A full-process intelligent control method and system for a glass bead production line

Abstract

The application relates to a full-process intelligent control method and system of a glass bead production line, and belongs to the technical field of control systems, which comprises the following steps: collecting multi-source heterogeneous sensing data, and performing preprocessing and time synchronization through an edge computing node; constructing a dynamic simulation system of the whole process of glass bead production based on a digital twin model; combining a digital twin model with a reinforcement learning agent mode to output deviations between preset quality targets, and adaptively generating a multi-actuator collaborative control strategy; evaluating indexes of generated glass beads online, and feeding back evaluation results to the reinforcement learning agent mode to realize closed-loop optimization of the control strategy; when abnormal working conditions or quality drift are detected, automatically triggering a fault tracing module based on causal reasoning, locating a disturbance source, and generating a process parameter compensation scheme, so that the whole process can maintain stable operation with high yield without human intervention; the application has the beneficial effect of improving intelligent management and control of glass bead production.

Description

Technical Field

[0001] This invention belongs to the field of control system technology, and specifically relates to a whole-process intelligent control method and system for a glass bead production line. Background Technology

[0002] Glass beads, as a versatile functional material, have important applications in road reflective markings, industrial abrasives, and decorative materials. However, as the market demands increasingly higher standards for core indicators of glass beads, such as particle size uniformity, sphericity, and strength, traditional glass bead production lines are gradually revealing problems such as reliance on manual operation, low process control precision, and limited production efficiency.

[0003] Existing glass bead production processes typically include multiple stages such as raw material pretreatment, melting, bead formation, cooling, and screening. Each stage often employs independent control, lacking overall process coordination. For example, raw material ratios rely on manual experience for adjustment, making it difficult to optimize in real time based on fluctuations in raw material composition. Key process parameters such as melting temperature and bead formation airflow velocity are mostly fixed values, failing to dynamically adapt to changes in production conditions, resulting in poor product quality stability and a high rate of defective products.

[0004] Meanwhile, traditional production line parameter monitoring relies heavily on manual inspections, resulting in delayed and inaccurate data collection, making it difficult to quickly detect and address anomalies in the production process. Furthermore, the lack of intelligent scheduling mechanisms between different processes easily leads to material accumulation or untimely supply issues, further impacting production efficiency.

[0005] In terms of control technology applications, existing control systems are mostly designed for single processes, with relatively limited functions and a lack of integrated analysis and intelligent decision-making capabilities for data across the entire process. Although some production lines have introduced simple automated equipment, poor compatibility between devices and insufficient data interoperability create "information silos," making it impossible to achieve intelligent management and control of the entire process. Summary of the Invention

[0006] This invention provides a fully intelligent control method and system for a glass bead production line, which addresses the technical problem of poor intelligent control technology in existing glass bead production. It adopts a multi-actuator collaborative control strategy for adjustment. When a certain indicator deviates significantly, the attention weight of the corresponding actuator approaches one and is prioritized for scheduling; when the deviation of a certain indicator is small, the attention weight of the corresponding actuator approaches zero, reducing disturbances, avoiding over-adjustment, and improving the intelligent control of glass bead production.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0008] A fully intelligent control method for a glass bead production line includes the following steps:

[0009] Collect multi-source heterogeneous sensor data and perform preprocessing and time synchronization through edge computing nodes;

[0010] A dynamic simulation system for the entire glass bead production process is constructed based on a digital twin model. The digital twin model integrates a physical mechanism model and a deep learning network to map the actual production line status online and predict the viscosity of the molten glass, the size distribution of the beads, and the probability of surface defects.

[0011] Based on the deviation between the output of the preset quality target by combining the reinforcement learning agent method and the digital twin model, an adaptive multi-actuator collaborative control strategy is generated. The multi-actuator collaborative control strategy synchronously adjusts the heating power zone of the melting furnace, the feed screw speed, the airflow forming pressure gradient and the cooling air curtain velocity field.

[0012] The particle size uniformity, sphericity, and optical uniformity of the generated glass beads are evaluated online, and the evaluation results are fed back to the reinforcement learning agent to achieve closed-loop optimization of the control strategy.

[0013] When abnormal operating conditions or quality drift are detected, the fault tracing module based on causal reasoning is automatically triggered to locate the source of disturbance and generate a process parameter compensation scheme to ensure that the entire process maintains stable operation with high yield without human intervention.

[0014] Optionally, real-time data acquisition can be achieved through quantitative monitoring of raw material components, monitoring of the distribution of temperature throughout the furnace, monitoring of the flow rate of molten glass delivery, monitoring of nozzle inlet and outlet pressure and pressure fluctuations, and real-time monitoring of multiple environmental parameters in the cooling zone.

[0015] The heterogeneity of data acquisition is characterized by heterogeneity in data source, format, rate, and accuracy.

[0016] Optionally, for the preprocessing of edge computing nodes, a time offset propagation correction method using graph neural networks is adopted. Since clock offsets have spatial correlations in large-scale heterogeneous edge networks, offset propagation is performed through graph neural network modeling, specifically as follows:

[0017] Each node broadcasts its current offset estimate.

[0018] The receiving node receives the information and performs weighted processing based on link quality;

[0019] Use the node update function to fuse its own historical estimates with neighbor messages;

[0020] Output the new offset estimate.

[0021] Optionally, for time synchronization of edge computing nodes, a clock correction method based on preprocessing feedback is used. This method corrects the clock by introducing preprocessing residual feedback. The approach is as follows:

[0022] Subtracting the consistency-indicating correction from the current parameters yields more stable new parameters. By using feedback from the surrounding environment, the system corrects its own state to achieve coordination.

[0023] Optionally, for digital twin models that integrate physical mechanism models and deep learning networks, a differential-algebraic-neural hybrid model is adopted, which allows the deep learning network to directly modulate the evolution rate of the physical mechanism model.

[0024] Specifically, this means that under new operating conditions, while maintaining the physical structure unchanged and the physical mechanism function effective, the neural network residual can quickly adapt to new deviations.

[0025] Optionally, for the deviation between preset quality targets, the degree of deviation of the system as a whole from the target is represented by the deviation energy. By using the weighted energy function of multidimensional quality deviation, the complex process quality requirements are transformed into scalar targets, and intelligent closed-loop control is executed to find the deviation, learn to correct it, and then automatically optimize the intelligent manufacturing closed-loop control.

[0026] Optionally, for multi-actuator collaborative control strategies, adjustments are made based on action increments:

[0027] The incremental control actions at any given moment are ultimately multiplied element by element to achieve selective activation, ensuring that control commands are both effective and safe.

[0028] When the metric deviation is large, the attention weight of the corresponding actuator is close to 1, and it is prioritized for scheduling.

[0029] When the deviation of the indicator is small, the attention weight of the corresponding actuator is close to 0, which reduces disturbance and avoids over-adjustment.

[0030] Optionally, the complete process of closed-loop optimization of the control strategy is as follows:

[0031] The production line continuously produces glass beads, and the online detection module simultaneously collects data and completes the evaluation of three indicators;

[0032] The standardized evaluation results are fed back to the reinforcement learning agent in real time;

[0033] The agent compares the evaluation results with the preset quality standards, and combines them with the current production parameters to generate the optimal adjustment strategy through an algorithm.

[0034] The actuator receives instructions and adjusts key parameters in the melting, forming, and cooling processes.

[0035] The adjusted production parameters are applied to the production of the next batch of glass beads, and the test data of the new batch of products are fed back to the agent, forming a cycle.

[0036] Optionally, fault diagnosis and parameter correction can be automatically completed by relying on the set algorithm to ensure high yield and stable output of the production line under complex working conditions. Historical data of fault tracing and compensation are fed back to the digital twin model and reinforcement learning agent to continuously optimize the prediction accuracy of the model and the adaptability of the control strategy.

[0037] A fully intelligent control system for a glass bead production line, comprising:

[0038] The multi-source heterogeneous data acquisition and preprocessing module is configured to acquire multi-source heterogeneous sensor data in real time during the glass bead production process, and to perform preprocessing and time synchronization processing on the acquired multi-source heterogeneous sensor data through edge computing nodes.

[0039] The digital twin dynamic simulation module is connected to the multi-source heterogeneous data acquisition and preprocessing module. It has a built-in digital twin model that integrates physical mechanism model and deep learning network, and is configured to build a dynamic simulation system for the entire glass bead production process.

[0040] The reinforcement learning collaborative control module is connected to the digital twin dynamic simulation module and is configured to adaptively generate a multi-actuator collaborative control strategy based on the reinforcement learning agent mechanism and the deviation between the prediction results output by the digital twin dynamic simulation module and the preset quality target.

[0041] The online quality assessment and feedback module is connected to the reinforcement learning collaborative control module and is configured to assess the indicators obtained from production online and feed the assessment results back to the reinforcement learning collaborative control module.

[0042] The fault tracing and parameter compensation module is connected to the online quality assessment and feedback module and configured to monitor the operating conditions of the glass bead production line in real time.

[0043] The beneficial effects of this invention are:

[0044] 1. This invention is based on a reinforcement learning agent approach combined with a digital twin model to output the deviation between preset quality targets. It adaptively generates a multi-actuator collaborative control strategy. The multi-actuator collaborative control strategy synchronously adjusts the heating power zone of the melting furnace, the screw speed of the feeding screw, the airflow forming pressure gradient, and the cooling air curtain velocity field. When the deviation of a certain indicator is large, the attention weight of the corresponding actuator is close to one, and it is given priority scheduling; when the deviation of a certain indicator is small, the attention weight of the corresponding actuator is close to zero, reducing disturbances, avoiding over-adjustment, and improving the intelligent management and control of glass bead production.

[0045] 2. This invention evaluates the particle size consistency, sphericity, and optical uniformity of the generated glass beads online and feeds the evaluation results back to a reinforcement learning agent to achieve closed-loop optimization of the control strategy. The original detection data for the three indicators are uniformly converted into standardized scores from zero to one. The reinforcement learning agent, aimed at improving the quality stability of the glass beads, consists of three parts: state perception, a reward function, and a policy network. The dynamic optimization of the control strategy involves the reinforcement learning agent continuously learning and optimizing based on the feedback quality data. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0048] Figure 2 This is a schematic diagram of the workflow of the present invention. Detailed Implementation

[0049] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0050] Example 1;

[0051] like Figure 1 As shown, this embodiment provides a fully intelligent control system for a glass bead production line, including:

[0052] The multi-source heterogeneous data acquisition and preprocessing module is configured to acquire in real time the glass raw material ratio data, melting furnace temperature field distribution data, material flow rate data, forming nozzle pressure data and cooling environment parameter data during the glass bead production process, and to perform preprocessing and time synchronization processing on the acquired multi-source heterogeneous sensor data through edge computing nodes.

[0053] The digital twin dynamic simulation module has a built-in digital twin model that integrates physical mechanism models and deep learning networks. It is configured to build a dynamic simulation system for the entire glass bead production process. The digital twin model maps the actual operating status of the glass bead production line online and predicts the viscosity of the molten glass, the size distribution of the beads, and the probability of surface defects during the production process.

[0054] The reinforcement learning collaborative control module is configured to adaptively generate a multi-actuator collaborative control strategy based on the reinforcement learning agent mechanism and the deviation between the prediction results output by the digital twin dynamic simulation module and the preset quality target. The multi-actuator collaborative control strategy is used to synchronously adjust the heating power zone of the melting furnace, the feed screw speed, the airflow forming pressure gradient, and the cooling air curtain velocity field.

[0055] The online quality assessment and feedback module is configured to evaluate the particle size consistency, sphericity, and optical uniformity of the glass beads produced online, and feed the assessment results back to the reinforcement learning collaborative control module to achieve closed-loop optimization of the control strategy.

[0056] The fault tracing and parameter compensation module is configured to monitor the operating conditions of the glass bead production line in real time. When abnormal operating conditions or quality drift are detected, the fault tracing process based on causal reasoning is automatically triggered to locate the disturbance source and generate a corresponding process parameter compensation scheme to ensure that the glass bead production line maintains stable operation with high yield without human intervention.

[0057] Example 2;

[0058] Based on Example 1, such as Figure 2 As shown, this embodiment provides a fully intelligent control method for a glass bead production line, including the following steps:

[0059] Step S1. Real-time acquisition of multi-source heterogeneous sensor data on glass raw material ratio, melting furnace temperature field distribution, material flow rate, forming nozzle pressure, and cooling environment parameters, and preprocessing and time synchronization through edge computing nodes;

[0060] Edge computing nodes preprocess multi-source heterogeneous data to achieve data time synchronization and rapid response, avoiding efficiency losses caused by parameter adjustment lag in traditional production.

[0061] Step S2. Construct a dynamic simulation system for the entire glass bead production process based on a digital twin model; wherein, the digital twin model integrates a physical mechanism model and a deep learning network to map the actual production line status online and predict the viscosity of the molten glass, the size distribution of the beads, and the probability of surface defects;

[0062] By integrating physical mechanisms and deep learning into a digital twin model, the viscosity of molten glass, the size distribution of beads, and the probability of surface defects can be predicted in advance, thus avoiding potential quality problems at the source.

[0063] By integrating physical mechanisms with a digital twin model of deep learning, the viscosity of molten glass, the size distribution of beads, and the probability of surface defects are predicted in real time. Based on reinforcement learning, the collaborative control of multiple process parameters is dynamically optimized, so that the particle size deviation of glass beads is controlled within ±0.5%, the sphericity is ≥99%, and the optical uniformity fluctuation is reduced by more than 60%, which is significantly better than the single-point automatic control method.

[0064] Step S3. Based on the reinforcement learning agent method combined with the digital twin model, the deviation between the output preset quality target is adaptively generated to generate a multi-actuator collaborative control strategy. The multi-actuator collaborative control strategy synchronously adjusts the heating power zone of the melting furnace, the feed screw speed, the airflow forming pressure gradient and the cooling air curtain velocity field.

[0065] Reinforcement learning agents adaptively generate control strategies, eliminating the need for repeated manual parameter adjustments, significantly reducing line changeover or process adjustment time, and making production line cycle time more stable.

[0066] Multiple actuators work together to adjust key parameters such as heating power zoning and molding pressure gradient, significantly improving particle size consistency and sphericity, and optical uniformity meets the requirements of high-precision applications.

[0067] It can accurately locate the root cause of abnormalities (such as nozzle blockage, temperature imbalance and cooling air deviation) within seconds and automatically generate collaborative control strategies to reduce unplanned downtime by more than 70% and ensure continuous high-efficiency production.

[0068] Step S4. Evaluate the particle size consistency, sphericity, and optical uniformity of the generated glass beads online, and feed the evaluation results back to the reinforcement learning agent to achieve closed-loop optimization of the control strategy;

[0069] By constructing an integrated intelligent control closed loop of "perception-simulation-decision-execution-feedback", it can continuously adapt to disturbances such as raw material fluctuations, equipment aging, or environmental changes without human intervention, increasing the production line yield to over 98% and reducing batch scrap due to process drift.

[0070] Online real-time evaluation of quality indicators and closed-loop feedback reduce human inspection errors, narrow the fluctuation range of core quality parameters, and significantly improve yield compared to traditional production methods.

[0071] Step S5. When abnormal operating conditions or quality drift are detected, the fault tracing module based on causal reasoning is automatically triggered to locate the disturbance source and generate a process parameter compensation scheme to ensure that the entire process maintains stable operation with high yield without human intervention.

[0072] The entire production line operation status can be traced, and the process parameters can be adjusted to meet the industrial mass production requirements of glass beads (such as micron-sized, high-performance hollow glass microspheres).

[0073] Example 3;

[0074] Based on Example 2, in step S1, the glass raw material ratio mainly includes basic raw materials such as quartz sand, soda ash, limestone and feldspar, as well as auxiliary raw materials such as clarifying agent and colorant. The accuracy of the ratio directly determines the chemical composition, strength and transparency of the glass. The real-time acquisition in this step focuses on the quantitative monitoring of raw material components.

[0075] The melting furnace is the core equipment in glass production. Raw materials need to be melted into a uniform glass liquid at a high temperature of 1500-1600℃. The uniformity of the temperature field directly affects the clarification effect of the glass liquid (avoiding bubble residue) and the uniformity of components (preventing stone defects). The core of this process is monitoring the temperature distribution of the entire furnace.

[0076] The molten glass, after being melted and clarified, needs to be transported to the forming stage through a feeding system. The stability of the feeding flow rate directly determines the uniformity of the thickness of the formed products (e.g., flat glass) and the consistency of weight (e.g., bottle and jar glass), avoiding forming defects such as material shortage and overflow. The key point of data collection in this stage is the real-time monitoring of the glass molten glass supply flow rate.

[0077] In the forming process of float glass, insulated glass, or special glass (such as optical glass), molten glass enters the forming mold or forming area through forming nozzles (or outlets). The stability of the nozzle pressure directly affects the outflow speed and forming shape of the molten glass. It is a key parameter to avoid ripples, deformation, and dimensional deviations in the formed products. The core of this process is to collect data on the nozzle inlet and outlet pressures and pressure fluctuations.

[0078] After molding, glass products need to undergo orderly cooling (annealing) to avoid internal stress caused by uneven cooling rate, which can lead to product cracking and reduced strength. The temperature, humidity and airflow speed of the cooling environment are the core factors affecting the cooling effect. Different types of glass products (such as ultra-thin glass and tempered glass) have significantly different requirements for cooling parameters. The focus of this step is to collect data on the real-time monitoring of multiple environmental parameters in the cooling area.

[0079] The heterogeneity is manifested in the following ways: First, the data sources are heterogeneous, with data coming from various types of sensors, including weighing, temperature, pressure, flow, spectroscopy, and infrared imaging, covering both contact and non-contact acquisition methods. Second, the formats are heterogeneous, including numerical data (e.g., flow rate, pressure), image data (e.g., infrared thermal imaging spectra), and spectral data (e.g., near-infrared spectral curves). Third, the rates are heterogeneous, with acquisition frequencies ranging from 1Hz (e.g., mixing uniformity) to 100Hz (e.g., nozzle pressure), resulting in data generation rates that differ by up to 100 times. Fourth, the accuracy is heterogeneous, with different accuracy requirements for different parameters (e.g., feed rate accuracy ±0.1%, humidity accuracy ±0.5%RH).

[0080] For the preprocessing of edge computing nodes, a time offset propagation correction method using graph neural networks is adopted. Since clock offsets have spatial correlation in large-scale heterogeneous edge networks, offset propagation is performed by modeling using graph neural networks.

[0081] ;

[0082] in, For nodes At any moment The Time offset estimation after layer iteration (i.e., the difference between the local clock and the global reference clock); For nodes At any moment The Layer offset estimation, used as data input; For nodes The set of neighbors (e.g., other edge nodes within the communication range that can exchange information); For neighboring nodes The Layer offset estimation; For nodes and nodes The edge weights between them reflect the communication quality (e.g., RTT delay, signal-to-noise ratio, and packet loss rate). For all neighboring nodes Send to node The message set, including neighboring nodes offset value and node Combine the link quality; This is a preprocessing function for neighboring nodes (e.g., a weighted summation of neighboring states). This is a node update function that receives its own state and messages from its neighbors, outputs a new offset estimate, representing the combined node... and nodes To make new decisions.

[0083] This formula allows each node to broadcast its current offset estimate. ;

[0084] Receiver node Received the information, and based on the link quality Perform weighted processing;

[0085] Use node update function Integrate its own historical estimates with neighbor information;

[0086] Output new offset estimate .

[0087] The first Each edge node at time... After the first layer or first Time offset estimate after the next iteration It is based on the current offset estimate Neighbor nodes Offset information (via function) (Weighted aggregation), combined with a nonlinear function To integrate and update.

[0088] For time synchronization of edge computing nodes, a clock correction method based on preprocessing feedback is used, specifically by introducing preprocessing residual feedback to correct the clock:

[0089] ;

[0090] in, For nodes At any moment Updated clock offset estimate; For nodes At any moment Old clock offset estimate before update; The learning rate (step size) is used to control the update magnitude. For nodes The output after preprocessing (e.g., feature vectors, compressed data); For neighboring nodes The preprocessed output set; For clock offset The gradient represents the effect of changing the offset on the consistency loss. The consistency loss function measures... The difference between its output and that of its neighbors.

[0091] Model parameters It is an indicator that needs to be calibrated (e.g., multi-source heterogeneous sensing data). It compares the degree of inconsistency between the current parameters and the states of its neighbors (e.g., the difference between a temperature reading and other surrounding temperature readings); gradient. This involves adjusting the parameters to reduce inconsistencies (e.g., if the temperature reading is too high, it needs to be lowered).

[0092] The essence of the formula is to subtract the correction amount that points to consistency from the current parameters to obtain more stable new parameters, and then correct its own state through feedback from the surrounding environment to achieve coordination.

[0093] Example 4;

[0094] Based on Example 2, in step S2, for the digital twin model that integrates the physical mechanism model and the deep learning network, a differential-algebraic-neural hybrid model is adopted, so that the deep learning network directly modulates the evolution rate of the physical mechanism model:

[0095] ;

[0096] in, This refers to the system state vector, such as the temperature field, viscosity field, and velocity field. To control input vectors, such as heating power, rotation speed, and cooling fan speed; Update the node function; For temperature gradient fields, such as spatial derivatives, which reflect the distribution of thermal stress, they can be scalars or vectors. These are neural network weights, which can be trained and optimized using data.

[0097] The physical mechanism function is a differential equation model built on physical quantities (such as mass, momentum, and energy). For neural network residuals, which are deep learning modules used to compensate for unmodeled complex nonlinear behaviors; The rate of change of state represents the direction of the system's dynamic evolution.

[0098] It indicates the trend of the system over time, such as how the temperature of the molten glass rises over time and how the flow rate changes;

[0099] It is a physical mechanism model based on physical quantities, including: the law of conservation of energy (such as the law of conservation of mass, momentum and energy), fluid dynamics (i.e., calculation using the Navier-Stokes equations), and viscosity model (VFT equations (implied in the physical mechanism function)). The model ensures physical consistency and interpretability, and is used to describe the force balance during temperature diffusion, molten glass flow, and bead formation processes within the furnace.

[0100] Neural network residual It is a learnable vector field designed to supplement aspects that physical mechanism models cannot accurately capture, namely batch variations in materials (e.g., the influence of impurities), performance drift caused by equipment aging, non-ideal boundary conditions (e.g., uneven local heat dissipation), and multiphase coupling effects (e.g., microscopic phenomena such as bubbles and crystallization). The input includes a temperature gradient field. It emphasizes the importance of spatial non-uniformity. For example, high-temperature gradient regions are more prone to defects, and neural network residuals can be directly loaded onto the evolution of physical mechanism models to achieve online correction.

[0101] This indicates that under new operating conditions (e.g., material change), as long as the physical structure remains unchanged, the physical mechanism function... It is still effective, while the neural network residual It can quickly adapt to new deviation patterns.

[0102] Example 5;

[0103] Based on Example 2, in step S3, the deviation between preset quality targets is represented by deviation energy to indicate the degree to which the system as a whole deviates from the target:

[0104] ;

[0105] in, As a scalar error metric, it represents the degree to which the system as a whole deviates from the target. It is used to quantify the extent to which the system deviates from the preset quality target at a certain moment. For a moment The quality deviation vector represents the difference between the actual / predicted value and the target of each quality indicator, reflecting a specific defect direction (e.g., too thick, too cold, uneven). Mass deviation vector The transpose of , This is the weight matrix. The square of the Euclidean norm;

[0106] It is the quality deviation vector With quality deviation vector transpose vector In the weight matrix The inner product is used to measure the energy of multidimensional deviation, representing the degree to which the system as a whole deviates from the preset quality target at the current moment.

[0107] It represents the weighted sum of squares of the deviations of each quality indicator, integrating the errors of all dimensions into a global indicator, realizing the transformation from multiple objectives to a single objective, which facilitates optimization; For nodes Edge weights; Indicates a node The mass deviation vector is squared to emphasize large errors (non-linear penalty), highlight the impact of large errors, and avoid positive and negative deviations canceling each other out, which is equivalent to energy loss. This represents the total number of quality indicators.

[0108] This formula is a weighted energy function for multidimensional quality deviations. It transforms complex process quality requirements into a simple, optimizable, and interpretable scalar objective, which is key to realizing intelligent closed-loop control. It enables intelligent manufacturing closed-loop control from finding deviations to learning to correct them and then to automatic optimization.

[0109] For multi-actuator coordinated control strategies, adjustments are made based on action increments:

[0110] ;

[0111] in, For at any time The increment of control actions; This is a biased attention mechanism used to determine which actuators need attention; For element-wise multiplication; As a basic response item, it can provide rapid, localized initial correction; This is a dynamic coupling compensation term used to handle complex interactions between actuators. This is a coupling matrix, representing the intensity of the linkage effect on the actuator when it is adjusted. For example, at high temperatures, the interaction between the heating and cooling zones is stronger. Automatically enhance related items, The deviation drive signal nonlinearly maps the quality deviation to the actuator coordinated drive signal.

[0112] Its function is to capture the nonlinear, time-varying coupling relationships between actuators; for example, reducing the heating power in the middle section not only affects the thickness but also changes the melt flow, and simultaneously adjusts the screw speed and gas flow pressure. These complex relationships are caused by... and Collaborative modeling.

[0113] time Control action increment Ultimately, selective activation is achieved through element-wise multiplication (Hadamard product), ensuring that control commands are both effective and safe.

[0114] When a certain indicator deviates significantly (e.g., thickness exceeds tolerance), the attention weight of the corresponding actuator is close to 1, and it is prioritized for scheduling.

[0115] When the deviation of a certain indicator is small (e.g., excellent thickness), the attention weight of the corresponding actuator is close to 0, reducing disturbances and avoiding over-adjustment.

[0116] Example 6;

[0117] Based on Example 2, in step S4, the specific steps for online evaluation are as follows:

[0118] Regarding particle size consistency, a laser diffraction method combined with dynamic image analysis is used for online detection. A laser particle size analyzer emits a laser to illuminate the glass bead flow, and the particle size distribution is calculated based on the angle and intensity of the scattered light, quickly outputting the particle size distribution (e.g., D10, D50, and D90). Simultaneously, a high-speed camera dynamic image analysis system is used to help capture particle morphology and eliminate interference from non-spherical impurities on the particle size data. The evaluation focuses on the coefficient of variation and distribution span; a smaller coefficient of variation and a narrower distribution span indicate better particle size consistency. For example, glass beads used for road markings need to ensure that the particle size is concentrated within a specific range to guarantee uniform reflectivity.

[0119] Regarding sphericity, an online evaluation is performed using a high-precision vision inspection system. This system consists of a high-resolution camera, a ring-shaped backlight, and a rotating conveyor. As the glass bead rotates with the conveyor, the camera captures its contour images from multiple angles. The software extracts the contour using an edge detection algorithm and then calculates the sphericity error using the least squares circle method, comparing the actual contour with the ideal circle. Generally, a sphericity ≥ 0.85 is considered acceptable, but optical-grade glass beads require even higher standards; in some scenarios, a sphericity of over 95% is necessary to avoid affecting the light refraction path.

[0120] Regarding optical uniformity, an evaluation is achieved by combining refractive index detection and internal defect imaging techniques. On one hand, an online Abbe refractometer is used to measure the refractive index of the glass bead in real time. If the refractive index fluctuation exceeds a set threshold, it indicates poor optical uniformity. On the other hand, a microscopic imaging and transmittance detection module captures internal bubbles, cracks, and impurities within the glass bead. These defects cause abnormal light scattering and refraction. The system further quantifies the optical uniformity level by analyzing the uniformity of transmittance and the proportion of defect area. For example, glass beads used in optical devices require a stable refractive index in the range of 1.5-1.65 and no obvious internal defects.

[0121] Data standardization involves converting the raw test data for the three indicators into a standardized score of 0-1. For example, sphericity is set to 1 point, and the score decreases linearly for every certain increase in sphericity error (in micrometers). Abnormal data, such as invalid images caused by equipment vibration or extreme particle size data due to sudden impurities in the raw materials, are removed to ensure the reliability of the feedback data.

[0122] The reinforcement learning agent aims to improve the quality stability of glass beads and consists of three parts: state awareness, reward function, and policy network. The state awareness module receives preprocessed quality data and collects current production parameters such as raw material ratio, melting temperature, and cooling rate. The reward function is set based on the quality score; if all three indicators meet the standards, a positive reward is given, and the greater the deviation, the stronger the penalty. The policy network, built based on a deep reinforcement learning algorithm, is responsible for outputting parameter adjustment instructions.

[0123] Dynamic optimization of the control strategy involves the reinforcement learning agent continuously learning and optimizing based on feedback quality data. For example, when a decrease in particle size consistency is detected, the agent analyzes the corresponding operating conditions in historical data to determine whether it is caused by uneven raw material mixing or fluctuations in outlet pressure. Subsequently, it outputs instructions to adjust the raw material delivery rate and fine-tune the outlet pressure. If the sphericity does not meet the standard, it often corresponds to excessively fast cooling or deviations in mold parameters. In this case, the agent will optimize the temperature gradient of the cooling channel and adjust the rotation speed parameters of the molding die.

[0124] The complete process of closed-loop optimization of the control strategy is as follows:

[0125] Step a: The production line continuously produces glass beads, and the online detection module simultaneously collects data and completes the evaluation of three indicators;

[0126] Step b: The standardized evaluation results are fed back to the reinforcement learning agent in real time;

[0127] Step c: The agent compares the evaluation results with the preset quality standards, and generates the optimal adjustment strategy based on the current production parameters through an algorithm;

[0128] Step d: The actuator receives instructions and adjusts key parameters in the melting, forming, and cooling processes;

[0129] Step e: The adjusted production parameters are applied to the production of the next batch of glass beads, and the test data of the new batch of products are fed back to the agent, forming a cycle.

[0130] Closed-loop optimization of control strategies can significantly improve the stability of glass bead production. For example, after applying a similar reinforcement learning optimization scheme, a company increased its production efficiency by 15%, reduced energy consumption, and significantly decreased the product quality defect rate. This is especially suitable for scenarios with stringent requirements for glass bead quality, such as optical devices and high-precision reflective materials.

[0131] Example 7;

[0132] Based on Example 2, in step S5, the triggering condition for abnormal operating conditions or mass drift is:

[0133] Abnormal operating conditions are determined by automatically triggering a fault tracing module based on causal reasoning when online evaluation indicators (particle size consistency, sphericity, optical uniformity) exceed preset thresholds, or when the collected sensor data (temperature field distribution, nozzle pressure) shows abrupt changes or exceeds limits.

[0134] Quality drift determination: When the evaluation indicators are not exceeded, but multiple batches of data show a significant trend of deviating from the target value, the system determines it as quality drift, initiates the intervention mechanism in advance, and automatically triggers the fault tracing module based on causal reasoning.

[0135] The fault tracing logic based on causal reasoning is as follows:

[0136] The causal reasoning-based fault tracing model integrates physical mechanism rules with the correlation patterns of production data, avoiding the misjudgment problems of traditional correlation analysis. The model maps abnormal phenomena with potential disturbance sources. For example, particle size fluctuations are associated with uneven temperature fields in the melting furnace and unstable screw speeds in the feeder; surface defects are associated with turbulent flow fields in the cooling air curtain. Through the simulation and backtracking function of the digital twin model, the evolution of process parameters before the anomaly occurs can be reproduced, accurately locating the specific position and degree of influence of the disturbance source.

[0137] The generation and execution of the process parameter compensation scheme are as follows:

[0138] The compensation scheme is based on the source tracing results, calls the historical optimization strategy library of the reinforcement learning agent, and generates targeted parameter adjustment instructions.

[0139] Adjustment commands are applied synchronously to multiple actuators, such as adjusting heating power in zones when the temperature field is uneven, and correcting the velocity field gradient when the air curtain is turbulent.

[0140] After the compensation plan is implemented, the system collects production data and quality indicators after parameter adjustment in real time to verify the compensation effect, forming a complete closed loop of fault triggering, source tracing and location, compensation execution and effect verification.

[0141] The goal of unmanned, stable operation is to ensure that the entire process requires no human intervention, relying on pre-defined algorithms to automatically complete fault diagnosis and parameter correction, thereby guaranteeing high-yield and stable output from the production line under complex operating conditions. Simultaneously, historical data on fault tracing and compensation is fed back to the digital twin model and reinforcement learning agent, continuously optimizing the model's predictive accuracy and the adaptability of the control strategy.

[0142] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope described in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fully intelligent control method for a glass bead production line, characterized in that, Includes the following steps: Multi-source heterogeneous sensor data is collected and preprocessed and synchronized with time through edge computing nodes; the multi-source heterogeneous sensor data includes glass raw material ratio, furnace temperature field distribution, feed flow rate, forming nozzle pressure, and cooling environment parameters. A dynamic simulation system for the entire glass bead production process is constructed based on a digital twin model. The digital twin model integrates a physical mechanism model and a deep learning network to map the actual production line status online and predict the viscosity of the molten glass, the size distribution of the beads, and the probability of surface defects. The digital twin model integrates a physical mechanism model with a deep learning network, employing a differential-algebraic-neural hybrid model. This allows the deep learning network to directly modulate the evolution rate of the physical mechanism model, as detailed below: ； in, The system state vector includes: temperature field, viscosity field, and velocity field; The control input vector includes: heating power, rotation speed, and cooling fan speed; Update the node function; It represents the temperature gradient field, its spatial derivative, and can reflect the distribution of thermal stress; it can be a scalar or a vector. These are neural network weights, which can be trained and optimized using data. It is a physical mechanism function, which is a differential equation model based on the physical quantities of mass, momentum and energy; For neural network residuals, which are deep learning modules used to compensate for unmodeled complex nonlinear behaviors; The rate of change of state represents the direction of the system's dynamic evolution, indicating the trend of the system's change over time, and reflecting how the temperature of the molten glass increases and how the flow velocity changes over time. It is a physical mechanism model based on physical quantities, including: the law of conservation of energy, fluid dynamics and viscosity model, which ensures that the model has physical consistency and interpretability, and is used to describe the force balance in the process of temperature diffusion, glass melt flow and bead formation in the furnace. Neural network residual It is a learnable vector field designed to supplement aspects that physical mechanism models cannot accurately capture, such as material batch variations, performance drift due to equipment aging, non-ideal boundary conditions, and multiphase coupling effects. The input includes a temperature gradient field. This emphasizes the importance of spatial non-uniformity; This indicates that under the new operating conditions, as long as the physical structure remains unchanged, the physical mechanism function... It is still effective, while the neural network residual It can quickly adapt to new deviation patterns; Based on the deviation between the output of the preset quality target by the reinforcement learning agent method and the digital twin model, an adaptive multi-actuator collaborative control strategy is generated. The multi-actuator collaborative control strategy is used to synchronously adjust the heating power zone of the melting furnace, the feed screw speed, the airflow forming pressure gradient and the cooling air curtain velocity field. The multi-actuator collaborative control strategy is adjusted based on action increments. ； in, For at any time The increment of control actions; This is a biased attention mechanism used to determine which actuators need attention; For element-wise multiplication; As a basic response item, it can provide rapid, localized initial correction; This is a dynamic coupling compensation term used to handle complex interactions between actuators. This is a coupling matrix, representing the intensity of the linkage effect on the actuator when it is adjusted. It reflects the strong interaction between the heating and cooling zones at high temperatures. This indicates an item that is automatically enhanced. The deviation drive signal maps the quality deviation nonlinearly to the actuator coordinated drive signal; Its function is to capture the nonlinear, time-varying coupling relationship between actuators; by reducing the heating power in the middle, it not only affects the thickness but also changes the melt flow, simultaneously adjusting the screw speed and airflow pressure, and this relationship is determined by... and Collaborative modeling; time Control action increment Selective activation is achieved through element-wise multiplication, ensuring that control commands are both effective and safe. When a certain indicator deviates significantly, the attention weight of the corresponding executor is 1, and it is prioritized for scheduling. When the deviation of a certain indicator is small, the attention weight of the actuator is 0 to reduce disturbances; The particle size uniformity, sphericity, and optical uniformity of the generated glass beads are evaluated online, and the evaluated indicators are fed back to the reinforcement learning agent to achieve closed-loop optimization of the control strategy. When abnormal operating conditions or quality drift are detected, the fault tracing and parameter compensation module based on causal reasoning is automatically triggered to locate the disturbance source and generate a process parameter compensation scheme to ensure that the entire process maintains stable operation with high yield without human intervention.

2. The intelligent control method for the entire process of a glass bead production line according to claim 1, characterized in that, For the preprocessing of the edge computing nodes, a time offset propagation correction method using graph neural networks is adopted. Since clock offsets have spatial correlations in large-scale heterogeneous edge networks, offset propagation is performed through graph neural network modeling, specifically as follows: ； in, For nodes At any moment The Time offset estimation after layer iteration; For nodes At any moment The Layer offset estimation, used as data input; For nodes The set of neighbors; Neighboring nodes The Layer offset estimation; For nodes and nodes The edge weights between nodes reflect the communication quality; For all neighboring nodes Send to node The message set will include neighboring nodes. offset value and node Combine the link quality; This is a preprocessing function for neighboring nodes; This is a node update function that receives its own state and messages from its neighbors, outputs a new offset estimate, representing the combined node... and nodes To make new decisions; The above formula is based on each node broadcasting its own current offset estimate. ; Receiver node Received the information, and based on the link quality Perform weighted processing; use node update functions. It integrates its own historical estimates with neighbor information to output a new offset estimate. .

3. The intelligent control method for the entire process of a glass bead production line according to claim 1, characterized in that, For time synchronization of edge computing nodes, a clock correction method based on preprocessing feedback is used. This method corrects the clock by introducing preprocessing residual feedback. The approach is as follows: ； in, For nodes At any moment Updated clock offset estimate; For nodes At any moment Old clock offset estimate before update; The learning rate or step size is used to control the update magnitude. For nodes The output after preprocessing; Neighboring nodes The preprocessed output set; For clock offset The gradient represents the effect of changing the offset on the consistency loss; it involves adjusting the parameters to reduce parameter inconsistency. The consistency loss function measures... The difference between the output of the current parameter and that of its neighbor is the degree of inconsistency between the current parameter and the state of the neighbor. The essence of the formula is to subtract the correction amount that points to consistency from the current parameters to obtain more stable new parameters, and then correct its own state through feedback from the surrounding environment to achieve coordination.

4. The intelligent control method for the entire process of a glass bead production line according to claim 1, characterized in that, The deviation between the preset quality targets is represented by deviation energy, indicating the degree of overall system deviation from the target. Using a weighted energy function of multidimensional quality deviations, complex process quality requirements are transformed into scalar targets. Intelligent closed-loop control is then executed, enabling intelligent manufacturing closed-loop control that identifies deviations, learns to correct them, and then automatically optimizes. Specifically: ； in, As a scalar error metric, it represents the degree to which the system as a whole deviates from the target. It is used to quantify the extent to which the system deviates from the preset quality target at a certain moment. For a moment The quality deviation vector represents the difference between the actual or predicted value of each quality indicator and the target, reflecting a specific defect direction, including: too thick, too cold, or uneven. Mass deviation vector The transpose of , This is the weight matrix. The square of the Euclidean norm; It is the quality deviation vector With quality deviation vector transpose vector In the weight matrix The inner product is used to measure the energy of multidimensional deviation, representing the degree to which the system as a whole deviates from the preset quality target at the current moment; It represents the weighted sum of squares of the deviations of each quality indicator, integrating the errors of all dimensions into a global indicator, realizing the transformation from multiple objectives to a single objective, which facilitates optimization; For nodes Edge weights; Indicates a node The mass deviation vector is squared to emphasize the impact of large errors and prevent positive and negative deviations from canceling each other out, which is equivalent to energy loss. This represents the total number of quality indicators.

5. The intelligent control method for the entire process of a glass bead production line according to claim 1, characterized in that, The complete operation flow of the closed-loop optimization of the control strategy is as follows: The production line continuously produces glass beads, and the online quality assessment and feedback module simultaneously collects data and completes the assessment of particle size consistency, sphericity, and optical uniformity. The evaluated metrics are fed back to the reinforcement learning agent in real time; The reinforcement learning agent compares and evaluates the indicators with preset quality standards, and generates the optimal adjustment strategy based on the current production parameters. The actuator receives instructions and adjusts key parameters in the melting, forming, and cooling processes. The adjusted production parameters are applied to the production of the next batch of glass beads, and the test data of the new batch of products is fed back to the agent, forming a cycle.

6. The intelligent control method for the entire process of a glass bead production line according to claim 1, characterized in that, The generation and execution of the process parameter compensation scheme are as follows: The compensation scheme is based on the source tracing results, calls the historical optimization strategy library of the reinforcement learning agent, and generates process parameter adjustment instructions; Process parameter adjustment commands are applied synchronously to multiple actuators, including: adjusting heating power in zones when the temperature field is uneven, and correcting the flow velocity gradient when the air curtain is turbulent. After the compensation plan is implemented, data on the adjusted process parameters and quality indicators are collected in real time to verify the compensation effect, forming a complete closed loop of fault triggering, source tracing and location, compensation execution and effect verification.

7. A fully intelligent control system for a glass bead production line, used to execute the fully intelligent control method for a glass bead production line according to any one of claims 1-6, characterized in that, include: The multi-source heterogeneous data acquisition and preprocessing module is configured to acquire multi-source heterogeneous sensor data in real time during the glass bead production process, and to perform preprocessing and time synchronization processing on the acquired multi-source heterogeneous sensor data through edge computing nodes. The digital twin dynamic simulation module is connected to the multi-source heterogeneous data acquisition and preprocessing module. It has a built-in digital twin model that integrates physical mechanism model and deep learning network, and is configured to build a dynamic simulation system for the entire glass bead production process. The reinforcement learning collaborative control module is connected to the digital twin dynamic simulation module and is configured to adaptively generate a multi-actuator collaborative control strategy based on the reinforcement learning agent mechanism and the deviation between the prediction results output by the digital twin dynamic simulation module and the preset quality target. The online quality assessment and feedback module is connected to the reinforcement learning collaborative control module and is configured to assess the indicators obtained from production online and feed the assessment results back to the reinforcement learning collaborative control module. The fault tracing and parameter compensation module is connected to the online quality assessment and feedback module and configured to monitor the operating conditions of the glass bead production line in real time.

Images