Optimized control method and system for production process of light shielding adhesive tape

Abstract

The application provides a production process optimization control method and system of light-shielding adhesive tape, and belongs to the technical field of industrial control, comprising the following steps: S100, acquiring a production process correlation data set, extracting key information from the production process correlation data set, and establishing a production process correlation graph; S200, acquiring a real-time production process control set, constructing a production process procedure intelligent agent cluster based on the real-time production process control set and the production process correlation graph, and generating a production process optimization control strategy through a virtual optimization strategy of the intelligent agent cluster; S300, performing digital twin verification on the production process optimization control strategy, generating a production process final control strategy, and performing real-time regulation and control based on the production process final control strategy; and S400, acquiring associated data after regulation and control, and updating the production process correlation graph based on the associated data after regulation and control. The method provides a control method for global and multidirectional optimization of the production process of light-shielding adhesive tape.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial control, in particular to a production process optimization control method and system of light shielding adhesive tape. BACKGROUND

[0002] In modern industrial production and daily life, light shielding adhesive tape, as a composite material with functions of light shielding, bonding, insulation, etc., is increasingly widely used. In the field of electronic equipment, it is used for frame sealing of liquid crystal display and light shielding protection of backlight module to ensure that the screen display effect is not disturbed by stray light; in automobile manufacturing, it is often used for light shielding fixation of lamp assemblies to ensure the accuracy and safety of driving lighting; in the building decoration industry, it can be used for local light shielding with glass curtain walls to adjust indoor lighting effect. With the increasing requirements of downstream industries for product precision and stability, the quality standards of light shielding adhesive tape are becoming more and more stringent, but there are still limitations in the production process control of light shielding adhesive tape.

[0003] On the one hand, single process parameter optimization, such as PID control of coating pressure, often ignores the coupling effect of coating, curing and slitting, leading to local optimization causing global deterioration, for example, high coating pressure to improve thickness uniformity will increase oven energy consumption.

[0004] On the other hand, conventional digital twin verification relies on historical working condition data, and the coverage rate of extreme disturbance is low, which often cannot predict complex risks, so that the generated optimization control instructions lack risk guarantee.

[0005] Therefore, it is urgent to build a new generation of production process global optimization framework integrating swarm intelligence game and adversarial twin verification to realize the self-sensing and self-decision global optimization control method of light shielding adhesive tape production process. SUMMARY

[0006] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a production process optimization control method of light shielding adhesive tape, which comprises:

[0007] S100, acquiring a production process associated data set, extracting key information from the production process associated data set, and establishing a production process associated graph;

[0008] S200, acquiring a real-time production process control set, constructing a production process procedure intelligent agent cluster based on the real-time production process control set and the production process associated graph, and generating a production process optimization control strategy through virtual optimization strategy of the intelligent agent cluster;

[0009] S300, performing digital twin verification on the production process optimization control strategy, generating a production process final control strategy, and performing real-time regulation and control based on the production process final control strategy.

[0010] S400, obtaining the regulated correlation data, and updating the production process correlation graph based on the regulated correlation data.

[0011] As a further scheme of the present application, a production process correlation data set is obtained, key information is extracted from the production process correlation data set, and a production process correlation graph is established, including:

[0012] The production process correlation data set is obtained based on a production monitoring device group, and data processing operations are performed on the production process correlation data set;

[0013] The production process key indicators in the production process correlation data set are extracted through a neural network, and a production process key matrix is established based on the production process key indicators;

[0014] The production linear weight, production nonlinear weight, and production time lag weight are calculated based on the production process key matrix, and the production process key matrix is imported into a PC algorithm, and a basic production process correlation graph is generated based on the PC algorithm;

[0015] The production process correlation graph is generated based on the basic production process correlation graph and the production linear weight, production nonlinear weight, and production time lag weight.

[0016] As a further scheme of the present application, a real-time production process control set is obtained, a production process procedure agent cluster is constructed based on the real-time production process control set and the production process correlation graph, and a production process optimization control strategy is generated through a virtual optimization strategy of the agent cluster, including:

[0017] The agent environment is initialized based on the real-time production process control set and the production process correlation graph, and a production process procedure agent cluster is generated;

[0018] Process agent observation data is obtained based on the production process procedure agent cluster, the process agent observation data is processed based on a triple specificity algorithm, and a production process control original strategy is obtained;

[0019] The production process control original strategy is optimized again, and a production process optimization strategy is obtained;

[0020] A real-time production process correlation set is obtained, a production utility target function is defined based on the real-time production process correlation set and the production process optimization strategy, and a multi-dimensional utility vector is obtained based on the production utility target function;

[0021] A collaborative control algorithm is executed based on the multi-dimensional utility vector, a multi-dimensional collaborative utility vector is obtained, the production utility target function is adjusted based on the multi-dimensional collaborative utility vector, and a production process optimization control strategy is obtained.

[0022] As a further scheme of the present application, the production process optimization strategy is acquired, comprising:

[0023] The first agent, the second agent and the third agent of the production process procedure are defined based on the real-time production process control set and the production process graph, and are combined into the agent cluster of the production process procedure, and the first agent observation state data, the second agent observation state data and the third agent observation data are acquired based on the agent cluster of the production process procedure;

[0024] The first agent observation state data is imported into the PSO algorithm, the second agent observation state data is imported into the DDPG algorithm, and the third agent observation data is imported into the Q-learning algorithm, and the production process control original strategy is acquired;

[0025] The production process control original strategy is secondarily optimized based on the Nash Q-learning algorithm, the pending production process optimization strategy is acquired, the physical constraint data is acquired, the pending production process optimization strategy is verified based on the physical constraint data, and after the verification is completed, the production process optimization strategy is acquired.

[0026] As a further scheme of the present application, the pending production process optimization strategy is verified based on the physical constraint data, and after the verification is completed, the production process control original strategy is acquired, comprising:

[0027] The physical constraint condition is acquired based on the physical constraint data;

[0028] If the pending production process optimization strategy satisfies the physical constraint condition, the pending production process optimization strategy is redefined as the production process control original strategy;

[0029] If the pending production process optimization strategy does not satisfy the physical constraint condition, the pending production process optimization strategy is corrected until the physical constraint condition is satisfied, and the production process control original strategy is acquired.

[0030] As a further scheme of the present application, the multi-dimensional utility vector is acquired based on the production utility target function, comprising:

[0031] The first utility data, the second utility data and the third utility data are acquired based on the production utility target function, and the multi-dimensional utility vector is generated based on the first utility data, the second utility data and the third utility data;

[0032] Introduce the multi-dimensional utility vector into the GA algorithm, generate an initial utility population based on the GA algorithm, and perform mutation and crossover operations on the initial utility population to obtain a secondary utility population, screen the secondary utility population based on a tournament strategy, and output the to-be-determined multi-dimensional utility vector based on the GA algorithm after screening is completed;

[0033] Extract N groups of data in the secondary utility population as a strategy candidate set, introduce the strategy candidate set and the to-be-determined multi-dimensional utility vector into the NSGA-II algorithm, optimize the to-be-determined multi-dimensional utility vector based on the NSGA-II algorithm, and obtain a multi-dimensional utility vector.

[0034] As a further scheme of the present application, the production process optimization control strategy is verified by digital twinning to generate a production process final control strategy, comprising:

[0035] A virtual production process working condition network is constructed, the production process optimization control strategy is introduced into the virtual production process working condition network, and virtual working condition data sets are generated based on the virtual production process working condition network;

[0036] A production process working condition prediction network is constructed, the virtual working condition data sets and the production process optimization control strategy are introduced into the production process working condition prediction network, and a strategy-quality prediction mapping table is obtained based on the production process working condition prediction network;

[0037] A historical working condition defect probability is obtained, a production process working condition entropy value calculation model is constructed based on the historical working condition defect probability and the strategy-quality prediction mapping table, and a strategy risk entropy data set is obtained based on the production process working condition entropy value calculation model;

[0038] A control strategy decision surface is established, the strategy risk entropy data set, the strategy-quality prediction mapping table and the production process optimization control strategy are introduced into the control strategy decision surface, and a production process final control strategy is generated based on the control strategy decision surface.

[0039] As a further scheme of the present application, the method further comprises:

[0040] The production process working condition network is established based on a GAN network and comprises a generator, a discriminator and a loss function, the generator is constructed by a ResNet-18 network, and the discriminator is constructed by a GAN network;

[0041] The production process working condition prediction network is established based on an LSTM network and comprises double channels, one group of the double channels is a process channel, and the other group of the double channels is a quality channel;

[0042] The control strategy decision surface is established based on a B-spline surface.

[0043] In still another aspect, the application also provides a production process optimization control system for the light-shielding adhesive tape, which comprises:

[0044] a graph module for obtaining a production process associated dataset and extracting key information in the production process associated dataset to construct a production process associated graph.

[0045] a primary strategy module for obtaining a real-time production process control set and constructing a production process procedure intelligent agent cluster of the real-time production process control set and the production process associated graph, and generating a production process optimization control strategy through a virtual optimization strategy of the intelligent agent cluster;

[0046] a secondary strategy module for digital twin verification and generating a production process final control strategy;

[0047] a regulation module for real-time regulation based on the production process final control strategy;

[0048] a reconstruction module for obtaining associated data after regulation and updating the production process associated graph.

[0049] Based on the above aspects, the application first constructs a production process associated graph by obtaining a production associated dataset and extracting key information in the production process associated dataset, thereby providing a low-dimensional and interpretable physical constraint input for subsequent processing. Secondly, the application constructs a production process procedure intelligent agent cluster of a real-time production process control set and the production process associated graph, and can use a combination of PSO algorithm, DDPG algorithm and Q-learing algorithm to process different types of observation state data in the production process procedure intelligent agent cluster, thereby providing an initial strategy initial value. Thirdly, the application performs digital twin verification on the production process optimization control strategy to generate a production process final control strategy, thereby eliminating low-quality strategies in the initial strategy and obtaining an optimal control strategy. Finally, the application updates the production process associated graph based on associated data after regulation, thereby achieving global control of the light-shielding adhesive tape process procedure and avoiding the situation of local triggering global deterioration caused by traditional single-process parameter optimization. In this way, the coupling effect is adapted, and through digital twin verification, risky strategies in the initial strategy can be excluded in all directions, thereby warning of the compounded risks and ensuring that the optimal control strategy output finally has process procedure global control and risk guarantee. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1is an execution flow schematic diagram of a production process optimization control method of a light shielding adhesive tape provided by an embodiment of the present application.

[0051] Figure 2 is an execution flow schematic diagram of step S200 in a production process optimization control method of a light shielding adhesive tape provided by an embodiment of the present application.

[0052] Figure 3 is a schematic diagram of a production process optimization control system of a light shielding adhesive tape provided by an embodiment of the present application.

[0053] Figure 4 is an execution flow schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0054] The accompanying drawings, which are included to provide a further understanding of the embodiments of the present application and are incorporated in and constitute a part of this application, illustrate embodiments of the present application and together with the description serve to explain the principles of the present application. In the drawings:

[0055] Therefore, the detailed description of the embodiments of the present application provided in the following description and in the accompanying drawings is merely exemplary and explanatory, and is not intended to limit the scope of the application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of the present application.

[0056] The present application will be described in detail below with reference to the accompanying drawings, Figure 1 is an execution flow schematic diagram of a production process optimization control method of a light shielding adhesive tape provided by an embodiment of the present application, Figure 2 is an execution flow schematic diagram of step S200 in a production process optimization control method of a light shielding adhesive tape provided by an embodiment of the present application, which will be described in detail below.

[0057] Specifically, a production process optimization control method of a light shielding adhesive tape comprises:

[0058] Step S100, acquiring a production process associated data set, extracting key information from the production process associated data set, and establishing a production process associated graph.

[0059] Further, the step S100 comprises the following specific steps:

[0060] Step S1000, acquiring a production process associated data set based on a production monitoring device group, and performing data processing operations on the production process associated data set.

[0061] Step S1001, extracting the production process key indicators in the production process correlation data set through the neural network, and establishing a production process key matrix based on the production process key indicators.

[0062] Step S1002, calculating the production linear weight, production nonlinear weight and production time lag weight based on the production process key matrix, and importing the production process key matrix into the PC algorithm to generate a basic production process correlation graph based on the PC algorithm.

[0063] Step S1003, generating the production process correlation graph based on the basic production process correlation graph and the production linear weight, production nonlinear weight and production time lag weight.

[0064] It can be understood that the process procedure flow of the light shielding adhesive tape can include process procedures such as raw material preparation and processing, coating processing, curing processing, die cutting processing, etc. The above four basic process procedures are described. For raw material preparation and processing, the substrate is selected according to product requirements and the black adhesive is prepared or selected according to the final use of the light shielding adhesive tape. For coating processing, the prepared black adhesive is uniformly coated on the surface of the substrate to form a light shielding function. During the coating process, the thickness of the coating needs to be controlled, which directly affects the light shielding effect. Then it is dried and cured. During the drying and curing process, it also needs to be controlled to prevent the phenomenon of blistering and cracking. For curing processing, the material after compounding is stored to ensure that the adhesive completes the final crosslinking reaction. For die cutting processing, the material after curing is cut on a slitting machine according to customer requirements. This method is based on the basic coating, curing and slitting production process procedures. In actual use, the actual production process procedures of the light shielding adhesive tape can be adjusted and expanded to match different implementation methods.

[0065] It can be understood that the production monitoring equipment group is used to monitor the multi-source data in the light shielding adhesive tape production process procedure. The production monitoring equipment can include industrial high-speed cameras, vibration sensors, temperature and humidity sensors, infrared thermographs, etc. The multi-source data collected by the monitoring equipment can include surface image data collected by industrial high-speed cameras, vibration frequency spectrum data of motor rotation collected by vibration sensors, temperature gradient data of ovens collected by infrared thermographs, etc. The multi-source data collected by the production monitoring equipment is formed into a production process correlation data set, and then the production process correlation data set can be subjected to data processing operations.

[0066] Further, the production process correlation data set is spatio-temporally aligned to provide a unified data basis for subsequent processes, solving the time dislocation and spatial offset problems of multi-source production process correlation data sets, which can be processed through a clock synchronization network.

[0067] In this embodiment, the data time of the production monitoring devices such as industrial high-speed cameras and vibration sensors can be unified to a range that meets the error by using an IEEE 1588 PTP clock synchronization network, and the time sequence confusion caused by protocol errors in traditional systems can be eliminated. Different interface protocols used by different production monitoring devices can use a double mechanism of hardware trigger alignment and software timestamp injection, such as a 2xxFPS image stream for a high-speed camera and a 5xkS / s spectrum for a vibration sensor, so that they are strictly matched with the displacement coordinates of the conveyor belt used in production in the time dimension, and a dynamic spatial mapping model is established to calculate the spatial offset of the data of each device station in real time according to the speed of the conveyor belt used in production, for example, to make the thickness measurement point at the outlet of the coating machine coincide with the temperature measurement point at the inlet of the oven in the coordinate system of the virtual production line, and finally make the timestamps of the multi-source production process correlation data set consistent, providing a time-space consistent observation benchmark for subsequent production process optimization.

[0068] Further, after the time-space alignment of the production process correlation data set, data cleaning is required, such as removing data with temperature jumps and vibration saturation. For example, if the temperature rise of adjacent sampling points is greater than 5 degrees Celsius, it is marked as abnormal and removed. This situation may be caused by poor contact of the thermocouple, and removing abnormal data of the production monitoring device group can eliminate distorted data from the source. After removal, related data may be missing, and at this time, it needs to be filled. For this purpose, linear interpolation filling, ARIMA time series prediction model reconstruction, LSTM time series prediction model reconstruction, and other methods can be used to fill the missing data to ensure data integrity.

[0069] In some possible embodiments, data normalization can also be performed. Softmax function can be used for normalization processing. Softmax function converts each element in the channel attention weight vector into a probability value. These probability values represent the relative importance of each channel in the fusion process, so that the data is normalized to a unified interval, simplifying subsequent processing.

[0070] Specifically, the multi-source high-dimensional production process correlation data set can be refined by a neural network, and quantified into quantifiable and interpretable process state features. Through neural network-driven feature engineering, the production process correlation data after data processing can be feature extracted. The extractable features can include, for example, calculating the standard deviation of the thickness measurement point to quantify the coating uniformity, extracting the temperature gradient of the oven to represent the heat distribution uniformity, or extracting the vibration energy integral in the xxx-xxx Hz frequency band in the vibration spectrum, etc. Through the above feature extraction, the multi-source high-dimensional production process correlation data set can be converted into a multi-source low-dimensional production process key indicator. Then, the production process key indicators are combined to establish a production process key matrix.

[0071] It can be understood that by extracting the production process key indicators in the production process correlation data set, the multi-source high-dimensional data can be reduced in dimension while retaining the physical meaning of the process, and the original data can be distilled into key data features, providing low-dimensional and interpretable inputs for subsequent establishment of production process correlation graphs.

[0072] Specifically, before step S1002 is performed, the corresponding historical production defect data is imported, so that the graph of the quantitative causal relationship between the process correlation parameters and the quality defects can be established, and an interpretable data basis is realized.

[0073] Further, the production linear weight, the production nonlinear weight, and the production time lag weight are used to represent the causal chain type of the subsequently established graph. For example, the correlation between the temperature rise rate and the light transmission exceeding the standard in the historical production defect data is a production nonlinear relationship, and the production linear weight thereof is obtained. For example, the coating gap and the edge warping probability are production linear relationships, and the production nonlinear weight thereof is obtained. For example, the generation of bubbles caused by increased vibration has a delay effect, and there is a delay between increased vibration and bubble generation, which is a production time lag relationship, and the production time lag weight thereof is obtained.

[0074] It can be understood that the PC algorithm used in the embodiment is a causal discovery method based on conditional independence test, which is used to identify the causal network structure between variables from observation data, which can break through the limitation of traditional statistical methods that can only judge correlation. In the specific use process, the key matrix of the production process and the historical production defect data are imported into the PC algorithm. First, a complete undirected graph is constructed, that is, all variables are connected to each other. Then, the minimum independent set is found by step-by-step deletion. For a pair of variables (X, Y), if there is a conditional set Z such that X Y|Z (X and Y are independent given Z), then the X-Y edge is deleted. Secondly, the partial causal orientation is determined by using the V structure rule, that is, when X→Z←Y and X and Y have no direct connection, if X Y but X not Y|Z, then a causal collision is formed. Finally, the global causal direction is determined based on the direction propagation rule, that is, to avoid new V structure and ring. This algorithm quantifies conditional independence and outputs the basic production process correlation graph.

[0075] It can be understood that through step S1000, the multi-source data of the production process correlation data set can be aligned in time and space, solving the time sequence confusion caused by protocol errors and spatial errors in traditional systems. Secondly, abnormal data points can be removed from the multi-source data through data cleaning, and the removed data points can be reasonably filled by using the filling method. Finally, the causal chain is obtained based on the PC algorithm, and different linear weights, nonlinear weights and time lag weights are quantified, and the production process correlation graph is constructed, providing a consistent and interpretable data basis for subsequent intelligent optimization.

[0076] Step S200, obtaining a real-time production process control set, constructing a production process procedure agent cluster based on the real-time production process control set and the production process correlation graph, and generating a production process optimization control strategy through the virtual optimization strategy of the agent cluster.

[0077] Further, the step S200 includes the following specific steps:

[0078] Step S2000, initializing the agent environment based on the real-time production process control set and the production process correlation graph, and generating a production process procedure agent cluster.

[0079] Step S2001, obtaining process agent observation data based on the production process procedure agent cluster, processing the process agent observation data based on a triple specificity algorithm, and obtaining a production process control original strategy.

[0080] Step S2002, performing secondary optimization on the production process control original strategy, and obtaining a production process optimization strategy.

[0081] Further, the steps S2000 to S2002 specifically include the following steps:

[0082] Step S2000-1, defining production process procedure first agent, production process procedure second agent and production process procedure third agent based on the real-time production process control set and the production process correlation graph, and combining them into the production process procedure agent cluster, and obtaining first agent observation state data, second agent observation state data and third agent observation data based on the production process procedure agent cluster.

[0083] Specifically, the real-time production process control set can include relevant data such as real-time thickness data, real-time temperature data, real-time tension data, etc. during the process of the production process procedure. The production process procedure first agent can be represented as a coating agent, the production process procedure second agent can be represented as a curing agent, and the production process procedure third agent can be represented as a slitting agent. By analyzing the production process correlation graph, the abstract process graph knowledge is converted into a mathematical boundary that can be understood by a machine. Then, by fusing the real-time production process control set, the observation state data is dynamically generated, forming a dual cognitive framework of physical rules and real-time state, i.e., coating agent observation state data, curing agent observation state data, and slitting agent state data, which provides cognitive data for subsequent algorithm processing. In this embodiment, coating, curing and slitting are used as examples, but in specific use, they can be redefined and other agents can be added to construct the production process procedure agent cluster.

[0084] Step S2000-2, importing the first agent observation state data into the PSO algorithm, importing the second agent observation state data into the DDPG algorithm, and importing the third agent observation data into the Q-learning algorithm to obtain the production process control original strategy.

[0085] Specifically, for the high-dimensional continuous parameter space of the coating procedure, the PSO algorithm (particle swarm optimization algorithm) is used. Through the various combinations formed by N particles in a short time, the fitness function of thickness uniformity and warping risk is used to output the optimal solution. For the time sequence control of the curing temperature field, the DDPG algorithm (deep deterministic policy gradient algorithm) is used. The Actor network in it generates fan frequency conversion instructions in real time, and the Critic network in it evaluates the trade-off between temperature field stability and energy consumption. For the discrete action requirement of the slitting procedure, the Q-learning algorithm (Q-learning algorithm) can be used. Through the ε-greedy strategy in it, the optimal solution is found in a short time. Through the triple specificity algorithms of PSO algorithm, DDPG algorithm and Q-learning algorithm, the limitations of traditional single algorithm are solved for different types of procedures. PSO can avoid the coating parameters from falling into local optimum, DDPG can solve the time delay control problem of the oven, and Q-learning can realize efficient search of discrete decisions in slitting, providing high-quality strategy initial value for subsequent processing.

[0086] It can be understood that the PSO algorithm can solve the high-dimensional continuous parameter optimization problem by simulating the cooperative foraging behavior of bird flocks, each particle represents a candidate solution such as the combination of coating gap and pressure, the position is updated by tracking the individual historical optimum and the group optimum, the speed update depends on the inertia term, i.e. the early exploration, the individual cognitive term, i.e. learning from the own optimum, and the social cognitive term, i.e. learning from the group optimum, the position iterative approximation global optimum, secondly, the DDPG algorithm combines policy gradient and Q-learning to solve the time series decision problem in continuous action space, the DDPG algorithm has an Actor-Critic double network architecture, the Actor network (policy network) generates continuous actions such as fan frequency values according to the state such as temperature field gradient, the Critic network (value network) evaluates the long-term return of the action, through the soft update of the target network and the exploration of the OU process noise, the strategy is gradually optimized, while Q-learning is based on time difference reinforcement learning, which optimizes the discrete action decision problem, it maintains a Q table to record the state-action value, such as the expected return of the "medium speed + medium tension" combination of the slitting machine, through the ε-greedy strategy to balance exploration and utilization, such as 30% random attempt of new action, the Q value update combines the immediate reward and the future discounted return.

[0087] Step S2000-3, based on the Nash Q-learning algorithm, the production process control original strategy is optimized twice to obtain the pending production process optimization strategy, the physical constraint data is obtained, the pending production process optimization strategy is verified based on the physical constraint data, after the verification is completed, the production process optimization strategy is obtained.

[0088] Specifically, after generating the initial production process control original strategy, the utility value of each combination in the production process optimization strategy is calculated based on the Nash Q-learning algorithm (Nash Q-learning algorithm), then the strategy is updated through the Boltzmann exploration balance, the Nash equilibrium point is determined by the convergence of the Q value, ensuring that the deviation of the single strategy will lose the utility value, solving the multi-objective conflict problem in the production process optimization strategy, such as the problem of improving the quality of coating and preventing the slitting machine from breaking the tape, achieving global optimization.

[0089] It can be understood that the Nash Q-learning algorithm solves the multi-agent game problem by extending the traditional Q-learning framework, and upgrades the state-action value function Q of a single agent to the Nash Q function under the joint action space, wherein each agent evaluates the expected cumulative reward of all joint actions of the agents under the state; the algorithm updates the Q value through iteration, replaces the maximum utility of the traditional Q-learning with the Nash equilibrium utility, that is, the Q value corresponding to the Nash equilibrium solution of the next state instead of max Q, and balances exploration and utilization by using the Boltzmann exploration strategy, and finally converges to the Nash equilibrium point, at which time any unilateral deviation of an agent from the strategy will reduce its own benefit, thereby coordinating the target conflicts of the coating, curing, slitting and other process agents in the masking tape production, and achieving multi-party collaborative optimization.

[0090] Further, step S2000-3 specifically comprises the following steps:

[0091] Step S2000-3-1, obtaining physical constraint conditions based on physical constraint data.

[0092] Specifically, the physical constraint data can be extracted based on the production process correlation graph, that is, the relevant data during data cleaning, so as to prevent the parameters in the generated strategy from exceeding the physical limit control parameters such as process equipment or process.

[0093] Step S2000-3-2, if the pending production process optimization strategy satisfies the physical constraint condition, the pending production process optimization strategy is redefined as the production process control original strategy.

[0094] Step S2000-3-4, if the pending production process optimization strategy does not satisfy the physical constraint condition, the pending production process optimization strategy is corrected until the production process control original strategy is obtained after satisfying the physical constraint condition.

[0095] Specifically, the types in the production process optimization strategy that do not satisfy the physical constraint condition are corrected, which may exceed the thermal safety boundary, exceed the equipment physical limit, power system constraint, etc. For the above various cases, it needs to be corrected, for example, the safety limiter correction can be used in the case of exceeding the equipment physical limit, that is, the exceeding parameter is forcibly corrected, such as the pressure limit is 1.2 MPa, but the parameter is 1.25 MPa, which is adjusted to 1.2 MPa.

[0096] Step S2003, obtaining a real-time production process correlation set, defining a production utility target function based on the real-time production process correlation set and the production process optimization strategy, and obtaining a multi-dimensional utility vector based on the production utility target function.

[0097] Step S2004, performing a collaborative control algorithm based on the multi-dimensional utility vector, obtaining a multi-dimensional collaborative utility vector, adjusting the production utility target function based on the multi-dimensional collaborative utility vector, and obtaining a production process optimization control strategy.

[0098] Further, obtaining the multi-dimensional utility vector based on the production utility target function specifically includes the following steps:

[0099] Step S2004-1, obtaining first utility data, second utility data, and third utility data based on the production utility target function, and generating a multi-dimensional utility vector based on the first utility data, the second utility data, and the third utility data.

[0100] Specifically, the real-time production process correlation set can include data such as real-time yield, real-time power, and real-time output rate. Yield utility modeling can be performed based on real-time yield, energy consumption utility modeling can be performed based on real-time power, and production capacity utility modeling can be performed based on real-time output rate. During the modeling process, normalization modeling is required, thereby establishing a suitable production utility target function. The first utility data is the yield weight of the production utility target function, the second utility data is the energy consumption utility weight, and the third utility data is the production capacity utility weight. The yield weight, the energy consumption utility weight, and the production capacity utility weight are combined to form a multi-dimensional utility vector.

[0101] Step S2004-2, importing the multi-dimensional utility vector into a GA algorithm, generating an initial utility population based on the GA algorithm, and performing mutation and crossover operations on the initial utility population to obtain a secondary utility population. The secondary utility population is screened based on a tournament strategy. After screening is completed, a to-be-determined multi-dimensional utility vector is output based on the GA algorithm.

[0102] Specifically, before executing the GA algorithm (genetic algorithm), the GA algorithm needs to be configured and initialized, such as defining energy consumption guides, yield guides, and yield guides. The dynamic fitness function of the energy consumption guide, the yield guide, and the yield guide quantifies the target collaborative value. For example, in the mode of the energy consumption guide, the energy consumption utility weighted X times guide algorithm tends to save energy solutions, or the quality is insufficient to trigger the yield utility minimum X guide algorithm to maintain the yield bottom line. The secondary utility population is generated through SBX crossover and Gaussian mutation, and the tournament is selected, such as reserving the top 30% of the secondary utility population as the secondary utility population. After iterative evolution, the final to-be-determined multi-dimensional utility vector is generated.

[0103] It can be understood that the genetic algorithm simulates the biological evolution mechanism, encodes the optimization problem solution as a chromosome, constructs an initial population, and then performs selection, retains individuals with high fitness, crossover, i.e. exchanges chromosome fragments to generate new solutions, mutation, i.e. random disturbance to prevent premature operation, and makes the population evolve generation by generation to approach the global optimal solution. It can solve high-dimensional, nonlinear, multi-peak complex optimization problems, such as process parameter multi-objective cooperation, and break through the limitations of traditional gradient method which is easy to fall into local optimum. In the production of shading tape, the conflicting objectives of yield, energy consumption and production capacity are dynamically balanced, and the efficient search of the optimal solution set is realized.

[0104] Step S2004-3, extract N groups of data in the secondary utility population as a strategy candidate set, import the strategy candidate set and the pending multi-dimensional utility vector into the NSGA-II algorithm, optimize the pending multi-dimensional utility vector based on the NSGA-II algorithm, and obtain a multi-dimensional utility vector.

[0105] Specifically, through the NSGA-II algorithm, the pending multi-dimensional utility vector is optimized again under the conflicting objectives of quality, energy consumption and production capacity, so as to make a more trade-off. First, a three-dimensional objective function is constructed to quantify the strategy utility. N groups of candidate solutions are stratified by non-dominated sorting. The first front layer contains strategies that are not dominated by any solution. The distribution density of the solution in the objective space is evaluated by congestion degree calculation. The solution in the sparse area is retained to maintain diversity. Finally, a multi-dimensional utility vector is output.

[0106] It can be understood that the GA algorithm compresses multiple objectives into a single objective fitness function, which essentially finds a single optimal solution under a predetermined preference. The NSGA-II algorithm retains all non-dominated solutions through non-dominated sorting. If solution A and solution B are not dominated by each other, they belong to the first front layer. The diversity of the solution set is maintained by congestion degree calculation. Finally, a complete set of Pareto optimal solutions is output. On the basis of the GA algorithm, the NSGA-II algorithm is further executed, thereby generating a more diversified and optimized multi-dimensional utility vector. With this multi-dimensional utility vector, the production utility objective function is adjusted, so that the production process optimization control strategy generated by the production utility objective function has more diversity on the premise of optimizing parameters, thereby providing better reference data for subsequent digital twin verification.

[0107] Step S300, performing digital twin verification on the production process optimization control strategy, generating a production process final control strategy, and performing real-time regulation and control based on the production process final control strategy.

[0108] Further, step S300 specifically includes the following steps:

[0109] Step S3000, a virtual production process condition network is constructed, the production process optimization control strategy is imported into the virtual production process condition network, and a virtual condition dataset is generated based on the virtual production process condition network.

[0110] Specifically, the production process condition network is established based on a GAN network and includes a generator, a discriminator, and a loss function.

[0111] Specifically, the GAN network (Generative Adversarial Network) learns the real data distribution through the game of the generator and the discriminator, constructs a virtual condition dataset covering extreme disturbances and small probability risks through the GAN network, takes the production process optimization control strategy as input, and is constructed through the combination of the generator, the discriminator, and the loss function, wherein the generator adopts a ResNet-18 network for learning the disturbance distribution law of the production process condition, the discriminator adopts a CNN network for improving the physical rationality of the generated data, and the loss function is designed and determined based on the actual situation for improving the training stability of the GAN network. Synthesize various industrial disturbances, including but not limited to sudden changes in temperature and humidity, sudden increases in vibration, etc., and add physical rules when synthesizing industrial disturbances to prevent generated data from exceeding physical limits, thereby outputting multiple virtual condition datasets in space and time, providing a more abundant scene library for subsequent strategy verification than directly using historical data.

[0112] Step S3001, a production process condition prediction network is constructed, the virtual condition dataset and the production process optimization control strategy are imported into the production process condition prediction network, and a strategy-quality prediction mapping table is obtained based on the production process condition prediction network.

[0113] Specifically, the production process condition prediction network is established based on an LSTM network and includes two channels, one of which is a process channel and the other of which is a quality channel. The process channel is used to predict the time series changes of thickness, temperature, tension, etc., and the quality channel is used to predict the defect changes of warping probability, light transmittance, etc.

[0114] Specifically, the LSTM network quantifies the risk of the control strategy based on the virtual condition dataset generated in step S3000 through the double-channel mechanism, realizes the function from passive monitoring to active defense, first inputs the virtual condition dataset and the production process optimization control strategy into the LSTM network, captures the cumulative effect of time series disturbances based on the encoder of the LSTM, such as too fast temperature rise rate that may cause the crosslinking degree of the glue layer to decrease, then analyzes the spatial distribution, outputs the multi-dimensional defect probability, such as the edge warping probability of 0.18 and the bubble generation probability of 0.35, and further outputs the prediction probability mapping table of the strategy and the quality in the production process optimization control strategy through counterfactual intervention analysis.

[0115] In step S3002, the historical working condition defect probability is obtained, a production process working condition entropy value calculation model is constructed based on the historical working condition defect probability and a strategy-quality prediction mapping table, and a strategy risk entropy data set is obtained based on the production process working condition entropy value calculation model.

[0116] Specifically, a unified decision scale across defect types is constructed by information entropy, traditional experience-driven fuzzy judgment is converted into a calculable mathematical index, a risk entropy model is established by KL divergence, a deviation value of each defect occurrence probability under a current strategy working condition is calculated based on a defect probability distribution of a historical normal working condition, for example, an edge warpage of 5%, a light transmittance exceeding a standard of 3%, and a bubble generation of 5%, for example, a temperature rise of 4 degrees Celsius may cause a light transmittance to rise by 7%, and a self-adaptive learning mechanism is added, for example, after the production of every 200 volumes of light shielding adhesive tape, the benchmark is updated and distributed, so that the model is real-time adapted to changes in the working condition, and the entropy weight in the production process working condition entropy value calculation model is defined in combination with the strategy-quality prediction mapping table, so as to establish the production process working condition entropy value calculation model, and a strategy risk entropy data set is generated by the production process working condition entropy value calculation model.

[0117] In step S3003, a control strategy decision surface is established, the strategy risk entropy data set, the strategy-quality prediction mapping table, and the production process optimization control strategy are imported into the control strategy decision surface, and a production process final control strategy is generated based on the control strategy decision surface.

[0118] Specifically, the control strategy decision surface is established based on a B-spline surface.

[0119] Specifically, the abstract multi-objective optimization is converted into an interactive geometric surface by establishing the control strategy decision surface, so as to generate a final real-time control strategy, and the production process procedure of the light shielding adhesive tape is real-time optimized and adjusted based on the final real-time control strategy, first, the strategy-quality prediction mapping table is taken as the X axis, the strategy risk entropy data set is taken as the Y axis, and the production process optimization control strategy is taken as the Z axis, a continuous decision space is generated by fitting a third-order B-spline surface, the process constraint is dynamically mapped based on the surface form, and the optimal path is automatically extracted based on a gradient ascent algorithm, so as to generate a final production process control strategy, and finally, the final production process control strategy is packaged as a control instruction and is sent to each device used for production in the process procedure.

[0120] In step S4000, associated data after regulation and control is obtained, and the production process association graph is updated based on the associated data after regulation and control.

[0121] In some possible embodiments, after the final regulation and control ends, a related step of updating the production process association graph generated in step S100 can be added.

[0122] In some possible embodiments, the production process correlation graph can be updated and reconstructed in a manner such as counterfactual reasoning and dynamic knowledge reconstruction. For example, actual yield rate data and real-time production process control set, i.e., correlation data after regulation, can be obtained by first performing deep analysis on the deviation data, and the correlation data can be analyzed by using a time series mining algorithm. For example, the Matrix Profile algorithm can be used to calculate the dynamic time warping distance between the actual data and the yield rate, so as to locate the maximum deviation time window. A deviation event report can be output based on a Bayesian diagnostic network, and then a random forest algorithm can be used for reasoning to output a diagnostic report of causal reasoning. Finally, the diagnostic report and the production process correlation graph are analyzed to obtain an updated production process correlation graph.

[0123] Figure 3 A schematic diagram of a production process optimization control system of a light-blocking adhesive tape is shown, and the production process optimization control method of the light-blocking adhesive tape is described in detail below.

[0124] Specifically, a production process optimization control system of a light-blocking adhesive tape includes:

[0125] A graph module is configured to obtain a production process correlation data set, extract key information in the production process correlation data set, and construct a production process correlation graph.

[0126] A primary strategy module is configured to obtain a real-time production process control set, construct a production process procedure agent cluster of the real-time production process control set and the production process correlation graph, and generate a production process optimization control strategy by using a virtual optimization strategy of the agent cluster.

[0127] A secondary strategy module is configured to perform digital twin verification and generate a production process final control strategy.

[0128] A regulation module is configured to perform real-time regulation based on the production process final control strategy.

[0129] A reconstruction module is configured to obtain correlation data after regulation and update the production process correlation graph.

[0130] The specific use and role of the embodiment are described below.

[0131] The method first acquires a production correlation data set, extracts key information from the production process correlation data set, and constructs a production process correlation graph, thereby providing low-dimensional and interpretable physical constraint inputs for subsequent processing. Secondly, by acquiring real-time production process control sets and constructing production process procedure agent clusters of real-time production process control sets and production process correlation graphs, different types of observation state data in the production process procedure agent clusters can be processed by a combination of PSO algorithm, DDPG algorithm and Q- learning algorithm, thereby providing initial strategy initial values. Then, the production process optimization control strategy is verified by digital twinning, and the final production process control strategy is generated, thereby eliminating low-quality strategies in the initial strategy and obtaining the optimal control strategy. Finally, the production process correlation graph can be updated based on the regulated correlation data, realizing global control of the shading tape process procedure and avoiding the local triggering of global deterioration caused by traditional single-process parameter optimization. In this way, the coupling effect is adapted, and through digital twinning verification, risky strategies in the initial strategy can be excluded in all directions, thereby warning of complex risks and ensuring that the optimal control strategy output has process procedure global control and risk guarantee control strategy, thereby providing a shading tape production process optimization control method and system.

[0132] Figure 4 A schematic diagram of an electronic device provided by some embodiments of the present application is shown, and the electronic device will be described in detail below.

[0133] Specifically, an electronic device includes:

[0134] At least one processor; and a memory in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method proposed in Embodiment One of the present application.

[0135] The various components of the electronic device will be described in detail below:

[0136] The processor is the control center of the electronic device, and can be one processor or a combination of multiple processing elements. For example, the processor is one or more central processing units (CPUs), application specific integrated circuits (ASICs), or one or more integrated circuits configured to implement an embodiment of the present application, such as one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs).

[0137] The processor can execute various functions of the electronic device by running or executing software programs stored in the memory and calling data stored in the memory.

[0138] The memory is used to store software programs for implementing the present application, and is controlled by the processor to execute. The specific implementation manner can refer to the above method embodiments, and will not be described here.

[0139] The memory can be a real-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only (CD-ROM), or other optical disk storage, optical disk storage (including compact disks, laser disks, optical disks, digital versatile disks, Blu-ray disks, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and capable of being accessed by a computer, but not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through the interface circuit of the electronic device, and the present application is not limited in this regard.

[0140] The above-described embodiments can be implemented in whole or in part by software, hardware (such as a circuit), firmware, or any combination thereof. When implemented in software, the above-described embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center by limited (for example, infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0141] It should be understood that the term "and / or" herein merely describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the existence of A alone, the existence of A and B, and the existence of B alone, where A and B can be singular or plural. In addition, the character " / " herein generally represents that the associated objects before and after are an "or" relationship, but can also represent an "and / or" relationship, which can be understood according to the context before and after.

[0142] It should be understood that in the embodiments of the present application, the size of the sequence number of the above-mentioned processes does not mean the order of execution, and the execution order of the processes should be determined according to their functions and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0143] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A production process optimization control method of a light shielding tape, characterized by, The method comprises: S100, acquiring a production process correlation data set, extracting key information from the production process correlation data set, and establishing a production process correlation graph; S200, acquiring a real-time production process control set, constructing a production process procedure intelligent agent cluster based on the real-time production process control set and the production process correlation graph, and generating a production process optimization control strategy through an intelligent agent cluster virtual optimization strategy; Intelligent agent environment initialization based on the real-time production process control set and the production process correlation graph generates a production process procedure intelligent agent cluster; Based on the production process procedure intelligent agent cluster, process intelligent agent observation data is obtained, the process intelligent agent observation data is processed based on a triple specificity algorithm, and a production process control original strategy is obtained; The production process control original strategy is optimized again to obtain a production process optimization strategy; Acquire real-time production process correlation set, define production utility target function based on real-time production process correlation set and production process optimization strategy, obtain multi-dimensional utility vector based on production utility target function; Based on the multi-dimensional utility vector, a collaborative control algorithm is executed to obtain a multi-dimensional collaborative utility vector, the production utility target function is adjusted based on the multi-dimensional collaborative utility vector, and a production process optimization control strategy is obtained; Based on the real-time production process control set and the production process correlation graph, a production process procedure first intelligent agent, a production process procedure second intelligent agent, and a production process procedure third intelligent agent are defined, the production process procedure first intelligent agent is represented as a coating intelligent agent, the production process procedure second intelligent agent is represented as a curing intelligent agent, and the production process procedure third intelligent agent is represented as a slitting intelligent agent, and they are combined into the production process procedure intelligent agent cluster. Based on the production process procedure intelligent agent cluster, first intelligent agent observation state data, second intelligent agent observation state data, and third intelligent agent observation data are obtained, the first intelligent agent observation state data is represented as coating intelligent agent observation state data, the second intelligent agent observation state data is represented as curing intelligent agent observation state data, and the third intelligent agent observation data is represented as slitting intelligent agent state data; The first intelligent agent observation state data is imported into the PSO algorithm, the second intelligent agent observation state data is imported into the DDPG algorithm, and the third intelligent agent observation data is imported into the Q-learning algorithm to obtain a production process control original strategy; Based on the Nash Q-learning algorithm, the production process control original strategy is optimized again to obtain a to-be-determined production process optimization strategy, physical constraint data is obtained, the to-be-determined production process optimization strategy is verified based on the physical constraint data, and after the verification is completed, the production process optimization strategy is obtained; S300, the production process optimization control strategy is verified by digital twinning, a production process final control strategy is generated, and real-time regulation and control are performed based on the production process final control strategy; S400, acquire the associated data after regulation and control, and update the production process correlation graph based on the associated data after regulation and control.

2. The method of claim 1, wherein the adhesive tape is a light shielding adhesive tape. Acquire a production process correlation dataset, perform key information extraction on the production process correlation dataset, and establish a production process correlation graph, including: Acquire a production process correlation dataset based on a production monitoring device group, and perform data processing operations on the production process correlation dataset; Extract production process key indicators from the production process correlation dataset using a neural network, and establish a production process key matrix based on the production process key indicators; Calculate production linear weights, production nonlinear weights, and production time lag weights based on the production process key matrix, and import the production process key matrix into a PC algorithm to generate a basic production process correlation graph based on the PC algorithm; Generate the production process correlation graph based on the basic production process correlation graph and the production linear weights, production nonlinear weights, and production time lag weights.

3. The method of claim 1, wherein the adhesive tape is a light shielding adhesive tape. Verify the pending production process optimization strategy based on physical constraint data, and after verification is complete, acquire the production process control original strategy, including: Acquire physical constraint conditions based on physical constraint data; If the pending production process optimization strategy meets the physical constraint conditions, redefine the pending production process optimization strategy as the production process control original strategy; If the pending production process optimization strategy does not meet the physical constraint conditions, modify the pending production process optimization strategy until the physical constraint conditions are met, and then acquire the production process control original strategy.

4. The method of claim 1, wherein the adhesive tape is a light shielding adhesive tape. Acquire a multi-dimensional utility vector based on the production utility target function, including: Acquire first utility data, second utility data, and third utility data based on the production utility target function, and generate a multi-dimensional utility vector based on the first utility data, second utility data, and third utility data; The first utility data represents a good product rate weight of the production utility target function, the second utility data represents an energy consumption utility weight, and the third utility data represents a production capacity utility weight; Import the multi-dimensional utility vector into a GA algorithm, generate an initial utility population based on the GA algorithm, and perform mutation and crossover operations on the initial utility population to acquire a secondary utility population. Screen the secondary utility population based on a tournament strategy, and after screening is complete, output a pending multi-dimensional utility vector based on the GA algorithm; Extract N groups of data from the secondary utility population as a strategy candidate set, import the strategy candidate set and the pending multi-dimensional utility vector into an NSGA-II algorithm, optimize the pending multi-dimensional utility vector based on the NSGA-II algorithm, and acquire a multi-dimensional utility vector.

5. The method of claim 1, wherein the method is characterized by: Perform digital twin verification on the production process optimization control strategy to generate a production process final control strategy, including: Construct a virtual production process working condition network, import the production process optimization control strategy into the virtual production process working condition network, and generate a virtual working condition dataset based on the virtual production process working condition network; Construct a production process working condition prediction network, import the virtual working condition dataset and production process optimization control strategy into the production process working condition prediction network, and acquire a strategy-quality prediction mapping table based on the production process working condition prediction network; The method further comprises: The production process condition network is established based on a GAN network and comprises a generator, a discriminator, and a loss function. The generator is constructed by a ResNet-18 network, and the discriminator is constructed by a GAN network; 6. The method of claim 5, wherein the adhesive tape is a light shielding adhesive tape. The production process condition prediction network is established based on an LSTM network and comprises double channels. One set of the double channels is a process channel, and the other set of the double channels is a quality channel; The control strategy decision surface is established based on a B-spline surface. The system comprises: A graph module, which is used to obtain a production process correlation data set, extract key information in the production process correlation data set, and construct a production process correlation graph; 7. A production process optimization control system for light shielding adhesive tape, for implementing the method of any one of claims 1 to 6, characterized by, A primary strategy module, which is used to obtain a real-time production process control set, construct a production process procedure intelligent agent cluster of the real-time production process control set and the production process correlation graph, and generate a production process optimization control strategy through virtual optimization of the intelligent agent cluster; A secondary strategy module, which is used to perform digital twin verification and generate a production process final control strategy; A regulation and control module, which is used to perform real-time regulation and control based on the production process final control strategy; A reconstruction module, which is used to obtain correlation data after regulation and control and update the production process correlation graph. ​ ​

Images