Intelligent scheduling management method and system for liquid crystal display screen production laminating equipment

Abstract

This invention relates to the field of production management technology and discloses an intelligent scheduling and management method and system for liquid crystal display (LCD) production lamination equipment. The method includes: step S1, acquiring lamination tasks of the target management equipment; after any lamination task is completed, collecting physicochemical data of the target management equipment in real time; analyzing the physicochemical data based on convolutional neural networks and multilayer perceptrons to generate residual environmental vectors; calculating adhesive environmental management values; generating single-task transition management values ​​and global risk accumulation management values ​​through long short-term memory networks; obtaining scheduling management index and dynamic management threshold by combining lamination equipment utilization rate; and finally generating management instructions by comparison. This can avoid interference from different adhesive flow levels, reduce the occurrence of LCD alignment slippage and local curing defects, and achieve the effect of scheduling and managing the lamination equipment.

Description

Technical Field

[0001] This invention relates to the field of production management technology, specifically to an intelligent scheduling and management method and system for liquid crystal display screen production lamination equipment. Background Technology

[0002] When existing LCD screens are manufactured using lamination equipment, optimization models are often built with time as the single dimension. Traditional optimization algorithms such as genetic algorithms and particle swarm optimization are used to minimize the production cycle or maximize equipment utilization, generating static task sequences. This scheduling and management method assumes that all tasks executed within the equipment time window are independent and equivalent, thus achieving intelligent management centered on efficiency.

[0003] However, the above-mentioned scheduling and management methods still have the following drawbacks when dealing with glass bonding tasks of various materials: Since the adhesives required for different glass materials have significant differences in viscosity, curing rate, and coefficient of thermal expansion, the process parameters required for the bonding equipment are different. For example, when the existing neural network algorithm schedules and manages the task, it may schedule a task that requires high viscosity and slow curing adhesive and a task that requires low viscosity and fast curing adhesive to be scheduled close together on the same bonding equipment. The leveling characteristics of the two adhesives will interfere with each other in the cavity of the bonding equipment. Even after basic cleaning, residual chemical molecules or unstable temperature fields will change the fluid dynamics environment during bonding, generating uneven stress on the precisely positioned glass, which will eventually lead to misalignment or poor local curing, resulting in a decrease in bonding quality and affecting the scheduling and management effect. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent scheduling and management method and system for liquid crystal display manufacturing lamination equipment, thus solving the aforementioned problems.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0006] A method for intelligent scheduling and management of LCD screen manufacturing lamination equipment includes:

[0007] include:

[0008] Step S1: Obtain the bonding task of the target management device. After any bonding task is completed, collect the physicochemical data of the target management device in real time. Analyze the physicochemical data based on the convolutional neural network and multilayer perceptron to generate a residual environmental vector. The target management device is the bonding device for the liquid crystal display screen.

[0009] Step S2: For each bonding task to be scheduled, the glue data of the bonding task is obtained in real time, and the glue data is calculated based on the back propagation algorithm to obtain the glue environment management value of the bonding task.

[0010] Step S3: Based on the residual environment vector of the target management object and the glue environment management value of any bonding task to be scheduled, after the bonding task is executed by the target management device, the risk of the process result deviating from the quality standard is analyzed based on the long short-term memory network, and the single task transition management value of the bonding task is generated.

[0011] Step S4: The single-task transition management values ​​of multiple matching tasks to be scheduled are imported into the long short-term memory network for analysis in the scheduling order to generate a global risk accumulation management value.

[0012] Step S5: Taking minimizing the global risk accumulation management value as the core optimization objective, the utilization rate and production cycle of the target management equipment are analyzed based on a neural network algorithm that combines convolutional neural network and multilayer perceptron, and the scheduling management index is obtained.

[0013] Step S6: Analyze the glue environment management value and global risk accumulation management value of all bonding tasks to be scheduled based on the neural network algorithm that combines long short-term memory network and convolutional neural network, and generate dynamic management threshold.

[0014] Step S7: Compare the scheduling management index with the dynamic management threshold to obtain management instructions.

[0015] Furthermore, based on convolutional neural networks and multilayer perceptrons, physicochemical data are analyzed to generate residual environmental vectors, including:

[0016] Spatiotemporal coupling analysis was performed on the preprocessed physicochemical data. By calculating the dynamic correlation between temperature, air pressure and residual glue molecule concentration at different times, a feature management dataset was generated.

[0017] The feature management dataset is input into a convolutional neural network to calculate the contribution weight of each feature to the glue residue analysis. Then, based on the contribution weight, a convolution operation is performed on the high contribution features to obtain the directional feature map.

[0018] The directional feature map and contribution weight are input together into the multilayer perceptron, so that high-contribution features maintain a higher influence in the multilayer mapping and generate a fusion feature management vector.

[0019] The fusion feature management vector is separated to obtain residual feature sub-vectors, thermodynamic feature sub-vectors, and surface energy feature sub-vectors. Then, based on the inherent laws of the three sub-vectors, a residual environment vector containing three dimensions of glue molecule residual concentration, thermodynamic equilibrium offset, and surface energy gradient is generated.

[0020] Furthermore, based on the backpropagation algorithm, the glue data is calculated to obtain the glue environment management values ​​for this bonding task, including:

[0021] Obtain historical qualified process parameters for the target management equipment;

[0022] The pre-processed adhesive data is compared with the historical qualified process parameters of the target management equipment to obtain the process scenario adaptability.

[0023] The dynamic decay law of adhesive viscosity with curing process and the inhibitory effect of thermal expansion coefficient on curing rate are analyzed to generate dynamic interference values ​​of parameters.

[0024] Based on the adaptability to the process scenario, feature scaling is applied to the dynamic interference values ​​of the parameters to generate a scenario-based feature enhancement matrix;

[0025] The scenario-based feature enhancement matrix is ​​analyzed based on the backpropagation algorithm. The error propagation path of backpropagation is adjusted with the parameter dynamic interference value as the weight to generate the initial value for glue environment management.

[0026] The initial values ​​for adhesive environment management are calibrated based on the dynamic interference values ​​of parameters and the adaptability of the process scenario, and then the adhesive environment management values ​​are generated.

[0027] Furthermore, based on the residual environment vector of the target management object and the glue environment management value of any bonding task to be scheduled, after the bonding task is executed by the target management device, the risk of the process results deviating from the quality standard is analyzed based on the long short-term memory network, and a single task transition management value for the bonding task is generated, including:

[0028] The residual environment vector and the adhesive environment management value are dynamically interacted to calculate the real-time adaptation loss degree of each dimension parameter and adhesive properties during the bonding process, and generate the ring-adhesive interaction loss coefficient.

[0029] When the cross-loss coefficient of the ring adhesive exceeds the critical management value, the synergistic failure threshold caused by the high residual adhesive concentration leading to a decrease in the curing rate of the new adhesive and the process abnormality threshold caused by the conflict between the thermodynamic balance offset and the thermal expansion coefficient of the adhesive are analyzed separately. The synergistic failure threshold and the process abnormality threshold are integrated to generate a multi-dimensional risk trigger threshold.

[0030] The cyclic adhesive interaction loss coefficient, multidimensional risk trigger threshold, residual environmental vector and adhesive environmental management value are sorted according to the bonding process time sequence, integrated into a time-series risk feature sequence, and imported into a long short-term memory network for analysis to generate a preliminary risk assessment value and risk transition probability value for a single task.

[0031] The initial risk assessment value for a single task is calibrated based on the risk transition probability value and the multi-dimensional risk trigger threshold to generate a single task transition management value.

[0032] Furthermore, the single-task transition management values ​​of multiple matching tasks to be scheduled are imported into the Long Short-Term Memory network for analysis in scheduling order to generate a global risk accumulation management value, including:

[0033] Extract the single-task transition management value of each matching task to be scheduled, analyze the risk transmission effect of the preceding matching task on multiple subsequent matching tasks according to the scheduling order, and generate the cross-order risk management attenuation coefficient.

[0034] Obtain the process duration of each matching task to be scheduled, and calculate the single task transition management value based on the process duration and the cross-sequence risk management attenuation coefficient to generate a time-varying management risk value.

[0035] Calculate the adaptation deviation between each time-varying management risk value and the residual environmental vector of the current target management device, and generate the risk-environment dynamic adaptation deviation rate.

[0036] Furthermore, the single-task transition management values ​​of multiple matching tasks to be scheduled are imported into the Long Short-Term Memory network for analysis in scheduling order to generate a global risk accumulation management value, which also includes:

[0037] The time-varying risk value, cross-sequence risk management attenuation coefficient, and risk-environment dynamic adaptation deviation rate are integrated according to the scheduling order. At the same time, the process type that fits the task is distinguished, and a dynamic correlation risk feature sequence containing four types of dynamic information, namely risk, transmission, environment, and process, is constructed.

[0038] The dynamic correlation risk feature sequence is imported into a long short-term memory network for analysis to generate risk management factors and preliminary cumulative values ​​of global risk.

[0039] The initial cumulative value of global risk is calibrated based on risk management factors and cross-order risk management attenuation coefficient to generate a two-dimensional cumulative risk value.

[0040] Analyze the risk contribution of each relevant task in the two-dimensional risk cumulative value, generate the risk cascading effect coefficient, adjust the two-dimensional risk cumulative value according to the risk cascading effect coefficient, and generate the global risk cumulative management value.

[0041] Furthermore, taking minimizing the cumulative global risk management value as the core optimization objective, a neural network algorithm based on the collaboration of convolutional neural networks and multilayer perceptrons is used to analyze the utilization rate and production cycle of the target management equipment, deriving a scheduling management index, including:

[0042] Obtain the time utilization rate of the target management equipment, the priority weight of the tasks to be scheduled, and the production cycle;

[0043] With minimizing the global risk accumulation management value as the core constraint, the dynamic relationship between time series utilization, priority weight and risk accumulation is analyzed to obtain the risk-efficiency-priority dynamic management coefficient;

[0044] The risk-efficiency-priority dynamic management coefficient, time series utilization rate, and production cycle are integrated in layers to generate a layered time series feature matrix. The layered time series feature matrix is ​​then imported into a convolutional neural network for calculation to generate the core feature vector of time series management.

[0045] The core feature vector of time series management is input into the multilayer perceptron and analyzed in combination with the risk-efficiency-priority dynamic management coefficient to generate a dynamic calibration efficiency evaluation value.

[0046] With the goal of minimizing the cumulative management value of global risk, the dynamic calibration effectiveness evaluation value is anchored and calibrated to obtain the scheduling management index.

[0047] Furthermore, based on a neural network algorithm that combines Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs), the glue environment management value and global risk accumulation management value of all scheduled bonding tasks are analyzed to generate dynamic management thresholds, including:

[0048] Obtain the process sensitivity attribute that fits the task;

[0049] Based on the glue environment management value, global risk accumulation management value and process sensitivity attribute of all bonding tasks to be scheduled, the mutual influence intensity of glue environment fluctuation and risk accumulation under different process sensitivity attributes is calculated to generate glue sensitivity management coefficient.

[0050] The glue sensitivity management coefficient, glue environment management value, and global risk cumulative management value are integrated into a risk-environment-sensitivity time series according to the scheduling order, and then imported into a long short-term memory network to calculate the actual contribution of each dimension to the global risk and generate time series risk contribution weights.

[0051] The time-series risk contribution weight, glue sensitivity management coefficient, and risk-environment-sensitivity time-series sequence are input into a convolutional neural network for calculation to generate a key feature-sensitivity matching map. The key feature-sensitivity matching map is then analyzed to generate a dynamic management threshold.

[0052] Furthermore, the scheduling management index is compared with the dynamic management threshold to obtain management instructions, including:

[0053] If the scheduling management index is greater than or equal to the dynamic management threshold, the scheduling management is deemed qualified, and a continued fitting instruction is issued.

[0054] If the scheduling management index is less than the dynamic management threshold, the scheduling management is deemed unqualified, and an alarm command is issued.

[0055] Furthermore, an intelligent scheduling and management system for LCD screen manufacturing lamination equipment, applied to the aforementioned intelligent scheduling and management method for LCD screen manufacturing lamination equipment, includes:

[0056] The integration unit is used to acquire the bonding task of the target management device. After any bonding task is completed, the physical and chemical data of the target management device are collected in real time. The physical and chemical data are analyzed based on the convolutional neural network and multilayer perceptron to generate a residual environmental vector. The target management device is the bonding device of the LCD screen.

[0057] The environmental management unit is used to acquire the glue data of each bonding task to be scheduled in real time, and calculate the glue environmental management value of the bonding task based on the backpropagation algorithm.

[0058] The deviation management unit is used to generate a single task transition management value for the bonding task after it is executed by the target management equipment, based on the residual environment vector of the target management object and the glue environment management value of any bonding task to be scheduled, and the risk of the process results deviating from the quality standard based on the analysis of the long short-term memory network.

[0059] The risk management unit is used to import the single-task transition management values ​​of multiple matching tasks to be scheduled into the long short-term memory network for analysis in the order of scheduling, and generate a global risk accumulation management value.

[0060] The optimization management unit is used to analyze the utilization rate and production cycle of the target management equipment based on a neural network algorithm that combines convolutional neural networks and multilayer perceptrons, with the core optimization objective of minimizing the cumulative global risk management value, and to derive the scheduling management index.

[0061] The threshold calculation unit is used to analyze the glue environment management value and global risk accumulation management value of all bonding tasks to be scheduled based on a neural network algorithm that combines long short-term memory network and convolutional neural network, and generate dynamic management thresholds.

[0062] The scheduling management unit is used to compare the scheduling management index with the dynamic management threshold to obtain management instructions.

[0063] In summary, the present invention has the following main beneficial effects:

[0064] In step S1, by collecting physicochemical data of the LCD bonding equipment in real time, key features of residual adhesive are extracted based on convolutional neural networks. Combined with multilayer perceptron, the data is integrated to generate a residual environment vector containing residual adhesive molecule concentration, thermodynamic equilibrium offset, and surface energy gradient. This accurately captures the internal environment of the equipment and avoids adhesive residue interference caused by lack of environmental perception. In step S2, based on the backpropagation algorithm, the adhesive environment management value is calculated by combining the historical qualified process parameters of the target management equipment. This helps to understand the compatibility between the adhesive of the scheduled task and the equipment process, ensuring that the adhesive required for different glass materials can match the basic process conditions of the equipment. In step S3, relying on long short-term memory networks, the interaction risk between the residual environment vector and the adhesive environment management value is analyzed to generate a single task transition management value. This helps to identify the process deviation risk during the execution of a single bonding task in advance and avoid LCD alignment slippage caused by task and environment mismatch.

[0065] Step S4 imports the single-task transition management value according to the scheduling order, and generates a global risk accumulation management value by combining the cross-sequence risk management attenuation coefficient, process duration, and residual environment vector. This quantifies the risk transmission and superposition effect between multiple tasks, solving the problem of glue interference caused by traditional scheduling. Step S5 focuses on minimizing the global risk accumulation management value. It combines convolutional neural networks and multilayer perceptrons to analyze equipment time utilization, production cycle, and task priority weights to generate a scheduling management index, achieving a balance between risk control and efficiency management. Step S6 generates a dynamic management threshold based on a long short-term memory network and convolutional neural network collaborative algorithm, combined with the bonding task process sensitivity attributes, to accurately distinguish between normal and high-risk scheduling scenarios. Step S7 outputs instructions by comparing the scheduling management index with the dynamic management threshold. If the result is satisfactory, bonding continues; if not, an alarm is triggered, ensuring that each round of scheduling meets the bonding quality requirements of the LCD screen. The overall solution can effectively reduce interference from different glue leveling characteristics, residual chemical molecules, or unstable temperature fields that cause local curing defects in the LCD screen, providing global scheduling management for the LCD screen bonding equipment. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the steps of an intelligent scheduling and management method for liquid crystal display screen manufacturing lamination equipment according to the present invention.

[0067] Figure 2 This is a schematic diagram of an intelligent scheduling and management system for a liquid crystal display screen manufacturing bonding equipment according to the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] refer to Figure 1 and Figure 2 A method for intelligent scheduling and management of LCD screen manufacturing lamination equipment, comprising:

[0070] Step S1: Obtain the bonding task of the target management device. After any bonding task is completed, collect the physical and chemical data of the target management device in real time. Analyze the physical and chemical data based on the convolutional neural network and multilayer perceptron to generate a residual environmental vector. The target management device is the bonding device of the liquid crystal display screen.

[0071] Physicochemical data include: temperature and air pressure inside the bonding area of ​​the LCD screen bonding equipment, concentration of residual adhesive molecules, etc.

[0072] Step S2: For each bonding task to be scheduled, the glue data of the bonding task is obtained in real time. The glue data is calculated based on the backpropagation algorithm to obtain the glue environment management value of the bonding task. The glue data includes: viscosity, curing rate, coefficient of thermal expansion, etc.

[0073] Step S3: Based on the residual environment vector of the target management object and the glue environment management value of any bonding task to be scheduled, after the bonding task is executed by the target management device, the risk of the process result deviating from the quality standard is analyzed based on the long short-term memory network, and the single task transition management value of the bonding task is generated.

[0074] Step S4: The single-task transition management values ​​of multiple matching tasks to be scheduled are imported into the long short-term memory network for analysis in the scheduling order to generate a global risk accumulation management value.

[0075] Step S5: Taking minimizing the global risk accumulation management value as the core optimization objective, the utilization rate and production cycle of the target management equipment are analyzed based on a neural network algorithm that combines convolutional neural network and multilayer perceptron, and the scheduling management index is obtained.

[0076] Step S6: Analyze the glue environment management value and global risk accumulation management value of all bonding tasks to be scheduled based on the neural network algorithm that combines long short-term memory network and convolutional neural network, and generate dynamic management threshold.

[0077] Step S7: Compare the scheduling management index with the dynamic management threshold to obtain management instructions.

[0078] By collecting physicochemical data from target management equipment in real time, extracting key features of residual adhesive using convolutional neural networks, and integrating the data with a multilayer perceptron, a residual environment vector is generated, including residual adhesive molecule concentration, thermodynamic equilibrium shift, and surface energy gradient. Then, the adhesive environment management value for the task to be scheduled is calculated based on the backpropagation algorithm. The single task transition management value and the global risk accumulation management value are analyzed through a long short-term memory network. The equipment utilization rate and production cycle are optimized by combining convolutional neural networks and multilayer perceptron collaborative algorithms to obtain the scheduling management index and generate a dynamic management threshold. Finally, the management command is output by comparing the two. This can avoid the interference of different adhesive leveling characteristics, alignment slip and local curing defects caused by unstable residual molecules or temperature fields. While taking into account the utilization rate of bonding equipment and the production cycle, the bonding quality is guaranteed, and the scheduling management effect is greatly improved.

[0079] In one embodiment, the analysis of physicochemical data based on convolutional neural networks and multilayer perceptrons generates a residual environment vector, including:

[0080] Spatiotemporal coupling analysis was performed on the preprocessed physicochemical data. By calculating the dynamic correlation between temperature, air pressure, and residual glue molecule concentration at different times, a feature management dataset was generated, which specifically includes:

[0081] The pre-processed physicochemical data are time-series records of multiple spatial monitoring points of the target management equipment, such as the bonding operation area, curing chamber, and material transport track. Each monitoring point collects data such as temperature, air pressure, and residual glue concentration at a frequency of 10 times per second, and each data is accompanied by a timestamp to mark the collection time.

[0082] During the analysis, a sliding time window was constructed based on a set length of 60 seconds. The window moved along the time axis in 30-second increments, ensuring that adjacent windows overlapped by 50% to maintain the continuity of the time series analysis. Subsequently, for the time series data of temperature and residual glue concentration within each window, lag correlation calculation was performed using cross-correlation analysis. Specifically, within a lag range of 0 to 300 seconds, a lag interval of 10 seconds was taken, and the cross-correlation coefficient between the temperature series and the concentration series after shifting the lag interval was calculated. Then, the lag value corresponding to the largest absolute value of the cross-correlation coefficient was selected from all the calculation results as the optimal correlation lag between temperature and concentration within that window, and this maximum value was recorded simultaneously as its correlation strength.

[0083] The data on air pressure and residual glue concentration are processed using the same calculation method as for temperature and concentration to obtain the optimal correlation lag and corresponding correlation strength between air pressure and concentration. Each window outputs these two sets of lag and strength data. The three-dimensional coordinates (x-axis, y-axis, z-axis positions) of each monitoring point are used as spatial attributes, and the optimal temperature-concentration lag, optimal air pressure-concentration lag, temperature-concentration correlation strength, and air pressure-concentration correlation strength of each monitoring point are used as feature parameters.

[0084] Using the three-dimensional coordinates (x-axis, y-axis, z-axis positions) of each monitoring point as a reference, the straight-line distance between any two monitoring points is calculated. This straight-line distance is the Euclidean distance in three-dimensional space. A neighborhood radius of 5 cm is set. For any given monitoring point, all other monitoring points within a 5 cm radius centered on the point itself are considered spatially adjacent to it. Simultaneously, a minimum clustering threshold of 3 points is set; a monitoring point is considered a core point only if there are at least 3 monitoring points within its 5 cm neighborhood (including itself). During clustering, starting with the first core point, all monitoring points within its 5 cm neighborhood are included in the same initial cluster. Then, these included clusters are checked. If a point is a core point, it continues to include other unclassified points within its 5-cm neighborhood into the cluster, recursively expanding until no new points can be added, forming a complete cluster. This logic is used to traverse all monitoring points, ultimately resulting in several clusters, each containing at least 3 monitoring points. Each cluster represents a spatial association pattern, and a unique identifier is assigned to each cluster, using the numbers 1, 2, and 3 in sequence. All monitoring points within the cluster are assigned the corresponding identifier as spatial pattern labels. If a monitoring point has fewer than 3 points within its 5-cm neighborhood and is not within the neighborhood of any core point, it is considered a noise point and is not included in any cluster.

[0085] The feature management dataset is obtained by integrating the start timestamp of the window, the optimal time delay of temperature-concentration, the optimal time delay of air pressure-concentration, the correlation strength of temperature-concentration, the correlation strength of air pressure-concentration, and the spatial pattern label.

[0086] The feature management dataset is input into a convolutional neural network to calculate the contribution weight of each feature to the glue residue analysis. Then, based on the contribution weight, high-contribution features are convolved to obtain a directional feature map. Specifically, this involves converting the temperature-concentration time lag, air pressure-concentration time lag, temperature-concentration correlation strength, air pressure-concentration correlation strength, and spatial labels in the feature management dataset into normalized values ​​in the 0-1 range, arranging them into a two-dimensional data matrix in timestamp order plus monitoring point spatial coordinates, and using this matrix as input to the convolutional neural network. The Pearson correlation coefficient between each feature and the glue residue concentration is calculated through neural network branches, and the larger the absolute value of the correlation coefficient, the higher its contribution weight for that feature. The feature map is then analyzed based on its contribution weight. The features are sorted from highest to lowest weight, and the top 50% are selected to form a high-contribution feature subset. A 3×3 convolution kernel is used, sliding along the time dimension of the two-dimensional data matrix with a stride of 1. Window-by-window operation is performed on the high-contribution feature subset. During window-by-window operation, for each 3×3 window, which contains feature values ​​of 3 consecutive time points and 3 adjacent monitoring points, each feature value in the window is multiplied by the weight at the corresponding position of the convolution kernel. All product results are then added to obtain the local feature association strength within the spatiotemporal window. Finally, a two-dimensional directional feature map is output, with the horizontal axis representing time and the vertical axis representing the spatial coordinates of the monitoring points. The value of each spatial coordinate represents the feature association strength at that spatiotemporal location.

[0087] The directional feature map and contribution weights are input into the multilayer perceptron to allow high-contribution features to maintain a higher influence in the multilayer mapping and generate a fusion feature management vector. Specifically, the directional feature map is expanded point by point in the order of time priority and spatial following: first, each time node is traversed from early to late according to the time axis, and then the spatial monitoring points are sorted under each time node. The feature association strength values ​​of all points are concatenated in sequence to form a one-dimensional feature sequence containing time-space-feature association strength. Each element in the one-dimensional feature sequence corresponds to a high-contribution feature and association strength at a unique spatiotemporal location.

[0088] For each element in a one-dimensional feature sequence corresponding to a high-contribution feature, the contribution matching weight of that feature is calculated based on a convolutional neural network. The element value is multiplied by the corresponding contribution matching weight to obtain a spatiotemporal feature weighted value. All spatiotemporal feature weighted values ​​are arranged in the original sequence order to form a weighted one-dimensional vector containing the spatiotemporal feature weighted values. This weighted one-dimensional vector is input into the input layer of a multilayer perceptron. The dimension of this weighted one-dimensional vector corresponds one-to-one with the number of neurons in the input layer. Each neuron receives one spatiotemporal feature weighted value. The weighted one-dimensional vector is then passed to multiple hidden layers. Each neuron in each layer performs a linear weighted summation of the received spatiotemporal feature weighted values. The summation result is then processed by the ReLU activation function to retain positive values ​​and suppress negative values. Finally, the output layer integrates the effective data processed by all hidden layers to generate a fused feature management vector. The fused feature management vector contains data dimensions such as the high-contribution feature weighted values ​​of each key spatiotemporal region and the global feature association strength.

[0089] The fusion feature management vector is separated to generate a residual environment vector containing three dimensions: residual glue molecule concentration, thermodynamic equilibrium offset, and surface energy gradient. Specifically, this involves: acquiring multiple sets of sample data with known residual glue molecule concentrations; recording the correspondence between the weighted values ​​of high-contribution features and known residual glue molecule concentrations under different operating conditions; establishing a numerical mapping table between the two; comparing the numerical mapping table to find the matching item corresponding to the weighted value of the high-contribution feature in the fusion feature management vector; and obtaining the temperature and pressure correlation high-contribution features of the target management equipment under a state of no glue residue and stable operation. The baseline value of the contribution feature is the thermodynamic equilibrium standard value. Then, the weighted values ​​of the high contribution features related to temperature and air pressure in the fused feature management vector are subtracted from the corresponding baseline values. The two differences together constitute a thermodynamic equilibrium offset. Based on the coordinates of the spatial monitoring points corresponding to each high contribution feature in the fused feature management vector, adjacent monitoring points are found, and the difference between the weighted values ​​of the high contribution features corresponding to adjacent monitoring points is calculated. This difference is divided by the three-dimensional Euclidean distance between two adjacent monitoring points to obtain the surface energy gradient. The residual concentration of glue molecules, the thermodynamic equilibrium offset, and the surface energy gradient are integrated to obtain the residual environment vector.

[0090] By analyzing the temporal physicochemical data of monitoring points such as the bonding work area and curing chamber, and through a 60-second sliding time window and cross-correlation analysis, the optimal correlation lag and intensity of temperature, air pressure, and residual glue concentration are obtained. Combined with spatial clustering, the correlation pattern is obtained, forming a feature management dataset. Then, the feature contribution weights are calculated by a convolutional neural network, and high-contribution features are extracted to generate a directional feature map. Combined with a multilayer perceptron, the spatiotemporally weighted features are integrated to obtain a fused feature management vector. Finally, the residual environment vector composed of residual glue molecule concentration, thermodynamic equilibrium offset, and surface energy gradient is separated. The residual environment vector can reflect the residual and temperature field state in the equipment in real time, avoid glue interference between adjacent tasks, reduce alignment slip and local curing defects, ensure bonding quality, and improve scheduling effectiveness.

[0091] In one embodiment, the adhesive data is calculated based on the backpropagation algorithm to obtain the adhesive environment management value for the bonding task, including:

[0092] Obtain historical qualified process parameters of the target management equipment. These parameters include the temperature, air pressure, viscosity range, adhesive curing rate range, and adhesive thermal expansion coefficient range within the bonding work area of ​​the bonding equipment when the glass bonding quality is qualified.

[0093] The pre-processed adhesive data is compared with the historical qualified process parameters of the target management equipment to obtain the process scenario adaptability. Specifically, this includes comparing the parameters in the adhesive data with the corresponding parameter ranges in the historical qualified process parameters. If the parameter is within the qualified parameter range, the matching value of the parameter is recorded as 1. If it is not, it is recorded as a value between 0 and 1 according to the deviation ratio. For example, if the viscosity is within the viscosity range, the viscosity matching value is 1. If the viscosity is lower than the lower limit of the viscosity range, the current viscosity is divided by the lower limit of the viscosity range to obtain the viscosity matching value.

[0094] The viscosity weight is set to 0.4, the curing rate weight to 0.3, and the coefficient of thermal expansion weight to 0.3. The matching values ​​of each parameter are multiplied by their corresponding weights, summed, and normalized to between 0 and 1 to obtain the process scenario adaptability.

[0095] Viscosity directly determines the uniformity of adhesive application, its flowability, and its wetting effect on the glass surface during the bonding process. Excessive viscosity can lead to uneven coating, while insufficient viscosity can cause adhesive overflow. It is a core parameter affecting the sealing performance of the bond and has the greatest direct impact on the quality of the bond, hence it is assigned the highest weight of 0.4. Curing rate affects the strength and efficiency of the cured adhesive layer. Abnormal rates can lead to incomplete curing or excessive cross-linking, but the impact is mainly reflected in the curing process, so its weight is set at 0.3. The coefficient of thermal expansion mainly affects the compatibility of the adhesive with materials such as glass during temperature changes, preventing delamination due to expansion differences. However, this effect is relatively delayed in the short-term bonding stage, so its weight is set at 0.3.

[0096] The dynamic decay law of adhesive viscosity with the curing process and the inhibitory effect of the coefficient of thermal expansion on the curing rate were analyzed to generate dynamic interference values ​​of parameters. Specifically, the adhesive viscosity was recorded every 10 seconds during the curing process, for a total of 30 sets of data. The difference between each set of viscosity and the initial viscosity at the start of curing was calculated, and the difference was the decay amount. The mean of the decay amounts of 5 consecutive sets was then calculated to obtain the smoothed decay value. These smoothed decay values ​​were arranged in chronological order to form a trend curve of viscosity decreasing with curing time. The adhesive curing rate when the coefficient of thermal expansion was in the standard range in the history of the target management equipment was obtained as the reference rate. The current adhesive curing rate was subtracted from the reference rate and then divided by the reference rate to obtain the inhibition rate. The decay value was found from the trend curve at the current time point. The decay value and the inhibition rate were multiplied, and the result was normalized to between 0 and 1, which is the dynamic interference value of the parameters.

[0097] Based on the process scenario adaptability, the dynamic interference value of the parameter is scaled to generate a scenario-based feature enhancement matrix. Specifically, this includes: subtracting the process scenario adaptability from 1 to obtain the scaling factor; using the process scenario adaptability as the first row of data in the matrix, and using the result of multiplying the dynamic interference value of the parameter by the scaling factor as the second row of data, forming a two-dimensional matrix with 2 rows and 1 column. This matrix is ​​the scenario-based feature enhancement matrix.

[0098] The scenario-based feature enhancement matrix is ​​analyzed using the backpropagation algorithm. The error propagation path of the backpropagation is adjusted using dynamic interference values ​​as weights to generate initial values ​​for adhesive environment management. Specifically, the scenario-based feature enhancement matrix is ​​input into the backpropagation network. Through forward propagation, the median of the acceptable range for each adhesive parameter is calculated. The median of each parameter is weighted and summed according to the weights of viscosity (0.4), curing rate (0.3), and thermal expansion coefficient (0.3) to obtain a standard value. The process scenario adaptability of the first row of the scenario-based feature enhancement matrix is ​​used as the basic predicted value. The standard value is subtracted from the basic predicted value to obtain the initial error. Then, the dynamic interference values ​​associated with the second row of the matrix are used as weights. The larger the interference value, the stronger the impact of the error on the correction of parameters in each layer of the network, and vice versa. After multiple rounds of error calculation and parameter adjustment iterations, the predicted value at which the error is minimized and stabilized is the initial value for adhesive environment management.

[0099] The initial value of adhesive environment management is calibrated based on the dynamic interference value of the parameters and the adaptability of the process scenario to generate the adhesive environment management value. Specifically, this involves: subtracting the adaptability of the process scenario from 1 and multiplying the result by the dynamic interference value of the parameters to obtain an adjustment coefficient in the 0-1 range; if the initial value of adhesive environment management is higher than 0.5, the initial value of adhesive environment management is subtracted from the initial value of adhesive environment management and multiplied by the adjustment coefficient, and the result is normalized to the 0-1 range to obtain the adhesive environment management value; if the initial value of adhesive environment management is lower than 0.5, the initial value of adhesive environment management is added to (1 minus the initial value of adhesive environment management) and multiplied by the adjustment coefficient, and the result is normalized to the 0-1 range to obtain the adhesive environment management value.

[0100] By comparing adhesive data with historical qualified process parameters, the process scenario adaptability is calculated with weights of viscosity (0.4), curing rate (0.3), and coefficient of thermal expansion (0.3), reflecting the differentiated impact of parameters on bonding quality. Furthermore, the dynamic decay law of viscosity and the inhibitory effect of the coefficient of thermal expansion on the curing rate are analyzed to generate dynamic interference values ​​for the parameters. A scenario-based feature enhancement matrix is ​​constructed based on the adaptability, and the error propagation is adjusted using the interference values ​​through a backpropagation algorithm. After calibration, the adhesive environment management value is obtained. This value accurately reflects the adaptability and dynamic changes of the adhesive in the equipment, avoiding interference from the flow leveling characteristics when different adhesive tasks (high viscosity, low viscosity, fast curing, slow curing, etc.) are scheduled adjacently, thus ensuring bonding quality.

[0101] In one embodiment, based on the residual environment vector of the target management object and the glue environment management value of any bonding task to be scheduled, after the bonding task is executed by the target management device, a single task transition management value for the bonding task is generated based on the risk of the process results deviating from the quality standard through long short-term memory network analysis. This includes:

[0102] The residual environmental vector and the adhesive environmental management value are dynamically interacted to calculate the real-time adaptation loss degree of each dimension parameter and adhesive properties during the bonding process, generating the ring-adhesive interaction loss coefficient. Specifically, the three dimensions of the residual environmental vector—adhesive molecule residual concentration, thermodynamic equilibrium offset, and surface energy gradient—are normalized to the 0-1 interval, with the weights of adhesive molecule residual concentration (0.4), thermodynamic equilibrium offset (0.3), and surface energy gradient (0.3) set to 0.3. The normalized values ​​of adhesive molecule residual concentration, thermodynamic equilibrium offset, and surface energy gradient are multiplied by their respective weights and summed. The summation result is then normalized to the 0-1 interval to obtain the environmental impact value. The environmental management value is subtracted from 1, multiplied by the environmental impact value, and the result is normalized to the 0-1 interval, which is the ring-adhesive interaction loss coefficient.

[0103] When the ring-linked adhesive cross-loss coefficient exceeds the critical management value, the synergistic failure threshold caused by excessively high residual adhesive concentration leading to a decrease in the curing rate of new adhesive, and the process anomaly threshold caused by the conflict between thermodynamic equilibrium offset and adhesive thermal expansion coefficient are analyzed separately. The synergistic failure threshold and process anomaly threshold are integrated to generate a multi-dimensional risk trigger threshold. Specifically, this includes: obtaining 100 ring-linked adhesive cross-loss coefficients, sorting them from smallest to largest, and taking the 90th percentile after sorting as the critical management value; when the ring-linked adhesive cross-loss coefficient exceeds the critical management value; and using a linear regression algorithm to calculate the relationship between residual adhesive concentration and curing rate. The curing rate decreases for every 1 unit increase in residual adhesive concentration. When the curing rate drops to the lower limit of the acceptable curing rate range, the corresponding residual adhesive concentration is the synergistic failure threshold. The thermal expansion coefficient of the adhesive for the current task to be scheduled is extracted. Combined with the thermodynamic equilibrium offset in the residual environment vector, the correlation between these two parameters is calculated using Pearson correlation analysis. When the correlation is negative and the absolute value exceeds 0.8, indicating a strong negative correlation, the corresponding thermodynamic equilibrium offset is the process anomaly threshold. The synergistic failure threshold and the process anomaly threshold are compared numerically, and the smaller value is selected as the final multidimensional risk triggering threshold.

[0104] The cyclic adhesive cross-loss coefficient, multidimensional risk trigger threshold, residual environment vector, and adhesive environment management value are sorted according to the bonding process time sequence and integrated into a time-series risk feature sequence. This sequence is then imported into a long short-term memory network for analysis to generate a preliminary risk assessment value and risk transition probability value for a single task. Specifically, the bonding process of the LCD production bonding equipment is divided into three stages: heating, pressurization, and curing. The heating stage is associated with temperature-related parameters, which are filled into the thermodynamic equilibrium offset and cyclic adhesive cross-loss coefficient in the residual environment vector. The pressurization stage involves surface pressure, which is filled into the surface energy gradient and multidimensional risk trigger threshold in the residual environment vector. The curing stage focuses on the adhesive curing effect, which is filled into the residual adhesive molecule residual concentration and adhesive environment management value in the residual environment vector. The above six parameters are arranged in the order of heating, pressurization, and curing to form values ​​such as [thermodynamic equilibrium offset, cyclic adhesive cross-loss coefficient, surface energy gradient, multidimensional risk trigger threshold, adhesive molecule residual concentration, adhesive environment management value], which is the time-series risk feature sequence.

[0105] The process involves acquiring the risk feature sequence of historical qualified product production by the target management equipment, inputting this historical qualified risk feature sequence into a trained LSTM, and comparing the current time-series risk feature sequence with the historical qualified risk feature sequence parameters one-to-one. For each corresponding parameter, the current parameter value is subtracted from the historical qualified parameter value to obtain the difference. The difference is then squared to obtain the squared error of a single parameter. The squared errors of all six parameters are summed and divided by 6 to obtain the average value, which is the mean squared error. The mean squared errors calculated from all historical qualified risk feature sequences are statistically analyzed, and the maximum value is taken as the upper limit of error. 1 is subtracted from (the current mean squared error value divided by the upper limit of error) to obtain the preliminary risk assessment value for a single task. The preliminary risk assessment value for a single task ranges from 0 to 1. The smaller the current mean squared error, the closer the assessment value is to 1, indicating a lower risk for the single task. The softmax algorithm is used to map the time-series risk feature sequence to the probabilities of low risk, medium risk, and high risk levels. The probability of each level is between 0 and 1, and the sum is 1. The value corresponding to the highest probability level is the risk transition probability value.

[0106] The initial risk assessment value for a single task is calibrated based on the risk transition probability value and the multidimensional risk trigger threshold to generate a single task transition management value. Specifically, this involves comparing the parameters in the residual environment vector that are associated with the multidimensional risk trigger threshold (such as the synergistic failure threshold corresponding to the residual glue molecule concentration and the process anomaly threshold corresponding to the thermodynamic equilibrium offset) with the multidimensional risk trigger threshold. If the parameter exceeds the threshold, the threshold exceedance coefficient is recorded as 1; otherwise, it is recorded as 0. The risk transition probability value is multiplied by the threshold exceedance coefficient to obtain the calibration coefficient. The calibration coefficient ranges from 0 to 1, and the larger the calibration coefficient, the stronger the calibration force.

[0107] If the risk transition probability value corresponds to a high risk level, subtract (the initial risk assessment value of the single task multiplied by the calibration coefficient) from the initial risk assessment value of the single task, and normalize the result to the 0-1 range, which is the single task transition management value; if the risk transition probability value corresponds to a medium risk level, multiply the initial risk assessment value of the single task by (1 minus the calibration coefficient and then divide by 2), and normalize the result to the 0-1 range, which is the single task transition management value; if the risk transition probability value corresponds to a low risk level, add (1 - the initial risk assessment value of the single task) multiplied by the calibration coefficient to the initial risk assessment value of the single task, and normalize the result to the 0-1 range, which is the single task transition management value.

[0108] By calculating the ring-resin interaction loss coefficient, the adaptation loss between the residual environmental vector and the adhesive environment management value is quantified. When the ring-resin interaction loss coefficient exceeds the critical management value, the collaborative failure threshold and process abnormality threshold are extracted to generate a multi-dimensional risk trigger threshold, identify high-risk critical points, and then integrate the parameters according to the heating, pressurizing, and curing stages into a time-series risk feature sequence. This sequence is input into a long short-term memory network to compare with historical qualified data, and obtain the preliminary risk assessment value and risk transition probability value for a single task. Finally, the threshold exceedance coefficient is combined to calibrate and generate a single task transition management value, which accurately reflects the interaction risk between the bonding task and the residual environment of the bonding equipment.

[0109] In one embodiment, the single-task transition management values ​​of multiple matching tasks to be scheduled are imported into a long short-term memory network for analysis in scheduling order to generate a global risk accumulation management value, including:

[0110] Extract the single-task transition management value of each pending fitting task, analyze the risk transmission effect of preceding fitting tasks on multiple subsequent fitting tasks according to the scheduling order, and generate a cross-order risk management attenuation coefficient. Specifically, this includes: arranging the single-task transition management value of each pending fitting task into an ordered sequence according to the scheduling order of the fitting tasks; taking the completion time of the previous fitting task as the starting point, calculating the time interval from that time to the start time of each subsequent pending fitting task, setting the attenuation coefficient as 1 for time intervals of 5 minutes or less, 0.8 for time intervals of 6-10 minutes, 0.6 for time intervals of 11-15 minutes, and 0.4 for time intervals of more than 16 minutes; multiplying the single-task transition management value of the preceding fitting task by its attenuation coefficient to a subsequent fitting task to obtain the risk transmission value of the preceding task to the subsequent fitting task; adding the risk transmission values ​​of all preceding tasks to the same subsequent fitting task, and then dividing by the number of preceding tasks involved in the addition to obtain the cross-order risk management attenuation coefficient of the subsequent fitting task.

[0111] The process duration of each bonding task to be scheduled is obtained. Based on the process duration and the cross-sequence risk management attenuation coefficient, the single task transition management value is calculated to generate a time-varying management risk value. Specifically, this includes: obtaining the process duration and standard process duration of the entire bonding process of the target management equipment; dividing the process duration by the standard process duration to obtain the duration ratio; multiplying the single task transition management value by the cross-sequence risk management attenuation coefficient to obtain the basic risk value; and then multiplying the basic risk value by the duration ratio and normalizing the result to the 0-1 range, which is the time-varying management risk value.

[0112] The adaptation deviation between each time-varying management risk value and the residual environmental vector of the current target management device is calculated to generate the risk-environment dynamic adaptation deviation rate. Specifically, this includes: normalizing the residual glue molecule concentration, thermodynamic equilibrium offset, and surface energy gradient of the residual environmental vector to the 0-1 interval; calculating the absolute difference between the time-varying management risk value and the values ​​of the three dimensions of the residual environmental vector; and multiplying the three absolute differences by their corresponding weights according to the weights of residual glue molecule concentration (0.4), thermodynamic equilibrium offset (0.3), and surface energy gradient (0.3), and then summing the results. The result is then normalized to the 0-1 interval, which is the risk-environment dynamic adaptation deviation rate.

[0113] By extracting the single-task transition management value of the bonding task to be scheduled, calculating the time interval between the preceding task and the subsequent task according to the scheduling order, and generating the cross-sequence risk management attenuation coefficient based on the interval setting, we can understand the risk transmission effect of the preceding bonding task on the subsequent bonding task. Then, by combining the ratio of task process time to standard process time, we can correct the single-task transition management value to obtain the time-varying management risk value. Finally, we can calculate the adaptation deviation rate between the time-varying management risk value and the residual environmental vector, match the environmental and task risks of the bonding equipment, and improve the scheduling management accuracy of the bonding equipment for LCD screens.

[0114] In one embodiment, the single-task transition management values ​​of multiple matching tasks to be scheduled are imported into a long short-term memory network for analysis in scheduling order to generate a global risk accumulation management value, and the method further includes:

[0115] The time-varying management risk value, cross-order risk management attenuation coefficient, and risk-environment dynamic adaptation deviation rate are integrated according to the scheduling order. At the same time, the process type of the bonding task is distinguished, and a dynamic correlation risk feature sequence containing four types of dynamic information, namely risk, transmission, environment, and process, is constructed. Specifically, the process type of the bonding task is obtained. If it is high-hardness glass bonding, the process is assigned a value of 1, and if it is low-hardness glass bonding, the process is assigned a value of 0. For each bonding task, the time-varying management risk value of the risk category, the cross-order risk management attenuation coefficient of the transmission category, and the risk-environment dynamic adaptation deviation rate of the environment category are extracted respectively. These are added to the process type to form a task quadruple containing four data points. According to the scheduling order, all task quadruples are connected in sequence to form a dynamic correlation risk feature sequence containing four types of information, namely risk, transmission, environment, and process.

[0116] The dynamic correlation risk feature sequence is imported into a Long Short-Term Memory (LSTM) network for analysis to generate risk management factors and preliminary cumulative global risk values. Specifically, this involves: normalizing the four data points of each task quadruple in the dynamic correlation risk feature sequence to the 0-1 interval; inputting the preprocessed dynamic correlation risk feature sequence into a trained LSTM; the LSTM captures the temporal risk correlation between tasks through memory units and outputs the risk contribution features of each task; mapping the risk contribution features to values ​​in the 0-1 interval using the softmax algorithm, which represents the risk management factor, indicating the weight of the corresponding task's impact on the global risk; multiplying the time-varying risk value of each task by its risk management factor; and summing the results of all tasks to obtain the preliminary cumulative global risk value.

[0117] The initial cumulative value of global risk is calibrated based on risk management factors and cross-order risk management attenuation coefficients to generate a two-dimensional cumulative risk value. Specifically, this involves multiplying the risk management factor of each task with the corresponding cross-order risk management attenuation coefficient to obtain the two-dimensional calibration weight of that task. The two-dimensional calibration weight ranges from 0 to 1, with a larger weight indicating a stronger combined effect of risk transmission and its own influence on that task. The two-dimensional calibration weights of all tasks are summed to obtain the total calibration coefficient. The initial cumulative value of global risk is multiplied by the total calibration coefficient, and the result is normalized to the 0-1 range to obtain the two-dimensional cumulative risk value.

[0118] The risk contribution of each aligned task in the two-dimensional risk cumulative value is analyzed, and a risk cascading effect coefficient is generated. The two-dimensional risk cumulative value is then adjusted based on this coefficient to generate a global risk cumulative management value. Specifically, this involves: calculating the proportion of the two-dimensional calibration weight of each aligned task to the total calibration coefficient of all tasks; this proportion represents the risk contribution percentage of that aligned task in the two-dimensional risk cumulative value; for each aligned task, multiplying the risk contribution percentages of all its preceding aligned tasks by the cross-order risk management attenuation coefficient from the preceding task to the current aligned task and summing the results to obtain the cascading effect coefficient for that task; summing the cascading effect coefficients of all aligned tasks to obtain the total risk cascading effect coefficient; and multiplying the two-dimensional risk cumulative value by (1 + total cascading effect coefficient) and normalizing the result to the 0-1 range to obtain the global risk cumulative management value.

[0119] By integrating time-varying risk management values, cross-sequence risk management attenuation coefficients, and risk-environment dynamic adaptation deviation rates according to the scheduling order, and combining them with the process type of LCD bonding tasks, a dynamic correlation risk feature sequence containing risk, transmission, environment, and process is constructed. This sequence is then imported into a long short-term memory network to capture the temporal risk correlation of tasks to generate risk management factors and preliminary global risk accumulation values. After two-dimensional calibration and weight correction, a two-dimensional risk accumulation value is obtained. Finally, the global risk accumulation management value is generated by adjusting the risk chain effect coefficient. This accurately identifies the superimposed interference risks of different adhesive tasks, such as high viscosity slow curing and low viscosity fast curing, and reduces the occurrence of alignment slippage and local curing defects during LCD bonding.

[0120] In one embodiment, minimizing the cumulative global risk management value is taken as the core optimization objective. A neural network algorithm based on a combination of convolutional neural networks and multilayer perceptrons is used to analyze the utilization rate and production cycle of the target management equipment, resulting in a scheduling management index, including:

[0121] The process involves obtaining the time-series utilization rate of the target management equipment, the priority weight of the tasks to be scheduled, and the production cycle. Specifically, this includes: setting the statistical period to the most recent hour, dividing it into 6 time intervals of 10 minutes, recording the actual time the target management equipment spends executing the bonding task within each interval, dividing the actual working time of each interval by 10 minutes to obtain the utilization rate of each interval, and the 6 utilization rates representing the utilization rate fluctuation value every 10 minutes within the most recent hour; classifying the bonding tasks to be scheduled into emergency bonding tasks and regular bonding tasks, assigning a priority weight of 1.2 to emergency bonding tasks and a priority weight of 1 to regular bonding tasks, and obtaining the production cycle of each bonding task, which is the period from the start to the end of bonding.

[0122] With minimizing the global risk accumulation management value as the core constraint, this study analyzes the dynamic correlation between time series utilization, priority weight, and risk accumulation to obtain the risk-efficiency-priority dynamic management coefficient. Specifically, this includes: normalizing the time series utilization and priority weight to the 0-1 range; subtracting the minimum value from the maximum value of the six utilization rates to obtain the maximum change range, subtracting (the global risk accumulation management value divided by the maximum change range) from 1, and normalizing the result to the 0-1 range to obtain the correlation between the time series utilization and the global risk accumulation management value; subtracting the global risk accumulation management value from 1 to obtain the matching value, subtracting (the absolute value of the matching value minus the priority weight) from 1, and normalizing the result to the 0-1 range to obtain the correlation between the priority weight and the risk accumulation value.

[0123] With risk accumulation as the core constraint, the correlation weight between the global risk accumulation management value and the global risk accumulation management value is set to 0.5, the correlation weight between the time series utilization rate and the global risk accumulation management value is set to 0.3, and the correlation weight between the priority weight and the risk accumulation value is set to 0.2. The global risk accumulation management value and the two correlation weights are multiplied by their respective weights and then added together. The calculation result is normalized to the 0-1 range, which is the risk-efficiency-priority dynamic management coefficient.

[0124] The risk-efficiency-priority dynamic management coefficient, time series utilization rate, and production cycle are integrated in layers to generate a layered time series feature matrix. The layered time series feature matrix is ​​then imported into a convolutional neural network for calculation to generate a core feature vector for time series management. Specifically, this includes: normalizing the production cycle to the 0-1 interval; arranging the risk-efficiency-priority dynamic management coefficient, time series utilization rate, and production cycle into a 3x6 two-dimensional matrix, with each layer corresponding to a type of parameter and each column corresponding to a time interval. The first layer is the coefficient layer, used to control the overall system: six time intervals share a single value, so the first row is filled with six identical risk-efficiency-priority dynamic management coefficients; the second layer is the utilization rate layer, used to reflect the status of the target management equipment: six time series utilization rates are filled into the second row in chronological order; and the third layer is the cycle layer, used to associate task requirements: according to the scheduling order, the production cycle of the corresponding task to be executed for each time interval is filled into the third row.

[0125] The two-dimensional matrix is ​​input into the convolutional neural network. A 3×3 convolutional kernel is used and slides along the time interval direction (column direction) with a stride of 1. The sum of the product of all values ​​in the window and the preset weights of the convolutional kernel is calculated to obtain the local feature value of the window. After sliding through all windows, there are a total of 4 windows, and 4 local feature values ​​are obtained. Then, the ReLU activation function is used to filter out the effective features. The filtered effective features are concatenated in order to form a one-dimensional time series management core feature vector.

[0126] The core feature vector of time-series management is input into a multilayer perceptron and analyzed in conjunction with the risk-efficiency-priority dynamic management coefficient to generate a dynamic calibration efficiency evaluation value. Specifically, this involves: integrating the one-dimensional core feature vector of time-series management with the risk-efficiency-priority dynamic management coefficient, adding the risk-efficiency-priority dynamic management coefficient as an extra element to the end of the vector to form a new input vector, thereby ensuring that the dimension of this vector completely corresponds to the number of neurons in the MLP input layer; passing each element in the input vector to the corresponding neuron, with each neuron multiplying the received element value by its preset weight, summing all the products, and using the ReLU activation function to retain positive values ​​and filter out negative values ​​to obtain the hidden layer output value; multiplying all the hidden layer output values ​​again by the preset weights of the output layer neurons and summing them to obtain the preliminary efficiency evaluation value; multiplying the preliminary efficiency evaluation value by the risk-efficiency-priority dynamic management coefficient, and normalizing the calculation result to the 0-1 interval to obtain the dynamic calibration efficiency evaluation value.

[0127] With the goal of minimizing the global cumulative risk management value, the dynamic calibration performance evaluation value is anchored and calibrated to obtain the scheduling management index. Specifically, this involves: acquiring all records corresponding to the dynamic calibration performance evaluation value in the target management equipment; filtering out all records with a dynamic calibration performance evaluation value ≥ 0.8 (representing high equipment performance); extracting the global cumulative risk management value corresponding to these records; summing these global cumulative risk management values ​​and dividing by the number of records to obtain the target risk value; subtracting the target risk value from the current global cumulative risk management value to obtain the risk difference; dividing the risk difference by the target risk value to obtain the risk deviation coefficient. If the risk difference is negative, the risk deviation coefficient is directly set to 0, indicating that the risk has met the standard; calculating the average of the three most recent dynamic calibration performance evaluation values ​​and using this average as the benchmark value for risk minimization; calculating the performance difference between the current dynamic calibration performance evaluation value and the benchmark value; multiplying the performance difference by the risk deviation coefficient to obtain the calibration correction amount; subtracting the calibration correction amount from the benchmark value; and normalizing the calculation result to the 0-1 range, which is the scheduling management index.

[0128] By minimizing the cumulative global risk management value, the system first obtains the time-series utilization rate of the target management equipment every 10 minutes over the past hour, the priority weight of the LCD screen bonding task (and the production cycle), calculates the risk-efficiency-priority dynamic management coefficient, and then constructs a hierarchical time-series feature matrix containing this coefficient, time-series utilization rate, and production cycle. The core feature vector of time-series management is extracted through a convolutional neural network, and a dynamic calibration efficiency evaluation value is generated by combining it with a multilayer perceptron. Finally, the target risk value corresponding to the efficiency of high bonding equipment is anchored, and the scheduling management index is calibrated to obtain the index. This process balances risk with equipment utilization rate and production cycle, and avoids interference from adhesive tasks such as high viscosity slow curing and low viscosity fast curing.

[0129] In one embodiment, a neural network algorithm combining a long short-term memory network and a convolutional neural network is used to analyze the glue environment management value and the global risk accumulation management value of all bonding tasks to be scheduled, generating a dynamic management threshold, including:

[0130] Obtain the process sensitivity attribute of the bonding task, specifically including: if the bonding task is high-hardness glass bonding, the process sensitivity attribute is assigned a value of 1.3, and if it is low-hardness glass bonding, the process sensitivity attribute is assigned a value of 0.8.

[0131] Based on the adhesive environment management value, global risk accumulation management value, and process sensitivity attribute of all bonding tasks to be scheduled, the mutual influence intensity of adhesive environment fluctuation and risk accumulation under different process sensitivity attributes is calculated to generate the adhesive sensitivity management coefficient. Specifically, this includes: dividing all bonding tasks to be scheduled into two groups according to the process sensitivity attribute, one group being high-hardness glass bonding tasks and the other being low-hardness glass bonding tasks, ensuring that the process sensitivity attribute of tasks within each group is consistent; for all bonding tasks within each group, the maximum and minimum values ​​of the adhesive environment management value are found, and the maximum value is subtracted from the minimum value to obtain the adhesive environment fluctuation range of that group; for each group, the mean of the global risk accumulation management value of all bonding tasks in that group is calculated, and the adhesive environment fluctuation range of that group is divided by the mean value to obtain the fluctuation-risk impact degree; the fluctuation-risk impact degree of each group is multiplied by the corresponding process sensitivity attribute value of that group, and the result is divided by 1.3 (here, 1.3 represents the process sensitivity attribute value of high-hardness glass bonding) and normalized to the 0-1 interval to obtain the adhesive sensitivity management coefficient of each group.

[0132] The glue sensitivity management coefficient, glue environment management value, and global risk cumulative management value are integrated into a risk-environment-sensitivity time series according to the scheduling order, and imported into a long short-term memory network to calculate the actual contribution of each dimension to the global risk and generate time series risk contribution weights. Specifically, according to the scheduling order of the fitting tasks, the glue sensitivity management coefficient, glue environment management value, and global risk cumulative management value of each fitting task are combined into a risk-environment-sensitivity triplet for that task. All the triplets of all tasks are connected in sequence according to the scheduling order to form a two-dimensional risk-environment-sensitivity time series. The three parameters of all triplets in the risk-environment-sensitivity time series are normalized to the 0-1 interval.

[0133] The preprocessed risk-environment-sensitivity time series is input into the trained LSTM. The LSTM captures the temporal correlation between the tasks through memory units and outputs the original contribution features of the three dimensions corresponding to each task. For the original contribution features of each task, the softmax algorithm is used to map them into three values ​​in the range of 0-1. The sum of the three values ​​is 1. These three values ​​are the temporal risk contribution weights of the three dimensions corresponding to the task. The larger the weight, the stronger the actual contribution of that dimension to the global risk.

[0134] The temporal risk contribution weight, glue sensitivity management coefficient, and risk-environment-sensitivity temporal sequence are input into a convolutional neural network for calculation, generating a key feature-sensitivity matching map. This map is then analyzed to generate a dynamic management threshold. Specifically, this involves merging the 3D temporal risk contribution weight, 1D glue sensitivity management coefficient, and 3D risk-environment-sensitivity triplet for each fitting task into an 8D feature vector. The 8D feature vectors of all fitting tasks to be scheduled are arranged in their scheduling order to construct a two-dimensional input matrix, where each row corresponds to a task and each column corresponds to a feature dimension. This two-dimensional input matrix is ​​then input... The convolutional neural network is fed into a one-dimensional convolutional neural network. This network uses a convolutional kernel with a width of 8 and a height of 3, and performs one-dimensional convolution strictly along the scheduling order (row direction). This allows it to specifically capture the local evolution patterns of global risk performance among multiple consecutive tasks. After convolution, the network is processed by an activation function to output a sequence risk feature vector that can characterize the risk level of different task segments. The sequence risk feature vector is then processed using a quantile algorithm, and its 90th quantile is taken. This value represents a critical level that can distinguish between most normal tasks and high-risk tasks based on historical high-risk patterns. This quantile is then normalized to the 0-1 interval to obtain the dynamic management threshold.

[0135] By grouping tasks based on their process sensitivity attributes for LCD bonding, the ratio of adhesive environment fluctuation amplitude to the average global risk cumulative management value for each group is calculated. This, combined with the process sensitivity attributes, generates an adhesive sensitivity management coefficient, revealing the strength of the adhesive fluctuation-risk correlation for different tasks. The adhesive sensitivity management coefficient, adhesive environment management value, and global risk cumulative management value are then integrated into a time-series sequence according to the scheduling order. A long short-term memory network is used to calculate the risk contribution weights for each dimension, and a convolutional neural network is employed to capture the continuous task risk evolution pattern. After normalizing the 90th percentile of the sequence risk feature vector, a dynamic management threshold is obtained. This dynamic management threshold accurately distinguishes between normal and high-risk scheduling, avoiding interference from adjacent adhesive tasks such as high-viscosity slow-curing and low-viscosity fast-curing adhesives, reducing LCD alignment slippage and localized curing defects, and ensuring bonding quality.

[0136] In one embodiment, the scheduling management index is compared with the dynamic management threshold to obtain a management instruction, including:

[0137] If the scheduling management index is greater than or equal to the dynamic management threshold, the scheduling management is deemed qualified, and a continued fitting instruction is issued.

[0138] If the scheduling management index is less than the dynamic management threshold, the scheduling management is deemed unqualified, and an alarm command is issued.

[0139] By comparing the scheduling management index with the dynamic management threshold to generate management instructions, it is possible to accurately determine whether the scheduling of the LCD screen bonding equipment is qualified, avoid interference from adhesive tasks such as high viscosity slow curing and low viscosity fast curing, reduce LCD screen alignment slippage, ensure bonding quality, and improve the scheduling and management effect of LCD screen bonding equipment.

[0140] In one embodiment, an intelligent scheduling and management system for liquid crystal display manufacturing lamination equipment is applied to the aforementioned intelligent scheduling and management method for liquid crystal display manufacturing lamination equipment, comprising:

[0141] The integration unit is used to acquire the bonding task of the target management device. After any bonding task is completed, the physical and chemical data of the target management device are collected in real time. The physical and chemical data are analyzed based on the convolutional neural network and multilayer perceptron to generate a residual environmental vector. The target management device is the bonding device of the LCD screen.

[0142] The environmental management unit is used to acquire the glue data of each bonding task to be scheduled in real time, and calculate the glue environmental management value of the bonding task based on the backpropagation algorithm.

[0143] The deviation management unit is used to generate a single task transition management value for the bonding task after it is executed by the target management equipment, based on the residual environment vector of the target management object and the glue environment management value of any bonding task to be scheduled, and the risk of the process results deviating from the quality standard based on the analysis of the long short-term memory network.

[0144] The risk management unit is used to import the single-task transition management values ​​of multiple matching tasks to be scheduled into the long short-term memory network for analysis in the order of scheduling, and generate a global risk accumulation management value.

[0145] The optimization management unit is used to analyze the utilization rate and production cycle of the target management equipment based on a neural network algorithm that combines convolutional neural networks and multilayer perceptrons, with the core optimization objective of minimizing the cumulative global risk management value, and to derive the scheduling management index.

[0146] The threshold calculation unit is used to analyze the glue environment management value and global risk accumulation management value of all bonding tasks to be scheduled based on a neural network algorithm that combines long short-term memory network and convolutional neural network, and generate dynamic management thresholds.

[0147] The scheduling management unit is used to compare the scheduling management index with the dynamic management threshold to obtain management instructions.

[0148] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent scheduling management method for a liquid crystal display screen production laminating device, characterized in that, The method comprises the following steps: Step S1, obtain the target management device's fitting task, and real-time collect the target management device's physical and chemical data after any one fitting task is completed, analyze the physical and chemical data based on a convolutional neural network and a multi-layer perception machine, and generate a residual environment vector, the target management device is a liquid crystal display screen fitting device, and the physical and chemical data comprises the temperature and air pressure in a fitting operation area inside the liquid crystal display screen fitting device and the concentration of residual glue molecules; Step S2, for each fitting task to be scheduled, real-time obtain the glue data of the fitting task, calculate the glue data based on a back propagation algorithm, and obtain the glue environment management value of the fitting task, the glue data comprises viscosity, solidification rate and thermal expansion coefficient; Step S3, according to the residual environment vector of the target management device and the glue environment management value of any one fitting task to be scheduled, simulate the fitting task after being executed on the target management device, analyze the risk of process result deviation from the quality standard based on a long short-term memory network, and generate a single task transition management value of the fitting task; Step S4, import the single task transition management values of a plurality of fitting tasks to be scheduled into the long short-term memory network in a scheduling order for analysis, and generate a global risk accumulation management value; Step S5, take minimizing the global risk accumulation management value as a core optimization target, analyze the utilization rate and production cycle of the target management device based on a neural network algorithm of the convolutional neural network and the multi-layer perception machine, and obtain a scheduling management index; Step S6, analyze the glue environment management value and the global risk accumulation management value of all fitting tasks to be scheduled based on a neural network algorithm of the long short-term memory network and the convolutional neural network, and generate a dynamic management threshold value; Step S7, compare the scheduling management index with the dynamic management threshold value, and obtain a management instruction.

2. The intelligent scheduling management method of the liquid crystal display screen production laminating equipment according to claim 1, characterized in that, The residual environment vector is generated by analyzing the physical and chemical data based on the convolutional neural network and the multi-layer perception machine, and the method comprises the following steps: Performing spatiotemporal coupling analysis on the preprocessed physical and chemical data, calculating the dynamic correlation degree of temperature, air pressure and residual glue molecule concentration at different times, and generating a feature management data set; Inputting the feature management data set into the convolutional neural network, calculating the contribution weight of each feature to glue residue analysis, and performing convolution operation on high-contribution features based on the contribution weight to obtain a directional feature spectrum; Inputting the directional feature spectrum and the contribution weight into the multi-layer perception machine together, so that the high-contribution features maintain higher influence in multi-layer mapping, and a fusion feature management vector is generated; Separating the fusion feature management vector to obtain a residual feature subvector, a thermodynamic feature subvector and a surface energy feature subvector, and generating a residual environment vector comprising three dimensions of glue molecule residual concentration, thermodynamic equilibrium offset and surface energy gradient based on the internal law mapping of the three subvectors.

3. The intelligent scheduling management method of the liquid crystal display screen production laminating equipment according to claim 1, characterized in that, The glue environment management value of the fitting task is obtained by calculating the glue data based on the back propagation algorithm, and the method comprises the following steps: Obtain the historical qualified process parameters of the target management device; Compare the preprocessed glue data with the historical qualified process parameters of the target management device to obtain a process scene adaptation degree; The dynamic attenuation law of viscosity of the glue with the curing process, the inhibition effect of the thermal expansion coefficient on the curing rate, and the generated parameter dynamic interference value are analyzed; Based on the process scene adaptation degree, the parameter dynamic interference value is scaled to generate a scene-based feature enhancement matrix; According to the back propagation algorithm, the scene-based feature enhancement matrix is analyzed, and the error transmission path of the back propagation is adjusted based on the parameter dynamic interference value to generate the initial value of the glue environment management; Based on the parameter dynamic interference value and the process scene adaptation degree, the initial value of the glue environment management is calibrated to generate the glue environment management value.

4. The intelligent scheduling management method of the liquid crystal display screen production laminating equipment according to claim 3, characterized in that, According to the residual environment vector of the target management device and the glue environment management value of any one to-be-scheduled bonding task, the risk of process result deviation from the quality standard after the execution of the bonding task on the target management device is simulated based on the long short-term memory network to generate a single-task transition management value of the bonding task, including: The residual environment vector and the glue environment management value are dynamically interacted to calculate the real-time adaptation loss degree of each dimension parameter and the glue characteristic in the bonding process to generate a ring-glue interaction loss coefficient; When the ring-glue interaction loss coefficient exceeds the critical management value, the synergistic failure threshold caused by the high residual glue concentration leading to the decrease of the curing rate of the new glue, the thermodynamic equilibrium offset, and the process abnormal threshold caused by the conflict of the glue thermal expansion coefficient are analyzed, and the synergistic failure threshold and the process abnormal threshold are integrated to generate a multi-dimensional risk trigger threshold; The ring-glue interaction loss coefficient, the multi-dimensional risk trigger threshold, the residual environment vector, and the glue environment management value are sorted according to the bonding process time sequence, integrated into a time sequence risk feature sequence, and imported into the long short-term memory network for analysis to generate a single-task risk preliminary evaluation value and a risk transition probability value; Based on the risk transition probability value and the multi-dimensional risk trigger threshold, the single-task risk preliminary evaluation value is calibrated to generate a single-task transition management value.

5. The intelligent scheduling management method of the liquid crystal display screen production laminating equipment according to claim 4, characterized in that, The single-task transition management values of a plurality of to-be-scheduled bonding tasks are imported into the long short-term memory network in the scheduling order for analysis to generate a global risk accumulation management value, including: The single-task transition management values of each to-be-scheduled bonding task are extracted, and the risk transmission effect of the previous bonding task on the subsequent bonding tasks is analyzed in the scheduling order to generate a cross-sequence risk management attenuation coefficient; The process time of each to-be-scheduled bonding task is obtained, and based on the process time and the cross-sequence risk management attenuation coefficient, the single-task transition management value is calculated to generate a time-varying management risk value; The adaptation deviation rate of each time-varying management risk value and the residual environment vector of the current target management device is calculated to generate a risk-environment dynamic adaptation deviation rate.

6. The intelligent scheduling management method of the liquid crystal display screen production laminating equipment according to claim 5, characterized in that, The single-task transition management values of a plurality of to-be-scheduled bonding tasks are imported into the long short-term memory network in the scheduling order for analysis to generate a global risk accumulation management value, including: The time-varying management risk value, the cross-sequence risk management attenuation coefficient, and the risk-environment dynamic adaptation deviation rate are integrated in the scheduling order, and the process types of the bonding tasks are distinguished to construct a dynamic associated risk feature sequence containing four types of dynamic information, i.e., risk, transmission, environment, and process. The dynamic correlation risk feature sequence is introduced into the long short-term memory network for analysis to generate a risk management factor and a preliminary cumulative value of global risk; The preliminary cumulative value of global risk is calibrated based on the risk management factor and the cross-sequence risk management attenuation coefficient to generate a two-dimensional risk cumulative value; The risk contribution of each task in the two-dimensional risk cumulative value is analyzed to generate a risk chain effect coefficient, and the two-dimensional risk cumulative value is adjusted based on the risk chain effect coefficient to generate a global risk cumulative management value.

7. The intelligent scheduling management method of the liquid crystal display screen production laminating device according to claim 6, characterized in that, The utilization rate and production cycle of the target management device are analyzed based on the neural network algorithm of convolutional neural network and multilayer perceptron cooperation to obtain a scheduling management index, including: The time sequence utilization rate of the target management device, the priority weight of the task to be scheduled, and the production cycle are obtained; The dynamic correlation between the time sequence utilization rate, the priority weight, and the risk accumulation is analyzed to obtain a risk-performance-priority dynamic management coefficient, with the minimization of the global risk cumulative management value as the core constraint; The risk-performance-priority dynamic management coefficient, the time sequence utilization rate, and the production cycle are integrated in layers to generate a layered time sequence feature matrix, which is introduced into the convolutional neural network for calculation to generate a time sequence management core feature vector; The time sequence management core feature vector is input into the multilayer perceptron, and the risk-performance-priority dynamic management coefficient is analyzed to generate a dynamic calibration performance evaluation value; The dynamic calibration performance evaluation value is anchored and calibrated to obtain the scheduling management index, with the minimization of the global risk cumulative management value as the target.

8. The intelligent scheduling management method of the liquid crystal display screen production laminating equipment according to claim 6, characterized in that, The glue environment management value and the global risk cumulative management value of all tasks to be scheduled are analyzed based on the neural network algorithm of long short-term memory network and convolutional neural network cooperation to generate a dynamic management threshold, including: The process sensitivity attribute of the task is obtained; The mutual influence strength of the glue environment fluctuation and the risk accumulation under different process sensitivity attributes is calculated based on the glue environment management value and the global risk cumulative management value of all tasks to be scheduled, and a glue sensitivity management coefficient is generated; The glue sensitivity management coefficient, the glue environment management value, and the global risk cumulative management value are integrated into a risk-environment-sensitivity time sequence in the scheduling order, and introduced into the long short-term memory network to calculate the actual contribution of each dimension to the global risk, generating a time sequence risk contribution weight; The time sequence risk contribution weight, the glue sensitivity management coefficient, and the risk-environment-sensitivity time sequence are input into the convolutional neural network for calculation to generate a key feature-sensitivity matching map, and the dynamic management threshold is generated by analyzing the key feature-sensitivity matching map.

9. The intelligent scheduling management method of the liquid crystal display screen production laminating device according to claim 8, characterized in that, The scheduling management index and the dynamic management threshold are compared to obtain a management instruction, including: If the scheduling management index is greater than or equal to the dynamic management threshold, it is determined that the scheduling management is qualified, and a continue bonding instruction is issued; If the scheduling management index is less than the dynamic management threshold, it is determined that the scheduling management is unqualified, and an alarm instruction is issued.

10. An intelligent scheduling management system of a liquid crystal display screen production laminating device, applied to the intelligent scheduling management method of the liquid crystal display screen production laminating device according to any one of claims 1-9, characterized in that, ​ The integration unit is used to obtain a fitting task of a target management device, to collect physical and chemical data of the target management device in real time after completion of any fitting task, to analyze the physical and chemical data based on a convolutional neural network and a multilayer perception machine, and to generate a residual environment vector, the target management device being a liquid crystal display screen fitting device, and the physical and chemical data including temperature and air pressure in a fitting operation area inside the liquid crystal display screen fitting device and concentration of residual glue molecules; The environment management unit is used to obtain glue data of each fitting task to be dispatched in real time, to calculate the glue data based on a back propagation algorithm, and to obtain a glue environment management value of the fitting task, the glue data including viscosity, solidification rate, and thermal expansion coefficient; The deviation management unit is used to simulate execution of any fitting task to be dispatched on the target management device based on a residual environment vector of the target management device and the glue environment management value of the fitting task, to analyze a risk of deviation of a process result from a quality standard based on a long short-term memory network, and to generate a single-task transition management value of the fitting task; The risk management unit is used to import the single-task transition management values of the multiple fitting tasks to be dispatched into the long short-term memory network in a scheduling order for analysis, and to generate a global risk accumulation management value; The optimization management unit is used to take minimization of the global risk accumulation management value as a core optimization target, to analyze utilization of the target management device and a production cycle based on a neural network algorithm of the convolutional neural network and the multilayer perception machine, and to obtain a scheduling management index; The threshold calculation unit is used to analyze the glue environment management values of all fitting tasks to be dispatched and the global risk accumulation management value based on a neural network algorithm of the long short-term memory network and the convolutional neural network, and to generate a dynamic management threshold value; The scheduling management unit is used to compare the scheduling management index with the dynamic management threshold value, and to obtain a management instruction.

Images