standard clusters in the dynamic statistical reference behavioral pattern library, and the deviation is calculated. This allows for a quantitative assessment of the degree to which its dynamic behavior conforms to historical health standards. Combined with the process parameter deviation range and material state distribution map obtained from step S104, possible root cause explanations and internal state evidence can be provided for any detected behavioral deviations. The comprehensive sealing quality assessment report integrates information such as behavioral deviation, suspected process parameter deviations, and abnormal material states to comprehensively judge the quality level and defect risk. Process optimization guidance suggestions, based on the root cause inferences in the report, propose targeted directions for process parameter adjustments. This step, through multi-source information fusion decision-making, achieves closed-loop quality control from process monitoring and anomaly diagnosis to optimization suggestion output.

[0103] This embodiment provides a method for inspecting and evaluating the sealing quality of LCD screens by deeply integrating multimodal perception, physical mechanism modeling, and data-driven learning, starting from dynamic process monitoring. It overcomes the limitations of traditional static or single-modal detection by constructing a multidimensional spatiotemporal state map of the sealing process and analyzing its behavioral patterns, achieving deep perception of the complex dynamic process of curing. Furthermore, it correlates observed behavior with process root causes and material states through a physical inversion model, and performs intelligent comparison using a dynamic reference library established from historical qualified data. Finally, it outputs a comprehensive decision including quality assessment and optimization suggestions. This method can not only detect potential defects earlier and more accurately, but also reveal the process causes behind the defects, thereby achieving predictive maintenance and precise process optimization.

[0104] Please see Figure 2 In some embodiments, based on the temporal multimodal image data of the sealing process, a multidimensional spatiotemporal state map of the sealing region is generated through spatiotemporal registration and data fusion algorithms, including:

[0105] S201. Perform timestamp alignment and spatial coordinate system unification on the visible light image sequence, near-infrared spectral image sequence and three-dimensional line laser scanning point cloud sequence to generate a time-synchronized and spatially aligned registered multimodal data stream.

[0106] S202. For each time frame in the registered multimodal data stream, perform the following fusion operation:

[0107] Extract the visible light image of the corresponding time frame from the visible light image sequence, calculate the local texture descriptor of the visible light image, and generate a texture feature map;

[0108] Near-infrared spectral images of corresponding time frames are extracted from the near-infrared spectral image sequence. Multiple preset characteristic spectral bands are selected, the reflection intensity of each characteristic spectral band is calculated, and a spectral feature map is generated.

[0109] Extract the 3D point cloud corresponding to the time frame from the 3D line laser scanning point cloud sequence, project the 3D point cloud onto the same imaging plane as the visible light image, calculate the elevation statistics of the region corresponding to each pixel after projection, and generate a 3D topographic feature map.

[0110] The texture feature map, spectral feature map and three-dimensional topography feature map are aligned pixel by pixel in space, and the texture feature value, reflection intensity value of multiple feature spectral bands and three-dimensional topography elevation statistics at each spatial pixel position are combined into a multi-dimensional fused feature vector.

[0111] S203. Based on the multi-dimensional fusion feature vectors at all spatial pixel locations, construct the fusion feature vector field of the current time frame;

[0112] S204. Arrange and index the fused feature vector field of each time frame in chronological order to form a multidimensional spatiotemporal state map of the sealing area. The structure of the multidimensional spatiotemporal state map enables querying and slice analysis of the multidimensional fused feature vector by time dimension and spatial location.

[0113] In step S201, timestamp alignment is achieved by configuring synchronous trigger signals for all acquisition devices (visible light camera, near-infrared spectral imager, 3D line laser scanner) or using a high-precision unified clock source, ensuring that each frame of data has an accurate and correlated timestamp, thereby aligning image sequences of different modalities to the same sampling time on the time axis. The spatial coordinate system involves establishing a unified two-dimensional or three-dimensional reference coordinate system. Preferably, using the pixel coordinate system of the visible light image as a reference, the spatial transformation relationship (such as a rigid body transformation matrix) from the near-infrared spectral image and the 3D line laser scanning point cloud data to this reference coordinate system is calculated using camera calibration and sensor extrinsic parameter calibration techniques. Applying these transformation relationships, data from different sources can be mapped to the same spatial reference frame, forming a registered multimodal data stream, where the data at each time point strictly corresponds in time and space.

[0114] In step S202, the fusion operation is performed independently for each time point of the registered multimodal data stream. For visible light images, local texture descriptors are used to quantify the texture characteristics of local regions of the image. They can be obtained by calculating statistics such as contrast and correlation of the gray-level co-occurrence matrix (GLCM), or by using algorithms such as Local Binary Pattern (LBP). The final generated texture feature map has each pixel value representing the texture complexity or pattern of that location region.

[0115] For near-infrared spectral images, multiple preset characteristic spectral bands are selected based on the characteristic absorption peaks of specific chemical bonds (such as epoxy groups and hydroxyl groups) in the curing process of the sealant material. Calculating the reflection intensity of each characteristic spectral band is to extract or average the spectral response value of that band. The generated spectral feature map is a set of images, and each image reflects the spatial distribution of the reflection intensity of a characteristic band.

[0116] For 3D point clouds, projection onto the visible light image plane typically employs perspective projection or orthographic projection models. After projection, the elevation statistics of the region corresponding to each pixel are calculated. The elevation statistics can be the average, median, or maximum value of the Z-coordinate (height) of all point clouds in that region, used to characterize the surface height or topographic undulation at the pixel location, generating a 3D topographic feature map.

[0117] Since the rigorous spatial registration in step S201 has already been performed, the pixel-by-pixel alignment of the texture feature map, spectral feature map, and 3D topography feature map in space is naturally feasible. At each spatial pixel location, the texture feature scalar value, the reflection intensity scalar values ​​of multiple feature spectral bands (forming a spectral feature vector), and the 3D topography elevation statistical scalar value at that location are sequentially concatenated to form a multi-dimensional fused feature vector.

[0118] In step S203, when constructing the fusion feature vector field for the current time frame, preferably, each pixel position in the image grid corresponding to the sealing region is filled with its corresponding multidimensional fusion feature vector. The fusion feature vector field can be viewed as a three-dimensional tensor in terms of data structure, where two dimensions correspond to spatial positions (rows and columns), and the third dimension corresponds to each component of the fusion feature vector, thus completely depicting the multimodal fusion state at each spatial point in the sealing region under the current time frame. This step organizes discrete pixel feature vectors into a continuous spatial field representation, facilitating spatial correlation analysis and calculation.

[0119] In step S204, the multidimensional spatiotemporal state map of the sealing area is logically a four-dimensional data volume, containing two spatial dimensions, one temporal dimension, and one feature dimension. By establishing an effective index structure (e.g., an index based on timestamps and spatial coordinates), the corresponding multidimensional fused feature vector can be efficiently queried by time point and spatial location, or slice analysis can be performed on data from specific time intervals or spatial regions, such as extracting the feature curve of a certain location point changing over time, or analyzing the feature distribution of the entire sealing area at a certain moment. This step ultimately generates a structured, queryable spatiotemporal data model, providing a directly operable data foundation for subsequent temporal dynamic feature analysis.

[0120] This embodiment ensures the inherent consistency of multi-source data through precise spatiotemporal registration. By independently extracting multimodal features and fusing them at the pixel level for each time frame, a comprehensive feature vector that simultaneously reflects texture, spectrum, and morphology is generated. These vectors are then organized into a spatial field and arranged in a time sequence, ultimately constructing a structured four-dimensional spatiotemporal data volume. The multidimensional spatiotemporal state map of the sealing region retains all the information dimensions of the original data, and through fusion and structuring, a standardized feature representation that is easy for machines to understand and analyze is generated, laying a solid data foundation for efficiently extracting the dynamic laws of the curing process from massive amounts of original data. This embodiment elevates the observation of the sealing dynamic process from separate multimodal signals to a unified spatiotemporal state model.

[0121] In some embodiments, temporal dynamic feature analysis is performed on the multidimensional spatiotemporal state map to extract behavioral pattern features of the sealing and curing process, including:

[0122] From the multidimensional spatiotemporal state map, extract the fused feature vector field of consecutive time frames in chronological order;

[0123] For each spatial cell location in the fused feature vector field, perform the following calculation:

[0124] Calculate the difference in texture feature values ​​between adjacent time frames for the spatial pixel location to obtain the gradient of texture feature change;

[0125] Extract the reflection intensity values ​​of multiple characteristic spectral bands of the spatial pixel location in consecutive time frames to form a time series of each characteristic spectral band, calculate the first difference of the time series, and obtain the spectral intensity change sequence.

[0126] Calculate the displacement vector of the spatial pixel location in the statistical value of the three-dimensional topography elevation between adjacent time frames to obtain the displacement of the three-dimensional topography change.

[0127] Based on the texture feature change gradient, spectral intensity change sequence and three-dimensional morphology change displacement of all spatial pixel locations, a dynamic change field of the sealing area within the observation time window is constructed. The dynamic change field includes the texture change field, spectral intensity change field and three-dimensional morphology displacement field.

[0128] Pattern recognition analysis is performed on the dynamic changing field to extract the behavioral pattern features of the sealing and curing process, which characterize the physical process of sealing and curing. These behavioral pattern features include the distribution pattern of the retraction velocity at the edge of the colloid, the trajectory pattern of the internal bubble movement, the abnormal pattern of the local spectral curing rate, and the uneven shrinkage pattern of the three-dimensional morphology.

[0129] In this embodiment, a time series is formed by fusing feature vector fields of consecutive time frames extracted sequentially from the multidimensional spatiotemporal state map. For each spatial pixel location in the field, the gradient of texture feature change can be directly obtained by calculating the difference in texture feature values ​​between adjacent time frames. This gradient value reflects the severity and direction of the texture change at that point over time. Here, the texture feature values, the reflection intensity values ​​of the feature spectral bands, and the statistical values ​​of the three-dimensional topography elevation all refer to the component values ​​of the corresponding dimensions extracted from the multidimensional fused feature vector corresponding to that pixel location.

[0130] For spectral features, the reflection intensity values ​​of each preset feature spectral band at the pixel location in continuous time frames can be extracted to form multiple time series. Calculating the first difference of these time series, i.e., the value of the next time step minus the value of the previous time step, yields a spectral intensity change sequence that characterizes the rate of change of reflection intensity of each band over time, which is directly related to the kinetic process of colloidal chemical reactions.

[0131] For three-dimensional topography, the displacement vector of the elevation statistics corresponding to the same pixel position between adjacent time frames is calculated. The displacement of three-dimensional topography changes includes not only the change in height, but also the displacement component in the horizontal direction, which is used to describe the local flow or contraction deformation of the colloidal surface.

[0132] Based on the texture feature change gradients, spectral intensity change sequences, and three-dimensional morphological displacements calculated from all spatial pixel locations, spatial distribution fields covering the entire sealing area can be constructed. The texture change field is a spatial distribution map of the texture feature change gradients at each pixel location. The spectral intensity change field can be a multidimensional field, with each spatial location corresponding to a vector, and the vector components being the statistics (such as mean and variance) of the change sequence of each characteristic spectral band within a specific time interval. The three-dimensional morphological displacement field records the displacement vector of each pixel location in three-dimensional space. These dynamic change fields together constitute a complete spatiotemporal description of the physicochemical state changes of the sealing area within the observation time window.

[0133] Pattern recognition analysis of dynamic changing fields extracts high-level semantic features that directly correspond to specific curing physical phenomena or defect mechanisms. For example, by analyzing the time integral of the displacement vector of the colloidal edge region in the three-dimensional morphology displacement field, the distribution pattern of the colloidal edge retraction velocity can be extracted, reflecting the uniformity of wetting and shrinkage between the colloidal material and the substrate or screen. By tracking the spatial anomaly regions of the rate of change of specific bands (such as bands related to the curing reaction) in the spectral intensity change field and their trajectories over time, local spectral curing rate anomaly patterns can be identified, indicating possible incomplete curing or temperature unevenness. By analyzing the divergence or convergence regions of displacement vectors in the three-dimensional morphology displacement field and combining them with time information, the trajectory patterns of internal bubble movement or the three-dimensional morphology shrinkage non-uniformity patterns can be identified. Pattern recognition analysis can be implemented by comprehensively utilizing image processing, spatiotemporal clustering, trajectory tracking, and anomaly detection algorithms. Preferably, a density-based spatiotemporal clustering algorithm is used to identify pixel regions with similar dynamic behaviors, thereby defining different behavioral patterns. The final extracted behavioral pattern features of the sealing curing process are the set of these identified patterns and their quantitative descriptive parameters (such as the spatial range, intensity, and evolution rate of the pattern).

[0134] This embodiment transforms static fusion features into a dynamic changing field by calculating temporal differences and displacements pixel by pixel, achieving precise quantification of the microscopic dynamics of the sealing process. Furthermore, through pattern recognition analysis, it extracts high-level behavioral patterns directly related to specific curing physical mechanisms (such as edge shrinkage, bubble movement, curing reaction, and uneven deformation) from these physical quantity changing fields. These behavioral pattern characteristics of the sealing and curing process condense complex multimodal temporal observation data into a series of characteristic indicators with clear physical meaning, providing high-level information input that can be directly used for causal reasoning and decision-making in subsequent process inversion and quality assessment.

[0135] In some embodiments, the colloidal rheology-thermosetting coupled inversion model is constructed through the following steps:

[0136] A forward physical simulation model of colloidal rheology-thermal curing was established. The inputs of the forward physical simulation model of colloidal rheology-thermal curing include dispensing path parameters, adhesive quantity parameters, temperature field parameters, and time parameters. The output of the forward physical simulation model of colloidal rheology-thermal curing is the predicted multidimensional spatiotemporal state map of the sealing area.

[0137] The colloidal rheology-thermosetting forward physics simulation model is encapsulated as a differentiable computational graph, making the output of the colloidal rheology-thermosetting forward physics simulation model differentiable with respect to the input parameters of the colloidal rheology-thermosetting forward physics simulation model.

[0138] Using the behavioral pattern characteristics of the sealing and curing process as observation constraints, the difference between the predicted output of the colloidal rheology-thermal curing forward physical simulation model and the observation constraints is constructed as a loss function;

[0139] The gradient descent optimization algorithm is used to iteratively optimize the loss function. The gradient is backpropagated through a differentiable computation graph to adjust the input parameters of the colloidal rheology-thermosetting forward physical simulation model until the loss function converges.

[0140] The deviation between the input parameters of the optimized colloidal rheology-thermosetting forward physical simulation model and the standard process parameters is output as the process parameter deviation range.

[0141] Meanwhile, the internal material state distribution obtained by running the colloidal rheology-thermosetting forward physical simulation model under optimized input parameters is output as a material state distribution map.

[0142] In this embodiment, the dispensing path parameter defines the spatial trajectory of the colloid deposition, the colloid quantity parameter determines the colloid volume per unit length, the temperature field parameter describes the thermal history of the curing environment, and the time parameter controls the simulation progress. By solving a set of partial differential equations coupled with the conservation of mass, momentum, and energy, or by using empirical constitutive relations and reaction rate equations, the colloid rheology-thermal curing forward physics simulation model can simulate the entire process of colloid from initial deposition and flow spreading to final curing under the aforementioned process parameters, and output a predicted multidimensional spatiotemporal state map of the sealing region. This map is structurally consistent with the multidimensional spatiotemporal state map constructed based on actual observation data.

[0143] The colloidal rheology-thermosetting forward physics simulation model is encapsulated as a differentiable computational graph. That is, each computational step in the entire simulation process is expressed as a computational node and its connections, making the final output prediction graph differentiable for small changes in all input parameters. This can be achieved using existing automatic differentiation frameworks, which allow the system to automatically calculate the gradient of the output with respect to any input parameter.

[0144] The observational constraints are provided by the behavioral patterns of the encapsulation curing process. For comparison, these behavioral patterns must be mapped to comparable quantities in the output domain of the forward physics simulation model. For example, the distribution pattern of the colloid edge retraction velocity extracted from actual observations is transformed into the morphological displacement rate distribution of the corresponding region in the simulation prediction map over the corresponding time period. The loss function is then used to quantify the difference between the mapped predicted features and the actual observed features; common construction methods include calculating the mean square error or cosine distance between the corresponding feature vectors.

[0145] When iteratively optimizing the loss function using the gradient descent optimization algorithm, in each iteration, the gradient of the loss function with respect to each input process parameter is calculated through the backpropagation mechanism of the differentiable computation graph. Then, along the direction of gradient descent, the values ​​of these input parameters are adjusted according to a preset learning rate. This process is repeated, allowing the predicted features generated by the forward physical simulation model running under the adjusted parameters to continuously approximate the actual observation constraints, until the value of the loss function no longer decreases significantly, reaching a convergence state.

[0146] Once the optimization converges, the input parameter values ​​of the colloidal rheology-thermosetting forward physical simulation model are compared with the preset standard process parameter values, and the difference is calculated. Due to the uncertainties or local optima that may exist in the optimization process, the deviation range of the process parameters can be estimated by statistically analyzing the results of multiple optimization runs or by using the covariance information of the optimization path, thus outputting a possible deviation range rather than a single value.

[0147] Running the colloidal rheology-thermosetting forward physics simulation model under optimized input parameters allows for the acquisition of physical fields within the model that reflect the instantaneous state of the material during the simulation. These fields include the spatial distribution of colloidal viscosity, the distribution of the degree of curing reaction, and the distribution of internal stresses caused by thermal stress or shrinkage. Extracting and visualizing the spatial distribution of these internal state variables in the sealing region yields a material state distribution map.

[0148] This embodiment establishes a complete computational framework for inferring potential process parameter deviations and internal material states from observed encapsulation dynamic behavior patterns by constructing a differentiable forward physical simulation model and utilizing a gradient descent optimization algorithm for inverse solution. The colloidal rheology-thermosetting coupled inverse model combines physical mechanism-based forward simulation with data-driven inverse optimization, enabling quality assessment to move beyond the identification of surface phenomena and delve into the process roots and internal material states that lead to abnormal behavior. This provides a quantitative theoretical basis and computational tools for achieving accurate process diagnosis and optimization.

[0149] In some embodiments, the behavioral pattern characteristics of the sealing and curing process are input into the colloidal rheology-thermal curing coupled inversion model, and the process parameter deviation range and material state distribution map that lead to the observed behavioral pattern are output through inverse solution, including:

[0150] The behavioral pattern characteristics of the sealing and curing process are used as the observation constraint input for the colloidal rheology-thermal curing coupled inversion model;

[0151] Within the colloidal rheology-thermal curing coupled inversion model, the observation constraints are quantified into target feature vectors corresponding to the output of the colloidal rheology-thermal curing forward physical simulation model.

[0152] The gradient descent optimization algorithm is executed to iteratively adjust the dispensing path parameters, adhesive quantity parameters, temperature field parameters, and time parameters, so as to minimize the loss function value between the predicted feature vector and the target feature vector output by the colloidal rheology-thermal curing forward physical simulation model.

[0153] When the loss function converges, record the optimized values ​​of the dispensing path parameters, glue quantity parameters, temperature field parameters, and time parameters at this point.

[0154] Calculate the difference between the optimized value and the preset standard process parameter value, and output the confidence interval of the difference as the process parameter deviation interval.

[0155] Meanwhile, internal state variables are extracted from the converged colloidal rheology-thermal curing forward physical simulation model. These internal state variables include the colloidal viscosity field, curing degree field, and stress field. The internal state variables are then mapped to spatial locations to generate a material state distribution map.

[0156] In this embodiment, the behavioral pattern characteristics of the sealing and curing process, after being input as observation constraints, first need to be quantized into a target feature vector that matches the output domain of the colloidal rheology-thermal curing forward physics simulation model. During the quantization process, the behavioral pattern characteristics (such as the spatial distribution of velocity values ​​contained in the colloidal edge retraction velocity distribution pattern) are extracted into a set of numerical descriptors that can characterize the core characteristics of the pattern. For example, the average retraction velocity of a specific region, the spatial uniformity index of the velocity distribution, or the curvature characteristics of the pattern boundary are calculated. These descriptors, arranged in order, constitute the target feature vector.

[0157] When iteratively adjusting dispensing path parameters, adhesive quantity parameters, temperature field parameters, and time parameters, the gradient descent optimization algorithm calculates an update direction and step size in each iteration based on the gradient of the loss function (a function that measures the difference between the predicted feature vector and the target feature vector) with respect to these process parameters. Dispensing path parameters may be represented as a sequence of control point coordinates, adhesive quantity parameters as flow rate or total quantity scalars, temperature field parameters as temperature-time curves or spatial temperature distribution parameters, and time parameters as the duration of critical process stages. By continuously fine-tuning these parameters along the reverse gradient direction, the predicted feature vector generated by the forward physical simulation model running under the new parameters gradually approaches the target feature vector, thereby driving a continuous decrease in the loss function value.

[0158] When the loss function converges, it indicates that further parameter adjustments cannot significantly reduce the discrepancy between prediction and observation. At this point, the recorded values ​​of dispensing path parameters, adhesive quantity parameters, temperature field parameters, and time parameters are the process parameter estimates that best explain the current observed behavior pattern.

[0159] The difference between the optimized values ​​and the preset standard process parameter values ​​is calculated to obtain the deviation of each parameter. Due to the uncertainty of the optimization process and the possibility of multiple approximate solutions, the confidence interval of the difference can be estimated by analyzing the fluctuation of parameter values ​​in the optimization path. For example, multiple perturbations and re-optimizations can be performed near the optimization convergence point, and the distribution of the obtained parameter values ​​can be statistically analyzed. Then, the mean and standard deviation can be calculated, and the confidence interval can be calculated by adding or subtracting a certain number of times the standard deviation from the mean. This interval reflects the range in which the process parameters causing the observed behavior pattern may deviate from the standard values.

[0160] Among the internal state variables, the colloidal viscosity field describes the flow resistance of the colloid at different locations and times, and its distribution directly affects the spread and shape retention of the colloid; the degree of cure characterizes the extent of the curing reaction, and its spatial non-uniform distribution may lead to insufficient local strength; the stress field reflects the internal stress caused by temperature changes and inconsistent shrinkage, and excessive stress is a potential cause of cracking or debonding. The values ​​of these internal state variables on the simulation mesh are mapped to the actual coordinates of the sealing area according to spatial relationships, and then visualized in the form of images or contour plots, thus generating a material state distribution map.

[0161] This embodiment details the inverse solution process of the colloidal rheology-thermosetting coupled inversion model. By quantifying behavioral pattern features into target vectors and performing meticulous gradient descent optimization, it achieves a precise mapping from complex observation data to specific process parameter deviations. Simultaneously, by calculating confidence intervals, the uncertainty of the inversion results is quantified, enhancing the reliability of the diagnostic conclusions. Furthermore, by extracting and visualizing internal state variables, the potential risk distribution within the material is intuitively revealed. This embodiment enables physical model-based inversion analysis not only to locate process anomalies but also to assess the credibility of these anomalies and gain insight into their internal material effects, providing solid, multi-dimensional data support for subsequent comprehensive quality assessment and precise optimization.

[0162] In some embodiments, a dynamic statistical reference behavior pattern library is established, including:

[0163] Collect multimodal image data of the encapsulation process of LCD screens that have been judged to be qualified through offline testing from historical production batches;

[0164] Based on the collected time-series multimodal image data of the sealing process, a corresponding set of historical qualified sealing curing process behavior patterns is generated;

[0165] The set of historical qualified sealing curing process behavior patterns is divided into time slices according to the process stages, which include the dispensing stage, the leveling stage and the curing stage.

[0166] For the historical qualified sealing and curing process behavior patterns in each process stage, a clustering algorithm is used to classify the patterns and generate multiple standard behavior pattern clusters for that process stage.

[0167] For each standard behavior pattern cluster, calculate the mean and standard deviation of all historical qualified sealant curing behavior pattern characteristics within the standard behavior pattern cluster in each feature dimension. Use the mean as the center of the standard behavior pattern cluster and the standard deviation of a preset multiple as the allowable fluctuation range of the standard behavior pattern cluster.

[0168] The standard behavior pattern clusters and their corresponding allowable fluctuation ranges are organized and stored according to the process stage to form a dynamic statistical reference behavior pattern library.

[0169] In this embodiment, offline testing and qualification are prerequisites for ensuring the quality of data entering the database, referring to LCD screens that pass final tests such as sealing performance and adhesive strength. The collected multimodal image data of the sealing process constitutes the raw material for building the knowledge base.

[0170] Each feature vector in the set of historical qualified sealing curing process behavior patterns is obtained by processing each piece of qualified data through a complete process of acquiring sealing process time-series multimodal image data, generating multidimensional spatiotemporal state maps, and performing time-series dynamic feature analysis.

[0171] The process stages are divided based on the physical characteristics of the colloidal state changes. The dispensing stage is mainly characterized by the extrusion and deposition of the colloidal material from the needle; the leveling stage is characterized by the spreading of the colloidal material under the action of surface tension, and the macroscopic morphology tends to stabilize; the core of the curing stage is the cross-linking reaction that occurs inside the colloidal material, resulting in enhanced mechanical properties. The division can be based on preset timestamps synchronized with equipment operation, or automatically achieved by analyzing abrupt changes in key indicators characterizing flow, spreading, and reaction in the behavioral pattern feature sequence.

[0172] Clustering algorithms are used to discover natural groupings of feature vectors in the feature space within each process stage. For example, the density-based DBSCAN algorithm can group density-connected sample points into the same cluster and effectively identify noise points, thereby automatically forming multiple standard behavior pattern clusters, each cluster representing a recurring qualified dynamic behavior pattern.

[0173] The mean of all samples within a standard behavior pattern cluster is calculated for each feature dimension to obtain a central vector characterizing the cluster's central tendency. The standard deviation of each feature dimension is calculated to quantify the dispersion of samples within the cluster around the center. The allowable range of variation is defined by expanding the central vector by a preset multiple of the standard deviation in each dimension, thus forming a multidimensional ellipsoidal or hypercube-shaped region to delineate the normal variation boundary of the qualified pattern.

[0174] The final structure of the dynamic statistical reference behavior pattern library consists of indexing and encapsulating all standard behavior pattern clusters generated under all process stages and their corresponding allowable fluctuation ranges according to stage labels. The dynamic statistical reference behavior pattern library can be stored using a key-value database or a dedicated data structure, ensuring that all corresponding standard clusters and their statistical parameters can be quickly retrieved based on the process stage.

[0175] This embodiment transforms discrete historical compliance experience into a structured dynamic statistical reference behavior pattern library through systematic data collection, feature extraction, staged clustering, and statistical quantification. It not only summarizes the diversity of compliance processes but also establishes quantitative standards and normal fluctuation ranges for each compliance pattern through statistical methods, thereby providing an intelligent comparison benchmark with both representativeness and tolerance for online detection.

[0176] In some embodiments, based on a dynamic statistical reference behavior pattern library, the sealing curing process behavior pattern characteristics of the currently inspected sealing area are matched and deviations are calculated, including:

[0177] The sealing and curing process behavior patterns of the current inspected sealing area are divided into the same process stages as those in the dynamic statistical reference behavior pattern library according to the time dimension.

[0178] For each process stage, retrieve all standard behavior pattern clusters and their allowable fluctuation ranges corresponding to that process stage from the dynamic statistical reference behavior pattern library;

[0179] Calculate the feature vector of the sealing curing process behavior pattern of the current inspected sealing area within this process stage, and the feature space distance between it and the central feature vector of each retrieved standard behavior pattern cluster;

[0180] The cluster of standard behavioral patterns with the smallest feature space distance is selected as the best matching cluster;

[0181] Determine whether the characteristic vector of the sealing curing process behavior pattern of the current inspected sealing area falls within the allowable fluctuation range of the best matching cluster in this process stage;

[0182] If it falls within the allowable fluctuation range, the standardized deviation between the feature vector of the sealing curing process behavior pattern of the current inspected sealing area and the central feature vector of the best matching cluster is calculated as the matching deviation of this process stage.

[0183] If it does not fall within the allowable fluctuation range, it is judged as no match, and the cross-boundary distance from the feature vector of the sealing curing process behavior mode of the current inspected sealing area to the boundary of the allowable fluctuation range of the best matching cluster is calculated as the matching deviation of this process stage.

[0184] Combine the matching deviations of all process stages in chronological order to generate the overall behavior pattern deviation sequence of the currently inspected sealing area.

[0185] In this embodiment, the division of process stages must be consistent with the construction of the dynamic statistical reference behavior pattern library. Based on the same time window or feature inflection point definition, the currently inspected continuous behavior pattern feature sequence is decomposed into corresponding sub-sequences such as dispensing, leveling, and curing.

[0186] When retrieving the dynamic statistical reference behavior pattern library, the system obtains all pre-stored standard behavior pattern clusters and their statistical parameters according to the current process stage label, including the central feature vector of each cluster and the standard deviation and preset multiple that define the allowable fluctuation range.

[0187] Feature space distance is used to quantify the similarity between the current feature vector and the center vectors of each candidate cluster. Specifically, it can include Euclidean distance, which is the square root of the sum of the squares of the differences in each dimension of the two vectors. When considering the covariance between features, Mahalanobis distance can also be used to obtain a more accurate similarity assessment.

[0188] Determining whether the current feature vector falls within the allowable fluctuation range of the best-matching cluster essentially involves performing a dimension-by-dimensional interval check. It checks whether the values ​​of the current vector in each feature dimension are all within the range of the cluster's central value plus or minus a preset multiple of the standard deviation. If all dimensions satisfy this condition, it is determined to fall within the range.

[0189] When a value is determined to fall within the allowable fluctuation range, a standardized deviation needs to be calculated to quantify its difference from the ideal central state. The standardized deviation transforms the differences in each dimension into dimensionless values ​​with normalized variance through a standardization process, and then calculates a scalar to characterize the degree of "imperfection" of the current state within the acceptable range. Preferably, the square root of the sum of the squares of the standardized differences in each dimension can be used for calculation.

[0190] If the current feature vector does not fall within the allowable fluctuation range, the outbound distance is calculated. The outbound distance measures the severity of the current state's deviation from the acceptable region and can be calculated as the nearest distance from the current vector to the boundary of the allowable fluctuation range (usually the surface of a hypercube or hyperellipsoid).

[0191] The overall behavior pattern deviation sequence is generated by arranging the matching deviations (standardized deviations or out-of-bounds distances) calculated for each process stage in sequence. This sequence constitutes a stage-by-stage quantitative description of the degree to which the dynamic behavior of the current sealing process conforms to historical acceptable standards. Its numerical magnitude and sequence shape directly indicate the stage and severity of the anomaly.

[0192] This embodiment achieves optimal pattern matching by calculating feature space distance, determines pass / fail status through interval testing, and uses standardized deviation and out-of-bounds distance for refined quantification of minor deviations within the pass range and significant anomalies outside the range. The final generated behavioral pattern deviation sequence transforms the complex dynamic behavior assessment into a series of interpretable and comparable numerical indicators, providing clear and quantitative input for subsequent comprehensive quality assessment and root cause localization based on process inversion results.

[0193] In some embodiments, the standardized deviation between the feature vector of the sealing curing process behavior pattern characteristics of the current inspected sealing area and the central feature vector of the best matching cluster is calculated as the matching deviation for this process stage, including:

[0194] Extract the feature vector of the sealing curing process behavior pattern of the current inspected sealing area within the process stage, and denote it as the current feature vector;

[0195] Extract the central feature vector of the best matching cluster, and denote it as the reference center vector;

[0196] Extract the standard deviation of each feature dimension within the allowable fluctuation range of the best matching cluster, and denot it as the standard deviation of each dimension;

[0197] Calculate the difference between the current feature vector and the reference center vector in each feature dimension to obtain the difference in each dimension;

[0198] Divide the difference between each feature dimension by the standard deviation of the corresponding feature dimension to obtain the standardized difference of that feature dimension.

[0199] The sum of squares of the standardized differences of all feature dimensions is taken, and then the square root of the sum of squares is taken to obtain the standardized deviation between the current feature vector and the reference center vector.

[0200] The calculated standardized deviation is used as the matching deviation output for the process stage.

[0201] In this embodiment, the current feature vector represents the quantitative result of the behavioral pattern characteristics of the inspected sealant within a specific process stage, while the reference center vector represents the typical qualified state of the best-matching cluster in the dynamic statistical reference behavioral pattern library. The standard deviation of each dimension comes from the historical statistics of the best-matching cluster, reflecting the natural fluctuation range of qualified samples of the cluster in the corresponding feature dimension.

[0202] Calculating the difference in each dimension involves subtracting the current feature vector from the reference center vector element by element to obtain the original deviation vector, which directly reflects the absolute difference between the current state and the ideal center in each feature.

[0203] Standardization is achieved by dividing the differences in each dimension by the standard deviation of that dimension. Standardization transforms the original absolute deviation into a relative deviation in units of the standard deviation of each dimension—the standardized difference. The standardized difference eliminates the incomparability caused by differences in the units and magnitudes of variation of different features, allowing for a comprehensive assessment of the deviations of features from different physical meanings on the same scale.

[0204] Taking the sum of squares of the standardized differences across all feature dimensions, and then taking the square root of the sum, essentially calculates the Euclidean modulus of the standardized deviation vector. This modulus is the standardized deviation, a non-negative scalar; the larger its value, the more severe the overall deviation of the current feature vector from the reference center vector in the standardized feature space.

[0205] The final output of the standardized deviation, as the matching deviation of this process stage, provides a dimensionless and comprehensive deviation measure. It not only considers the difference between the current state and the qualified center, but also implicitly considers the significance of the difference relative to historical normal fluctuations by dividing by the standard deviation, thus making the deviation assessment more objective and robust.

[0206] This embodiment introduces standardization based on standard deviation and Euclidean modulus calculation to comprehensively transform multi-dimensional and heterogeneous behavioral characteristic differences into a single, interpretable standardized deviation index. This index effectively overcomes the evaluation bias that may result from using raw distances directly due to inconsistent feature scales. It enables fair and consistent quantitative comparison of behavioral deviations under different process stages and different feature combinations, providing accurate and reliable normalized input for subsequent quality level determination and comprehensive decision-making.

[0207] In some embodiments, by combining the deviation range of process parameters with the material state distribution diagram, a comprehensive evaluation report on sealing quality and process optimization guidance suggestions are output, including:

[0208] The overall behavior pattern deviation sequence is input into the quality assessment rule engine, which predefines the deviation threshold and defect type mapping relationship for different process stages.

[0209] Based on whether the matching deviation of each process stage in the overall behavior pattern deviation sequence exceeds the deviation threshold of the corresponding stage, it is determined whether there is a defect and the process stage to which the defect belongs.

[0210] When a defect is identified, the suspected abnormal process parameters causing the defect are located by combining the deviation range of process parameters. Suspected abnormal process parameters include deviation of dispensing path parameters, deviation of adhesive quantity parameters, deviation of temperature field parameters, or deviation of time parameters.

[0211] At the same time, by combining the material state distribution map, abnormal state areas inside the material can be identified. These abnormal state areas include areas of insufficient curing, stress concentration areas, or areas rich in impurities.

[0212] Based on the process stage to which the defect belongs, the suspected abnormal process parameters, and the abnormal state area inside the material, a quality assessment report is generated that includes the defect type, defect severity level, defect spatial location, root cause process parameter inference, and description of the abnormal material state.

[0213] Based on the root cause process parameters in the quality assessment report, corresponding optimization suggestions are retrieved from the preset process parameter adjustment knowledge base. The optimization suggestions include suggestions on the direction and amount of adjustment for suspected abnormal process parameters.

[0214] The quality assessment report and optimization suggestions are combined to output a comprehensive sealing quality assessment report and process optimization guidance suggestions.

[0215] In this embodiment, the quality assessment rule engine contains a set of predefined logical rules that map the overall behavior pattern deviation sequence characteristics of different process stages (such as the deviation value of a specific stage and the location of the peak) to specific defect types. For example, a high deviation in the leveling stage may be associated with uneven glue line width, and a deviation of a specific pattern in the curing stage may point to bubbles or cracks.

[0216] Deviation thresholds are usually determined based on statistical analysis of the deviation distribution of historical qualified samples. For example, a certain percentile of the historical deviation value in that period can be taken as the threshold, and if it is exceeded, it is judged as abnormal.

[0217] When locating suspected abnormal process parameters, the system cross-references the process stage where the defect occurred with the deviation range of the process parameter. For example, if the defect occurs in the dispensing stage, and the deviation range of the process parameter shows a significant negative deviation in the dispensing path parameter or glue quantity parameter, then that parameter is listed as a suspected abnormal process parameter.

[0218] Identifying abnormal states within a material is achieved through image analysis and data processing of the material state distribution map. For example, in a curing degree field, pixel areas below a set threshold are marked as insufficiently cured areas; in a stress field, stress concentration areas are identified by finding regions with local stress maxima.

[0219] The generated quality assessment report is a structured document that systematically integrates defect type classification, severity level classification based on deviation values, spatial location of defects marked on the sealing image, a list of root cause process parameters inferred from the deviation range of process parameters, and textual descriptions and screenshots of abnormal areas in the material state distribution map.

[0220] The pre-defined process parameter adjustment knowledge base stores optimization rules derived from the experience of process experts or successful historical debugging cases. For example, for the root cause of "discontinuous glue lines due to deviations in dispensing path parameters," the knowledge base may contain "fine-tuning the dispensing path to increase the overlap rate" and specific compensation calculation formulas. The retrieval process involves matching and querying the knowledge base based on the root cause description in the quality assessment report.

[0221] The final output of the sealing quality comprehensive evaluation report and process optimization guidance suggestions is the product of merging and formatting the aforementioned quality evaluation report and the specific optimization suggestions retrieved. Its form can be a comprehensive document containing text, charts and operation instructions or an interactive interface, directly serving the decision-making and process adjustment on the production site.

[0222] This embodiment uses a rule engine to associate quantitative deviations with specific defects, locates the root causes of processes and material anomalies through cross-analysis, and generates targeted optimization suggestions based on a knowledge base, ultimately forming a closed-loop decision output. The comprehensive sealing quality assessment report and process optimization guidance suggestions integrate the multi-source and heterogeneous information (behavioral deviations, parameter deviations, material states) generated in all previous steps and transform it into a comprehensive assessment and clear suggestions that can directly guide production actions. This achieves a complete leap from "detection and diagnosis" to "assessment and decision-making," greatly improving the intelligence level and response efficiency of quality control.

[0223] Please see Figure 3In a second aspect, this embodiment also provides a visual inspection and evaluation system 1 for the sealing quality of a liquid crystal display screen, applicable to the method described in the first aspect. The system includes a multimodal temporal image acquisition module 11, a multidimensional spatiotemporal state map generation module 12, a behavior pattern feature extraction module 13, a process inversion and state analysis module 14, a dynamic reference pattern library management module 15, and a quality assessment and decision output module 16. The multimodal temporal image acquisition module 11 is used to acquire multimodal image data of the sealing process, which includes visible light image sequences, near-infrared spectral image sequences, and three-dimensional line laser scanning point cloud sequences synchronously acquired within the complete process window from dispensing to curing. The multidimensional spatiotemporal state map generation module 12 is used to generate a multidimensional spatiotemporal state map of the sealing area based on the multimodal image data of the sealing process through spatiotemporal registration and data fusion algorithms. The multidimensional spatiotemporal state map includes fused texture features, spectral features, and three-dimensional morphological features at each time point. The behavior pattern feature extraction module 13 is used to perform temporal dynamic feature analysis on the multidimensional spatiotemporal state map and extract... The sealing curing process behavior pattern characteristics are obtained by calculating the spectral intensity change field and three-dimensional morphological displacement field between continuous time frames. The process inversion and state analysis module 14 is used to input the sealing curing process behavior pattern characteristics into the colloidal rheology-thermal curing coupled inversion model. Through inverse solution, it outputs the process parameter deviation range and material state distribution map that lead to the observed behavior pattern. The colloidal rheology-thermal curing coupled inversion model is built based on the forward physical simulation model. The dynamic reference pattern library management module 15 is used to establish and maintain the dynamic statistical reference behavior pattern library. The dynamic statistical reference behavior pattern library generates standard behavior pattern clusters and their allowable fluctuation ranges for each process stage by accumulating historical qualified sealing process time-series multimodal image data and performing cluster analysis. The quality assessment and decision output module 16 is used to match and calculate the deviation of the sealing curing process behavior pattern characteristics of the current inspected sealing area based on the dynamic statistical reference behavior pattern library. Combined with the process parameter deviation range and material state distribution map, it outputs a comprehensive sealing quality assessment report and process optimization guidance suggestions.

[0224] In this embodiment, the multimodal temporal image acquisition module 11 consists of a visible light industrial camera, a near-infrared spectral imager, and a three-dimensional line laser scanner and their synchronization controller, used to capture synchronous multi-source image data covering the entire encapsulation process in real time on the production line. The multidimensional spatiotemporal state map generation module 12, the behavior pattern feature extraction module 13, the process inversion and state analysis module 14, the dynamic reference pattern library management module 15, and the quality assessment and decision output module 16 are jointly deployed in an industrial control computer or server, implementing their corresponding functions through software algorithms. These modules are sequentially connected to form an automated processing pipeline from raw data acquisition to final decision output. The dynamic reference pattern library management module 15 is also responsible for the online updating and version management of the library, ensuring that the reference standard can continuously evolve with the optimization of the production process. The quality assessment and decision output module 16 provides a human-machine interface for report visualization and confirmation and issuance of optimization suggestions. This system realizes online, automatic, in-depth diagnosis and closed-loop quality control of the LCD screen encapsulation and curing process, transforming the method described in the first aspect into stable and reliable industrial testing equipment.

[0225] By adopting the above technical solutions, this invention differs from existing technologies and possesses the following beneficial effects: By acquiring multimodal image data of the sealing process and generating a multidimensional spatiotemporal state map of the sealing area, a multidimensional and continuous characterization of the entire sealing and curing process is achieved. Through temporal dynamic feature analysis of the multidimensional spatiotemporal state map, behavioral pattern characteristics of the sealing and curing process, representing the physicochemical changes, are extracted, thereby elevating observation from static appearance to the dynamic mechanism level. Furthermore, through the inverse solution of the colloidal rheology-thermal curing coupled inversion model, the deviation range of process parameters leading to the observed behavioral pattern and the distribution map of the internal material state can be deduced from the observed behavioral pattern, establishing a quantitative correlation between abnormal phenomena, process root causes, and the internal state of the material. Simultaneously, a dynamic statistical reference behavioral pattern library based on historical qualified data provides an intelligent matching and deviation calculation benchmark for the behavioral pattern of the currently inspected sealing area. Finally, combining the deviation range of process parameters, the material state distribution map, and the deviation of the behavioral pattern, a comprehensive sealing quality assessment report and process optimization guidance suggestions, including specific defect diagnosis, root cause inference, and targeted adjustment suggestions, are output. The above technical solution overcomes the limitations of existing technologies in terms of insufficient analysis of the dynamic process of sealing and difficulty in tracing the root causes of the process. It realizes closed-loop quality control from deep process perception, mechanism inversion diagnosis to intelligent decision output, and significantly improves the accuracy, foresight and process optimization capabilities of LCD screen sealing quality inspection.

[0226] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0227] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0228] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A visual inspection and evaluation method for the sealing quality of a liquid crystal display screen, characterized in that, include: Acquire multimodal image data of the sealing process, which includes visible light image sequences, near-infrared spectral image sequences, and three-dimensional line laser scanning point cloud sequences synchronously acquired within the complete process window from dispensing to curing; Based on the temporal multimodal image data of the sealing process, a multidimensional spatiotemporal state map of the sealing area is generated through spatiotemporal registration and data fusion algorithms. The multidimensional spatiotemporal state map includes fused texture features, spectral features and three-dimensional morphological features at each time point. The multidimensional spatiotemporal state map is subjected to temporal dynamic feature analysis to extract the behavior pattern features of the sealing and curing process. The behavior pattern features of the sealing and curing process are obtained by calculating the spectral intensity change field and the three-dimensional morphology displacement field between consecutive time frames. The characteristics of the sealing and curing process behavior pattern are input into the colloidal rheology-thermal curing coupled inversion model. The process parameter deviation range and material state distribution map that lead to the observed behavior pattern are output through inverse solution. The colloidal rheology-thermal curing coupled inversion model is built on the basis of the forward physical simulation model. A dynamic statistical reference behavior pattern library is established. The dynamic statistical reference behavior pattern library generates standard behavior pattern clusters and their allowable fluctuation ranges for each process stage by accumulating historically qualified sealing process time-series multimodal image data and performing cluster analysis. Based on the dynamic statistical reference behavior pattern library, the behavior pattern characteristics of the sealing and curing process of the current inspected sealing area are matched and the deviation is calculated. Combined with the process parameter deviation range and material state distribution map, a comprehensive sealing quality evaluation report and process optimization guidance suggestions are output. The colloidal rheology-thermosetting coupled inversion model is constructed through the following steps: A forward physical simulation model of colloidal rheology-thermal curing is established. The inputs of the colloidal rheology-thermal curing forward physical simulation model include dispensing path parameters, adhesive quantity parameters, temperature field parameters, and time parameters. The output of the colloidal rheology-thermal curing forward physical simulation model is the predicted multidimensional spatiotemporal state map of the sealing area. The colloidal rheology-thermosetting forward physics simulation model is encapsulated as a differentiable computational graph, making the output of the colloidal rheology-thermosetting forward physics simulation model differentiable from the input parameters of the colloidal rheology-thermosetting forward physics simulation model. Using the behavioral pattern characteristics of the sealing and curing process as observation constraints, the difference between the predicted output of the colloidal rheology-thermal curing forward physical simulation model and the observation constraints is constructed as a loss function; The loss function is iteratively optimized using a gradient descent optimization algorithm. The gradient is backpropagated through the differentiable computation graph to adjust the input parameters of the colloidal rheology-thermal curing forward physical simulation model until the loss function converges. The deviation between the input parameters of the optimized colloidal rheology-thermocuring forward physical simulation model and the standard process parameters is output as the process parameter deviation range. Simultaneously, the internal material state distribution obtained by running the colloidal rheology-thermocuring forward physical simulation model under optimized input parameters is output as the material state distribution map.

2. The visual inspection and evaluation method for the sealing quality of a liquid crystal screen according to claim 1, characterized in that, Based on the temporal multimodal image data of the sealing process, a multidimensional spatiotemporal state map of the sealing region is generated through spatiotemporal registration and data fusion algorithms, including: The visible light image sequence, the near-infrared spectral image sequence, and the three-dimensional line laser scanning point cloud sequence are time-stamped and spatially aligned to form a unified spatial coordinate system, generating a time-synchronized and spatially aligned registered multimodal data stream; For each time frame in the registered multimodal data stream, perform the following fusion operation: Extract the visible light image of the corresponding time frame from the visible light image sequence, calculate the local texture descriptor of the visible light image, and generate a texture feature map; Extract the near-infrared spectral images of the corresponding time frames from the near-infrared spectral image sequence, select multiple preset characteristic spectral bands, calculate the reflection intensity of each characteristic spectral band, and generate a spectral feature map; Extract the three-dimensional point cloud of the corresponding time frame from the three-dimensional line laser scanning point cloud sequence, project the three-dimensional point cloud onto the same imaging plane as the visible light image, calculate the elevation statistics of the region corresponding to each pixel after projection, and generate a three-dimensional topography feature map. The texture feature map, the spectral feature map, and the three-dimensional topography feature map are aligned pixel by pixel in space, and the texture feature value, the reflection intensity value of multiple feature spectral bands, and the three-dimensional topography elevation statistics value at each spatial pixel position are combined into a multi-dimensional fusion feature vector. Based on the multidimensional fusion feature vectors at all spatial pixel locations, construct the fusion feature vector field for the current time frame; The fused feature vector fields of each time frame are arranged and indexed in chronological order to form a multidimensional spatiotemporal state map of the sealing region. The structure of the multidimensional spatiotemporal state map enables the multidimensional fused feature vectors to be queried and sliced ​​for analysis according to time dimension and spatial location.

3. The visual inspection and evaluation method for the sealing quality of a liquid crystal screen according to claim 2, characterized in that, The multidimensional spatiotemporal state map is subjected to temporal dynamic feature analysis to extract behavioral pattern features of the sealing and curing process, including: From the multidimensional spatiotemporal state map, extract the fused feature vector field of consecutive time frames in chronological order; For each spatial cell location in the fused feature vector field, perform the following calculation: Calculate the difference in texture feature values ​​between adjacent time frames for the spatial pixel location to obtain the texture feature change gradient; Extract the reflection intensity values ​​of multiple characteristic spectral bands of the spatial pixel location in consecutive time frames to form a time series of each characteristic spectral band, calculate the first difference of the time series, and obtain the spectral intensity change sequence. Calculate the displacement vector of the spatial pixel location in the statistical value of the three-dimensional topography elevation between adjacent time frames to obtain the displacement of the three-dimensional topography change. Based on the texture feature change gradient, the spectral intensity change sequence, and the three-dimensional morphology change displacement of all spatial pixel locations, a dynamic change field of the sealing region within the observation time window is constructed. The dynamic change field includes the texture change field, the spectral intensity change field, and the three-dimensional morphology displacement field. Pattern recognition analysis is performed on the dynamic field to extract the behavioral pattern features of the sealing curing process, which characterize the physical process of sealing curing. These behavioral pattern features include the distribution pattern of the colloid edge retraction velocity, the internal bubble movement trajectory pattern, the local spectral curing rate anomaly pattern, and the three-dimensional morphology shrinkage non-uniformity pattern.

4. The visual inspection and evaluation method for the sealing quality of a liquid crystal screen according to claim 1, characterized in that, The behavioral characteristics of the sealing and curing process are input into the colloidal rheology-thermal curing coupled inversion model. Through inverse solving, the model outputs the process parameter deviation ranges and material state distribution maps that lead to the observed behavioral patterns, including: The behavioral pattern characteristics of the sealing and curing process are used as the observation constraint input for the colloidal rheology-thermal curing coupled inversion model; Within the colloidal rheology-thermal curing coupled inversion model, the observation constraints are quantified into target feature vectors corresponding to the output of the colloidal rheology-thermal curing forward physical simulation model; The gradient descent optimization algorithm is executed to iteratively adjust the dispensing path parameters, the adhesive amount parameters, the temperature field parameters, and the time parameters, so as to minimize the loss function value between the predicted feature vector output by the colloidal rheology-thermal curing forward physical simulation model and the target feature vector; When the loss function converges, record the optimized values ​​of the dispensing path parameters, the glue quantity parameters, the temperature field parameters, and the time parameters at this time. Calculate the difference between the optimized value and the preset standard process parameter value, and output the confidence interval of the difference as the process parameter deviation interval; Simultaneously, internal state variables are extracted from the converged colloidal rheology-thermal curing forward physical simulation model. These internal state variables include the colloidal viscosity field, curing degree field, and stress field. The internal state variables are then mapped to spatial locations to generate the material state distribution map.

5. The visual inspection and evaluation method for the sealing quality of a liquid crystal screen according to claim 1, characterized in that, Establish a dynamic statistical reference behavior pattern library, including: Collect multimodal image data of the encapsulation process of LCD screens that have been judged to be qualified through offline testing from historical production batches; Based on the collected sealing process time-series multimodal image data, a corresponding set of historical qualified sealing curing process behavior patterns is generated; The set of historical qualified sealing and curing process behavior patterns is divided into time slices according to the process stages, which include the dispensing stage, the leveling stage, and the curing stage. For the historical qualified sealing and curing process behavior pattern characteristics within each process stage, a clustering algorithm is used to classify the patterns and generate multiple standard behavior pattern clusters for that process stage. For each of the standard behavior pattern clusters, the mean and standard deviation of all historical qualified sealant curing behavior pattern features within the standard behavior pattern cluster are calculated in each feature dimension. The mean is used as the center of the standard behavior pattern cluster, and the standard deviation is used as a preset multiple as the allowable fluctuation range of the standard behavior pattern cluster. The standard behavior pattern clusters and their corresponding allowable fluctuation ranges are organized and stored according to process stages to form the dynamic statistical reference behavior pattern library.

6. The visual inspection and evaluation method for the sealing quality of a liquid crystal screen according to claim 1, characterized in that, Based on the aforementioned dynamic statistical reference behavior pattern library, the sealing and curing process behavior pattern characteristics of the currently inspected sealing area are matched and the deviation is calculated, including: The sealing and curing process behavior pattern characteristics of the currently inspected sealing area are divided into the same process stages as those in the dynamic statistical reference behavior pattern library according to the time dimension. For each of the aforementioned process stages, retrieve all standard behavior pattern clusters corresponding to that process stage and their allowable fluctuation ranges from the dynamic statistical reference behavior pattern library; Calculate the feature vector of the sealing curing process behavior pattern of the currently inspected sealing area within this process stage, and the feature space distance between it and the central feature vector of each retrieved standard behavior pattern cluster; The cluster of standard behavioral patterns with the smallest feature space distance is selected as the best matching cluster; Determine whether the feature vector of the sealing curing process behavior pattern of the currently inspected sealing area falls within the allowable fluctuation range of the best matching cluster in this process stage; If it falls within the allowable fluctuation range, then the standardized deviation between the feature vector of the sealing curing process behavior pattern of the current inspected sealing area and the central feature vector of the best matching cluster is calculated as the matching deviation of this process stage. If it does not fall within the allowable fluctuation range, it is determined as no match, and the cross-boundary distance from the feature vector of the sealing curing process behavior pattern of the current inspected sealing area to the allowable fluctuation range boundary of the best matching cluster is calculated as the matching deviation of this process stage. The matching deviations of all process stages are combined in chronological order to generate the overall behavior pattern deviation sequence of the currently inspected sealing area.

7. The visual inspection and evaluation method for the sealing quality of a liquid crystal display screen according to claim 6, characterized in that, The standardized deviation between the feature vector of the sealing curing process behavior pattern of the currently inspected sealing area and the central feature vector of the best matching cluster is calculated as the matching deviation for this process stage, including: Extract the feature vector of the sealing curing process behavior pattern of the currently inspected sealing area within the process stage, and denote it as the current feature vector; Extract the center feature vector of the best matching cluster, and denote it as the reference center vector; Extract the standard deviation corresponding to each feature dimension within the allowable fluctuation range of the best matching cluster, and denot it as the standard deviation of each dimension; Calculate the difference between the current feature vector and the reference center vector in each feature dimension to obtain the difference in each dimension; Divide the difference between each feature dimension by the standard deviation of the corresponding feature dimension to obtain the standardized difference of that feature dimension. The sum of squares of the standardized differences for all feature dimensions is taken, and then the square root of the sum of squares is taken to obtain the standardized deviation between the current feature vector and the reference center vector. The calculated standardized deviation is output as the matching deviation of the process stage.

8. The visual inspection and evaluation method for the sealing quality of a liquid crystal display screen according to claim 6, characterized in that, Based on the aforementioned process parameter deviation range and material state distribution diagram, a comprehensive evaluation report on sealing quality and process optimization guidance suggestions are output, including: The overall behavior pattern deviation sequence is input into the quality assessment rule engine, which predefines the deviation threshold and defect type mapping relationship for different process stages. Based on whether the matching deviation of each process stage in the overall behavior pattern deviation sequence exceeds the deviation threshold of the corresponding stage, it is determined whether there is a defect and the process stage to which the defect belongs. When a defect is determined, the suspected abnormal process parameters causing the defect are located by combining the process parameter deviation range. The suspected abnormal process parameters include dispensing path parameter deviation, glue quantity parameter deviation, temperature field parameter deviation, or time parameter deviation. At the same time, based on the material state distribution map, abnormal state regions inside the material are identified, including regions of insufficient curing, stress concentration regions, or impurity enrichment regions. Based on the process stage to which the defect belongs, the suspected abnormal process parameters, and the abnormal state region inside the material, a quality assessment report is generated that includes the defect type, defect severity level, defect spatial location, root cause process parameter inference, and description of the abnormal material state. Based on the root cause process parameters in the quality assessment report, corresponding optimization suggestions are retrieved from a preset process parameter adjustment knowledge base. The optimization suggestions include suggestions on the direction and amount of adjustment for the suspected abnormal process parameters. The quality assessment report and the optimization suggestions are combined to output the comprehensive sealing quality assessment report and process optimization guidance suggestions.

9. A visual inspection and evaluation system for the sealing quality of a liquid crystal display screen, characterized in that, The system applicable to the method of any one of claims 1 to 8, the system comprising: The multimodal time-series image acquisition module is used to acquire multimodal image data of the sealing process. The multimodal image data of the sealing process includes visible light image sequences, near-infrared spectral image sequences, and three-dimensional line laser scanning point cloud sequences that are acquired synchronously within the complete process window from dispensing to curing. The multidimensional spatiotemporal state map generation module is used to generate a multidimensional spatiotemporal state map of the sealing area based on the temporal multimodal image data of the sealing process, through spatiotemporal registration and data fusion algorithms. The multidimensional spatiotemporal state map includes fused texture features, spectral features and three-dimensional morphological features at each time point. The behavior pattern feature extraction module is used to perform temporal dynamic feature analysis on the multidimensional spatiotemporal state map and extract the behavior pattern features of the sealing and curing process. The behavior pattern features of the sealing and curing process are obtained by calculating the spectral intensity change field and the three-dimensional morphology displacement field between consecutive time frames. The process inversion and state analysis module is used to input the behavioral pattern characteristics of the sealing and curing process into the colloidal rheology-thermal curing coupled inversion model, and output the process parameter deviation range and material state distribution map that lead to the observed behavioral pattern through inverse solution. The colloidal rheology-thermal curing coupled inversion model is built on the basis of the forward physical simulation model. The dynamic reference pattern library management module is used to establish and maintain a dynamic statistical reference behavior pattern library. The dynamic statistical reference behavior pattern library generates standard behavior pattern clusters and their allowable fluctuation ranges for each process stage by accumulating historically qualified sealing process time-series multimodal image data and performing cluster analysis. The quality assessment and decision output module is used to match and calculate the deviation of the sealing curing process behavior pattern characteristics of the current inspected sealing area based on the dynamic statistical reference behavior pattern library. Combined with the process parameter deviation range and material state distribution map, it outputs a comprehensive sealing quality assessment report and process optimization guidance suggestions.

Images