A production line intelligent interaction method and system based on digital twinning and internet of things fusion

Abstract

The application discloses a production line intelligent interaction method and system based on digital twinning and Internet of Things fusion, and the method comprises the following steps: collecting running state and process beat data in a production process to construct a digital twinning framework and generate a virtual-real mapping digital twinning model; based on the digital twinning model, a reinforcement learning algorithm is used to train a global optimization strategy for the production process, and a global optimization strategy is output; according to the global optimization strategy, the temperature field distribution, pressure change curve and beat connection relationship of each process in the production process are simulated and deduced to generate an optimized process parameter combination; the execution effect deviation of the optimized process parameter combination is detected, and early warning diagnosis information containing fault types is generated; and according to the early warning diagnosis information, a fault solution is generated. By using the embodiment of the application, the process parameter self-optimization and fault active early warning collaborative control of the laminated glass production line can be realized, and the product quality consistency and production line operation reliability are improved.

Description

Technical Field

[0001] This invention belongs to the field of laminated glass production technology, and in particular to an intelligent interactive method and system for production lines based on the integration of digital twins and the Internet of Things. Background Technology

[0002] Laminated glass production involves multiple continuous processes, including raw glass cleaning, lamination, pre-pressing, autoclave forming, and cutting. The temperature, pressure, and cycle time parameters of each process are highly coupled, and the precision of process control directly affects the quality of the finished product. Current production management generally combines manual experience-based adjustments with single-machine automation. Process parameter settings rely on repeated trial and error by engineers, and it's difficult to unify operating standards between different shifts, leading to large fluctuations in product quality and unstable yields. As production lines expand, equipment aging and multi-process coordination issues are further exacerbated, making it difficult for traditional offline quality inspection and periodic maintenance to detect process deviations and equipment degradation trends in a timely manner. In recent years, digital twin technology has been gradually applied in manufacturing, but most solutions only achieve status monitoring and 3D visualization, lacking deep integration with optimization algorithms such as reinforcement learning. IoT sensing data is often only simply aggregated and displayed, failing to drive autonomous optimization and proactive early warning of the production line. More importantly, existing methods generally separate process optimization from equipment health management, ignoring the potential impact of process parameter adjustments on equipment load. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent interactive method and system for production lines based on the integration of digital twins and the Internet of Things, in order to overcome the shortcomings of the existing technology, and to realize the self-optimization of process parameters and collaborative control of proactive fault warning in laminated glass production lines, thereby improving product quality consistency and production line operation reliability.

[0004] One embodiment of this application provides an intelligent interaction method for production lines based on the integration of digital twins and the Internet of Things, the method comprising: By deploying multi-source sensors at each process node of the laminated glass production line, the operating status and process cycle data of the production process are collected in real time to build a digital twin framework that includes a physical layer, a virtual layer, a data layer and an interactive control service layer. The equipment geometry, operating status and process flow of the physical production line are mapped and updated in real time in the virtual space to generate a digital twin model that maps the virtual and real. Based on the digital twin model, a reinforcement learning algorithm is used to train a global optimization strategy for the production process. The digital twin model is used as an environment interaction platform to dynamically simulate the production line operation status and output the global optimization strategy after reinforcement learning iteration. Based on the global optimization strategy, the temperature field distribution, pressure change curve and cycle connection relationship of each process in the production process are simulated and deduced. With product quality consistency as a constraint, the optimal combination of heating temperature, pressure parameters and rolling time is searched to generate the optimized process parameter combination. The optimized process parameter combination is subjected to performance deviation detection. At the same time, the digital twin model is combined with the pre-built equipment health assessment model to perform trend analysis on the operating status data of key equipment components, predict potential equipment failure hazards, and generate early warning and diagnostic information containing failure types. Based on the early warning and diagnostic information, a digital twin model is used for fault tracing and spatial positioning. Combined with a pre-set expert knowledge base, a repair plan corresponding to the fault type is matched to generate a fault solution that includes fault location identification and maintenance suggestions.

[0005] Another embodiment of this application provides an intelligent interactive system for a production line based on the integration of digital twins and the Internet of Things, the system comprising: The data acquisition module is used to collect real-time data on the operating status and cycle time of the production process through multi-source sensors deployed at each process node of the laminated glass production line. This is to build a digital twin framework that includes a physical layer, a virtual layer, a data layer, and an interactive control service layer. The module also maps and updates the equipment geometry, operating status, and process flow of the physical production line in the virtual space in real time, generating a digital twin model that maps the physical and virtual components. The simulation module is used to train a global optimization strategy for the production process based on the digital twin model using a reinforcement learning algorithm, dynamically simulate the production line operation status using the digital twin model as an environmental interaction platform, and output the global optimization strategy after reinforcement learning iteration optimization. The search module is used to simulate and deduce the temperature field distribution, pressure change curve and cycle connection relationship of each process in the production process according to the global optimization strategy. With product quality consistency as a constraint, it searches for the optimal combination of heating temperature, pressure parameters and rolling time to generate the optimized process parameter combination. The prediction module is used to detect the deviation of the execution effect of the optimized process parameter combination. At the same time, it uses the digital twin model in combination with the pre-built equipment health assessment model to perform trend analysis on the operating status data of key equipment components, predict potential equipment failure hazards, and generate early warning and diagnostic information containing failure types. The generation module is used to perform fault backtracking and spatial positioning using a digital twin model based on the early warning and diagnostic information, and to combine a pre-set expert knowledge base to match the repair scheme corresponding to the fault type, thereby generating a fault solution that includes fault location identification and maintenance suggestions.

[0006] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0007] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0008] Compared with existing technologies, the present invention provides an intelligent interaction method for production lines based on the integration of digital twins and the Internet of Things, which can realize the self-optimization of process parameters and collaborative control of proactive fault warning in laminated glass production lines, thereby improving product quality consistency and production line operation reliability. Attached Figure Description

[0009] Figure 1 A hardware structure block diagram of a computer terminal for an intelligent interaction method for a production line based on the integration of digital twins and the Internet of Things, provided for an embodiment of the present invention; Figure 2 A flowchart illustrating an intelligent interaction method for a production line based on the integration of digital twins and the Internet of Things, provided as an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent interactive system for a production line based on the integration of digital twins and the Internet of Things, provided as an embodiment of the present invention. Detailed Implementation

[0010] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] The present invention first provides an intelligent interaction method for production lines based on the integration of digital twins and the Internet of Things. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0012] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for an intelligent interaction method for a production line based on the integration of digital twins and the Internet of Things, provided as an embodiment of the present invention. (See diagram below.) Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0013] See Figure 2 The embodiments of the present invention provide an intelligent interaction method for production lines based on the integration of digital twins and the Internet of Things, which may include the following steps: S201 uses multi-source sensors deployed at each process node of the laminated glass production line to collect real-time data on the operating status and process cycle time during the production process. This is used to construct a digital twin framework that includes a physical layer, a virtual layer, a data layer, and an interactive control service layer. The equipment geometry, operating status, and process flow of the physical production line are mapped and updated synchronously in the virtual space in real time to generate a digital twin model that maps the virtual and real worlds. Specifically, multiple types of sensors deployed at various process nodes of the production line can be used to collect equipment operating parameters, product position signals, and process completion status signals in real time, generating multi-source sensor raw datasets. The core of this step is to build a distributed sensing network across all process nodes of the laminated glass physical production line. This network relies on multiple types of sensors to comprehensively collect three core data categories: equipment operation, product flow, and process progress. This provides a realistic and continuous underlying data source for subsequent digital twin modeling. The specific implementation method is as follows: The laminated glass production line comprises six core process nodes: raw glass cutting, edge grinding and cleaning, tempering heating, lamination and pre-pressing, autoclave curing, and finished product inspection. Each process node corresponds to different process parameters and equipment characteristics. Therefore, a differentiated sensor deployment scheme is adopted to achieve full-dimensional data coverage. All sensors adopt industrial-grade protection, are suitable for the complex working environment of high temperature in the tempering furnace and high humidity in the autoclave, and have an average mean time between failures (MTBF) of over 50,000 hours, ensuring the stability of long-term continuous data collection. The sensors are divided into three categories according to their functions. The first category is for equipment operating parameter acquisition, including temperature sensors, pressure sensors, motor current sensors, and roller speed sensors. These are deployed in the heating section of the tempering furnace, the autoclave cavity, and the drive motors and transmission roller groups of each process, respectively, to collect core equipment operating parameters. The temperature sampling frequency for the tempering furnace is 1 time per second, the pressure sampling frequency for the autoclave is 2 times per second, and the motor current sampling frequency is 10 times per second, ensuring that dynamic fluctuations in equipment operation can be captured. The second category is for product position signal acquisition. Through-beam photoelectric sensors are deployed at the entrance and exit of each process. When a glass workpiece passes by, a level change is triggered, accurately recording the moment the workpiece enters or exits the process. The sampling response time is less than 1 millisecond, ensuring the time accuracy of position detection. The third category is for process completion status signal acquisition. These sensors interface with the control system output ports of each process's equipment, collecting process completion trigger signals, equipment fault alarm signals, and manual pause signals to determine the processing progress of a single process and the overall operating status.

[0014] All sensors are connected to the edge acquisition gateway via an industrial bus. The collected data covers three main categories: equipment operating parameters for the entire process, product position trigger signals, and process completion status signals. Equipment operating parameters include core indicators such as the temperature of each section of the tempering furnace, the internal pressure of the autoclave, the operating current of each motor, and the speed of the drive rollers. Product position signals include the entry and exit times of each workpiece. Process completion status signals include single workpiece processing completion indicators, equipment standby indicators, and abnormal shutdown indicators. All collected raw data is arranged sequentially by collection timestamp, without any filtering or correction, fully preserving the original signal characteristics to form a multi-source sensor raw dataset. In the example, a single production line can collect approximately 120,000 raw sensor data points per hour, covering 32 collection points across six processes, fully supporting subsequent data processing and model building.

[0015] The edge-side preprocessing of the original dataset from multi-source sensing is performed, the filtering algorithm is used to remove signal noise, and the signals from different sensing devices are uniformly converted into standard format data frames through protocol parsing to generate a standardized sensing data frame sequence. The core of this step is to perform noise reduction, cleaning, and format standardization of raw data at the edge of the production line. This eliminates signal noise caused by electromagnetic interference and mechanical vibration in the industrial environment, resolves the incompatibility issues of communication protocols between different brands of sensors, and outputs standardized data with a unified format and high reliability. This provides a high-quality data foundation for subsequent feature extraction. The specific implementation method is as follows: Edge-side preprocessing is deployed in the edge computing gateway on the production line. This eliminates the need to upload all raw data to the cloud; data cleaning and format conversion can be completed locally, reducing data transmission bandwidth consumption and improving real-time processing. Preprocessing consists of two core stages: noise filtering and protocol parsing. The noise filtering stage uses a first-order low-pass filter algorithm to remove high-frequency noise from the analog signal. The core expression of the algorithm is y_n=α*x_n+(1-α)*y_n-1, where y_n is the filtered output value at the current moment, x_n is the original sampled value at the current moment, y_n-1 is the filtered output value at the previous moment, and α is the filter coefficient, ranging from 0 to 1. A smaller coefficient results in a stronger filtering effect but also a more significant signal lag. For slow variables such as tempering furnace temperature and autoclave pressure, where changes are gradual, the filter coefficient α is set to 0.2, effectively filtering out spike noise caused by electromagnetic interference while preserving the true trend of signal changes. For fast variables such as motor current and speed, the filter coefficient α is set to 0.6, ensuring signal response speed while filtering out noise. For digital switch signals such as photoelectric switches, anti-jitter filtering is used, with a 20-millisecond anti-jitter duration. Only when the signal level changes continuously for more than the anti-jitter duration is it considered a valid trigger, thus avoiding false triggering signals caused by mechanical vibration.

[0016] The protocol parsing stage unifies the communication protocols of different sensing devices. Different types of sensors in the field support different industrial communication protocols, with variations in protocol format, data encoding, and unit definitions, making direct fusion computation impossible. The edge gateway has a built-in multi-protocol parsing engine that can disassemble the raw messages of different protocols, extracting core information such as device identifier, parameter type, value, and acquisition time, and then repackage them into standard data frames according to a unified frame structure. The standard data frame uses a fixed structure, containing seven fields in sequence: data frame sequence number, acquisition timestamp, unique device identifier, parameter type encoding, parameter value, parameter unit, and data validity identifier. All fields use a unified data type and byte length to ensure direct reading and parsing by subsequent modules. The protocol parsing process also performs data validity verification, marking invalid data that exceeds the physical range or has abnormal timestamps, and directly discarding it in the subsequent feature extraction stage to further improve data quality. After filtering and protocol conversion, all data is arranged sequentially according to the acquisition time, forming a continuous, standardized sequence of sensor data frames. In the example, the proportion of outliers in the preprocessed data is less than 0.1%, the data format uniformity reaches 100%, and the single-frame data processing latency is less than 5 milliseconds, which fully meets the timeliness requirements of real-time mapping of digital twins.

[0017] Based on standardized sensor data frame sequences, the operating status features of each process equipment, including start / stop status and load rate, are extracted, and the process cycle time features, including product processing time and waiting time, are extracted to generate a set of equipment status and process cycle time features. The core of this step is to extract key feature indicators that can characterize equipment operating status and production cycle efficiency from standardized time-series data, transforming the raw numerical signals into feature quantities with clear business meaning. This provides structured feature input for the state mapping of the digital twin model and subsequent optimization algorithms. The specific implementation method is as follows: Feature extraction is also performed on the edge gateway side, divided into two main extraction directions: equipment operating status features and process cycle time features. The feature extraction process automatically filters out abnormal data. For feature values ​​exceeding a reasonable range, the feature values ​​from the previous window are used to complete the extraction, preventing abnormal data from interfering with the synchronization accuracy of the subsequent twin model. Equipment operating status features include two core indicators: start / stop status and load rate. Start / stop status is determined by the equipment motor current threshold, with 5% of the rated current set as the start / stop threshold. When the motor current continuously exceeds the threshold for more than 2 seconds, the equipment is determined to be in operating status; when the current continuously falls below the threshold, the equipment is determined to be in standby status. This determination method can accurately distinguish between idle operation and shutdown status, avoiding misjudgments. The load rate is used to characterize the equipment's load saturation level, calculated using the formula η = I_avg / I_rated × 100%, where η is the equipment load rate, I_avg is the average operating current of the motor within the statistical period, and I_rated is the rated current of the motor. A higher load rate indicates higher equipment utilization, while a low load rate indicates a large amount of idle time. In the example, the average operating current of the tempering process when operating at full load is 28.5A, and the rated current of the motor is 33.5A. It can be calculated that the load rate of this process is about 85%, which is in the high-efficiency operating range; the load rate of the edge grinding process is about 72%, and there is a certain amount of idle waiting time.

[0018] The process cycle time characteristic includes two indicators: product processing time and waiting time. Product processing time is calculated by the time difference between the photoelectric signals of the same workpiece entering and exiting the process, representing the actual processing time of a single product within that process, and is a core indicator for measuring process efficiency. Waiting time is calculated by the time difference between the workpiece's entry time and the previous workpiece's exit time, representing the queuing time of the workpiece in the process entry buffer. The longer the waiting time, the tighter the process capacity and the lower the cycle time matching between preceding and following processes. In the example, the processing time of a single laminated glass in the tempering process is 240 seconds, and the average waiting time of the workpiece is 15 seconds, indicating that the tempering process has sufficient capacity and a high cycle time matching degree. The processing time of the lamination process is 180 seconds, and the average waiting time is 45 seconds, indicating that the output speed of the preceding process is higher than the processing speed of the lamination process, and there is a slight cycle time mismatch problem.

[0019] All start / stop status, load rate, processing time, and waiting time corresponding to all processes are structured and organized according to process type and timestamp. Each time window corresponds to a set of full process feature data. The statistical time window of the feature set is set to 30 seconds. The full process feature data is updated every 30 seconds, which ensures the real-time performance of the features and avoids computational redundancy caused by overly fragmented data. Finally, a set of equipment status and process cycle time features is formed, which can be directly used to drive the synchronous update of the digital twin model status.

[0020] Based on the equipment status and process cycle characteristics, a digital twin framework including a physical layer, a virtual layer, a data layer, and an interactive control service layer is constructed in the virtual space. The equipment geometry, operating status, and process flow of the physical production line are mapped and updated synchronously in the virtual space in real time, generating a digital twin model that maps the virtual and real worlds.

[0021] The core of this step is to build a four-layer digital twin system, mapping the geometry, real-time status, and production process of the physical production line to the virtual space in all dimensions. This achieves a one-to-one correspondence between the virtual and physical worlds and real-time synchronization, providing an interactive virtual platform for subsequent process simulation, optimization, and fault diagnosis. The specific implementation method is as follows: The digital twin framework adopts a four-layer architecture, with clearly defined responsibilities and data interoperability between layers. Standardized interfaces facilitate data interaction between layers, and the decoupled design between layers ensures that adding new functional modules requires no changes to the overall framework, providing excellent scalability and adaptability, collectively forming a complete digital twin system. The first layer is the physical layer, corresponding to the physical laminated glass production line, including all processing equipment, sensors, control systems, and work-in-process products. It is the source of all data and the final execution carrier of optimization strategies. The physical layer transmits collected data upwards through edge gateways and simultaneously receives control commands downwards. The second layer is the virtual layer, a digital mirror of the physical production line in virtual space. It first uses 3D modeling technology to recreate the geometry, dimensions, and spatial layout of all equipment, ensuring a complete correspondence between the virtual scene and the spatial position of the physical production line. Modeling accuracy is controlled to the millimeter level, accurately reproducing the shape of the equipment, transmission structure, and the spatial form of key components. The third layer is the data layer, the data hub of the entire framework, responsible for storing all sensor data, feature data, model parameters, and simulation results. It also serves as the intermediary for the interaction between virtual and physical data, enabling the upward synchronization of physical data and the downward transmission of control commands. The fourth layer is the interactive control service layer, which is an application-oriented functional interface layer. It provides service interfaces such as process simulation, optimization deduction, fault diagnosis, and visualization, and supports the subsequent invocation of functions such as reinforcement learning optimization and process parameter search.

[0022] Once the framework is built, real-time virtual-physical mapping and synchronization are executed. Driven by equipment status and process cycle time feature sets, the synchronization frequency is set to 100 milliseconds per update, meaning the virtual layer's equipment status is updated with the latest feature data every 100 milliseconds. Geometric structure mapping remains static and constant, ensuring that the virtual and physical devices completely correspond in shape and position. Operational status mapping synchronizes real-time load rates and start / stop status to the virtual devices, visually displaying their operational status through color changes and numerical annotations. For example, running status is displayed in green, standby status in yellow, and fault status in red, with the load rate value annotated next to the corresponding device in real time. Process flow mapping, based on process cycle time data, synchronously simulates the flow of glass workpieces in the virtual scene. The workpiece movement speed and processing dwell time are completely consistent with the physical production line, fully replicating the entire production process.

[0023] Through continuous real-time synchronization and dynamic updates, the model in the virtual space completely corresponds to the physical production line in three dimensions: geometric structure, operating status, and production process. The data latency is less than 200 milliseconds, and the state synchronization accuracy reaches 99.9%, ultimately forming a digital twin model that maps the virtual and real worlds. This model can serve as a visual monitoring platform to intuitively display the entire production process, and it can also serve as a simulation environment to support subsequent reinforcement learning training, process parameter extrapolation, and fault simulation analysis. It is the core carrier for intelligent interaction across the entire production line.

[0024] S202, Based on the digital twin model, a reinforcement learning algorithm is used to train a global optimization strategy for the production process. The digital twin model is used as an environment interaction platform to dynamically deduce the production line operation status and output the global optimization strategy after reinforcement learning iteration. Specifically, the current production line operating status can be extracted from the digital twin model as the initial observation value. The state space of reinforcement learning is defined as the load rate, cycle time deviation and buffer accumulation of each process equipment, and the action space is the adjustment amount of the operating speed of each equipment, thus generating reinforcement learning environment configuration parameters. The core of this step is to initialize the reinforcement learning training environment, clarify the observable state dimensions and executable action range of the agent, and build a closed-loop training environment using a digital twin model as the interaction carrier. This provides basic configuration support for subsequent policy iteration and optimization. The specific implementation method is as follows: Reinforcement learning is a type of machine learning method that uses an agent to interact with its environment and continuously optimize decision-making strategies guided by reward signals. It is well-suited for dynamic optimization scenarios such as production line scheduling, which involve continuous states and multiple coupled variables. Digital twin models completely replicate the equipment structure, operating logic, and process flow of the physical production line. They can simulate state changes caused by parameter adjustments without interfering with actual production, making them a natural interactive environment for reinforcement learning. Initial observations are taken from the real-time mapping data of the digital twin model at the current moment, representing a snapshot of the actual operation of the production line at the start of training. All state data is synchronized with the physical production line at the millisecond level, ensuring that the initial training conditions closely match actual working conditions.

[0025] The state space defines all environmental features that the agent can observe. This study selects five core processes—cleaning, tempering, lamination, pre-pressing, and autoclave—as optimization nodes for a laminated glass production line. Each process corresponds to three state indicators, collectively forming a continuous state space. The first indicator is the equipment load rate, denoted as `load_i`, representing the proportion of the actual running time of the equipment in the i-th process to the statistical period. Its value ranges from 0 to 1; a value closer to 1 indicates the equipment is operating close to full load, while a value too low indicates excessive idle time. The second indicator is the cycle time deviation, denoted as `beat_dev_i`, in seconds, representing the difference between the actual single-piece processing time of the i-th process and the standard cycle time. A positive value indicates the process is slower than the baseline cycle time, while a negative value indicates it is faster. Excessive deviation can lead to supply and demand mismatches between preceding and following processes. The third indicator is the buffer accumulation, denoted as `buf_i`, in pieces, representing the number of work-in-process items in the buffer area at the i-th process entrance. Excessive accumulation can cause production congestion, while insufficient accumulation may cause material shortages and shutdowns in subsequent processes. In the example, the initial observation state at the start of training includes fifteen state values ​​for five processes. The load rate of the tempering process is 0.82, the cycle deviation is +3 seconds, and the accumulation of 12 pieces in the inlet buffer is 12. The other processes also have their own real-time state values, and all values ​​are within a reasonable operating range.

[0026] The action space defines all the control operations that the agent can perform on the production line. In this case, the adjustment amount of the operating speed of each process's equipment is used as the action variable, denoted as `act_i`, with a value range of -0.15 to +0.15. This represents that the equipment operating speed can be adjusted down by 15% to up by 15% from the baseline speed. Excessive adjustment could lead to uncontrolled process parameters or equipment overload; therefore, upper and lower limits are set to ensure the engineering feasibility of the actions. After defining the state space and action space, parameters such as the state dimension, action dimension, upper and lower limits of each variable's value, and single-step inference time step are uniformly organized to generate reinforcement learning environment configuration parameters. This parameter set serves as a unified benchmark for subsequent policy network construction and environment interaction inference, ensuring parameter consistency during the training process.

[0027] Using a digital twin model as an environment interaction platform for reinforcement learning, the current observation state is input into the policy network, the policy network outputs action decisions and applies them to the virtual devices in the digital twin model, driving the digital twin model to perform dynamic inference, and generating the next state and production line output results after the inference. The core of this step is to leverage the high-fidelity simulation capabilities of digital twins to virtually implement policy decisions and state evolution, completing the state transition of single-step reinforcement learning, and providing interactive data for subsequent reward calculation and policy updates. The specific implementation method is as follows: Compared to direct trial and error debugging on physical production lines, using digital twin models as an interaction platform enables cost-free and risk-free iterative testing, avoiding product scrap, equipment damage, and production stoppages caused by improper parameter adjustments, and significantly improving the iteration efficiency of optimization strategies. The policy network is the core decision-making unit of the reinforcement learning agent, constructed using a multi-layer fully connected neural network. The network input is the current state vector, and the output is the corresponding action vector. The network's weight parameters are continuously iterated and optimized during training, gradually improving the quality of decision-making.

[0028] The complete process of single-step interactive simulation is as follows: First, the current state vector s_t is input into the policy network π_θ, where θ represents the current weight parameters of the policy network. After forward computation, the network outputs a continuous action vector a_t, where each element corresponds to the speed adjustment of a process. In the example, after inputting an initial observation vector containing fifteen states, the policy network outputs the speed adjustment for five processes: cleaning process +0.04, tempering process +0.05, lamination process +0.03, pre-pressing process 0, and autoclave process -0.02. This represents that the corresponding processes are increased in speed by 4%, 5%, and 3% respectively based on the baseline speed, the pre-pressing process maintains its original speed, and the autoclave process is reduced in speed by 2%.

[0029] The action decisions are then synchronously applied to the corresponding virtual devices in the digital twin model. The model performs a dynamic simulation of the entire process according to a preset 30-second single-step time step. The simulation process fully follows the process flow logic and equipment operation rules of the physical production line. The temperature field changes in the tempering furnace, the bonding reaction in the lamination process, and the material flow in the buffer zone are all simulated and calculated according to real physical laws to ensure the authenticity and credibility of the simulation results. After the single-step simulation is completed, the model automatically outputs the state vector s_t+1 for the next moment and simultaneously calculates the production line output results within that time step, including three core indicators: product output per unit time, average utilization rate of equipment in the entire process, and comprehensive energy consumption per unit product. In the example, after the single-step simulation, the load rate of the tempering process increased to 0.86, the cycle deviation decreased to +1 second, the buffer accumulation decreased to 9 pieces, and the output per unit time increased by 6.2% compared to the baseline. This fully presents the changes in the production line state brought about by the action adjustment. All simulation data together constitute the sample data of the single-step interaction, which can be directly used for subsequent reward calculations.

[0030] The comprehensive reward value is calculated based on the production line output results after simulation. The reward function includes a weighted sum of three dimensions: product output, equipment utilization rate, and energy consumption. The reward value is then used to update the weight parameters of the policy network through the near-end policy optimization algorithm to generate the updated policy network. The core of this step is to construct a multi-objective fusion reward-incentive mechanism, quantify the merits of individual decisions, and then use a stable and efficient policy optimization algorithm to iterate network parameters, allowing the decision-making strategy to gradually evolve towards the global optimum. The specific implementation method is as follows: The reward function is the core guiding mechanism of reinforcement learning, determining the optimization direction of the agent. This study adopts a multi-dimensional weighted summation comprehensive reward function, taking into account three production objectives: capacity improvement, efficient equipment utilization, and energy consumption reduction. The formula is expressed as R=w_1 R_output+w_2 R_util-w_3 R_energy. In the formula, R is the overall reward value; a higher value indicates better overall efficiency of the action decision. w_1, w_2, and w_3 are the weight coefficients for the three optimization dimensions, with the sum of all weights equal to 1, corresponding to the priorities of output, equipment utilization, and energy consumption, respectively. Based on the actual operational needs of the laminated glass production line, w_1=0.5, w_2=0.3, and w_3=0.2 are set, representing that capacity improvement is the primary optimization goal, followed by equipment utilization efficiency, and energy consumption is the third optimization dimension. The weight settings can be flexibly adjusted according to the operational needs of different production lines.

[0031] All three reward components are calculated using a normalized method. R_output is the output reward component, calculated as the ratio of the current step's output per unit time to the baseline output; a higher ratio results in a higher reward. R_util is the equipment utilization reward component; the highest reward is obtained when the average load rate of the entire process is within the optimal operating range of 0.7 to 0.9. Deviating from this range results in a linear decrease in reward with the deviation, preventing equipment overload or excessive idleness. R_energy is the energy consumption penalty component, calculated as the ratio of the current step's energy consumption per unit of product to the baseline energy consumption; a higher ratio results in a greater penalty, hence it is included as a negative value in the overall reward formula. In the example, after a single-step simulation, the output per unit time is 1.062 times the baseline value, the average equipment load rate is 0.78 (within the optimal range), and the energy consumption per unit of product is 1.021 times the baseline value. Substituting these values ​​into the formula yields an overall reward value of 1.037, a significant improvement compared to the baseline reward value of 1, indicating that this set of action decisions has a positive optimization effect.

[0032] The proximal policy optimization algorithm is the policy update algorithm used in this study. This algorithm avoids drastic policy oscillations by limiting the magnitude of parameter updates in each round, and boasts advantages such as stable training, high sample utilization, and simple hyperparameter tuning, making it highly suitable for optimization scenarios with high stability requirements, such as industrial production lines. The core of the algorithm is to prune the objective function, constraining the probability ratio of the new and old policies within the range of 1-ε to 1+ε. In this study, a pruning threshold of ε=0.2 was set to ensure smooth and controllable policy updates. During the update process, the advantage function is first calculated using samples collected through multi-step interactions to measure the superiority or inferiority of the current action relative to the average level. Then, the pruned objective function is maximized through gradient ascent, completing the iterative update of the policy network weights θ. After the update, new weight parameters θ_new are obtained, corresponding to the generated updated policy network, further improving the network's decision accuracy compared to before the update.

[0033] Repeat the iterative process of state observation, action decision-making, deduction, reward calculation and network update until the reward value converges or reaches the preset number of iterations, and finally output the global optimization strategy after reinforcement learning iteration optimization.

[0034] The core of this step is to continuously optimize the decision-making ability of the strategy network through multiple rounds of closed-loop iteration until the strategy performance reaches a stable optimal state, and finally output a globally optimized strategy that can be directly used for production line control. The specific implementation method is as follows: A single iteration consists of five stages: state observation, action decision output, digital twin dynamic inference, comprehensive reward calculation, and policy network parameter update. With each iteration, the weight parameters of the policy network are optimized, improving decision-making capabilities. During the iteration process, each round uses the next state obtained from the previous round's inference as the initial observation state for the new round, forming a continuous interaction trajectory and continuously providing training samples for the algorithm.

[0035] The training is set with two termination conditions. The first condition is the reward convergence condition. When the average comprehensive reward value fluctuates by less than 1% over 20 consecutive iterations, the strategy performance is considered to have reached a stable state. Continuing to iterate will not bring significant performance improvement, and training can be terminated early. The second condition is the maximum number of iterations condition. The preset maximum number of iterations is 1000. Even if the reward does not converge, training will stop after reaching this number to avoid unlimited consumption of computing resources. Iteration can be terminated if either of the two conditions is met.

[0036] In the example, the policy network weights are randomly initialized in the early stages of iteration, resulting in highly random action decisions and drastic fluctuations in the overall reward value, which may even fall below the baseline value. As the number of iterations increases, the policy network gradually learns the production line operation rules under the guidance of the reward signal, and the reward value shows a steady upward trend with a continuously narrowing fluctuation range. When the iteration reaches the 782nd round, the average reward value of the 20 consecutive rounds stabilizes at around 1.12, with a fluctuation range of 0.78%, which meets the convergence condition, and the training automatically terminates.

[0037] After training terminates, the decision rules corresponding to the final strategy network constitute the global optimization strategy. This strategy can quickly output an optimal equipment speed adjustment scheme that balances capacity, equipment utilization, and energy consumption based on the real-time operating status of the production line at any given time, achieving collaborative optimization across the entire process. Compared to traditional manual scheduling methods, this optimization strategy can improve the overall output efficiency of the production line by more than 12%, while reducing unit product energy consumption by approximately 5%. Equipment operation remains within its optimal load range, effectively extending equipment lifespan. This global optimization strategy also provides a global scheduling framework and operating speed benchmark for subsequent refined optimization of process parameters.

[0038] S203, based on the global optimization strategy, simulate and deduce the temperature field distribution, pressure change curve and cycle connection relationship of each process in the production process, and with product quality consistency as a constraint, search for the optimal combination of heating temperature, pressure parameters and rolling time to generate the optimized process parameter combination. Specifically, the target operating rate and temperature setting range of each process can be extracted from the global optimization strategy, and a virtual process simulation environment including the heating section of the tempering furnace, the lamination chamber and the autoclave can be built in the digital twin model. The core of this step is to extract the constraint boundaries of process parameters from the global optimization scheme at the production line level, and to build a high-precision simulation platform for the three core processes based on the virtual simulation capabilities of the digital twin model. This ensures that subsequent process parameter derivations fully conform to the hardware capabilities and operating cycle requirements of the production line, providing a standardized virtual experimental environment for multiphysics simulation and parameter optimization. The specific implementation method is as follows: The global optimization strategy is a production line-level operational optimization scheme output by reinforcement learning algorithms. It covers macro-level operational rules such as equipment load distribution, process cycle time matching, and total energy consumption control. Among them, the target operating rate and temperature setting range are two core boundary parameters that directly affect process simulation. The target operating rate represents the output of qualified products per unit time of a single process, measured in pieces per minute. This parameter directly determines the cycle time connection between preceding and following processes. If the rate is too high, it will cause insufficient material supply in the preceding process or material accumulation and blockage in the following process. If the rate is too low, it will reduce the production efficiency of the entire production line. The temperature setting range limits the adjustable temperature upper and lower limits of thermal processes to avoid quality defects such as product cracking and adhesion failure caused by parameters exceeding the limits, while taking into account equipment lifespan and energy consumption control requirements. In the example, the target operating rate of the tempering furnace heating section extracted from the global optimization strategy is 12 pieces per minute, with a temperature setting range of 680 degrees Celsius to 720 degrees Celsius; the target operating rate of the laminating chamber is 11 pieces per minute, with an ambient temperature setting range of 22 degrees Celsius to 26 degrees Celsius and an ambient relative humidity range of 30% to 40%; the target operating rate of the autoclave is matched with that of the preceding and following processes, with an operating temperature setting range of 120 degrees Celsius to 140 degrees Celsius and an operating pressure range of 1.0 MPa to 1.4 MPa. All parameter boundaries meet the rated operating limits of the production line equipment.

[0039] The virtual process simulation environment is built within the virtual layer of the digital twin model. It is a 1:1 high-precision digital replica of the core processes of the physical production line, completely restoring the physical characteristics of the equipment geometry, internal heating element arrangement, roller conveyor system, and fluid pipeline routing of the tempering furnace heating section, laminating chamber, and autoclave. It also embeds corresponding heat conduction, fluid mechanics, and timing logic simulation engines. The tempering furnace heating section simulation module has three independent heating units (upper, middle, and lower), allowing for independent adjustment of heating power and set temperature in each zone, simulating the surface temperature evolution of glass workpieces during roller conveying. The laminating chamber simulation module replicates a constant temperature and humidity clean operating environment, simulating the bonding and rolling process of glass substrates and PVB films, and can output the influence of rolling pressure and speed on laminating quality. The autoclave simulation module covers the entire process flow of heating, pressurizing, heat preservation and pressure holding, and cooling and depressurization, and can output complete pressure-time and temperature-time curves. The completed virtual process simulation environment strictly follows the constraints extracted by the global optimization strategy for all parameter values, ensuring that the subsequent simulation results are fully adapted to the actual operating capabilities of the physical production line and that there are no invalid parameter schemes that exceed the limits of the equipment hardware.

[0040] In a virtual process simulation environment, the temperature field distribution of each heating zone of the tempering furnace is simulated and calculated, the pressure change curve of the autoclave is dynamically simulated, and the timing relationship between the laminator and the preceding and following processes is deduced to generate a multiphysics simulation result set. The core of this step is to conduct multi-dimensional synchronous simulation calculations in a virtual simulation environment to acquire three types of core simulation data: temperature field, pressure curve, and cycle time sequence. This constructs a multi-dimensional sample space covering thermophysics, mechanics, and production logic, providing complete data support for subsequent optimization of process parameters. The specific implementation method is as follows: The temperature field distribution simulation of the tempering furnace uses the finite difference method to solve the unsteady heat conduction equation, and the corresponding governing equation is ρc. T / t=λ ²T+q, where ρ represents the density of the glass material (kg / m³); c represents the specific heat capacity of the glass at constant pressure (joules / kg / degrees Celsius); T represents the internal node temperature of the glass (degrees Celsius); t represents the heat transfer time (seconds); λ represents the thermal conductivity of the glass (watts / m³ / degrees Celsius); and q represents the internal heat source density applied by the heating element (watts / m³ / degrees Celsius). The simulation process involves inputting set temperature values ​​for the upper, middle, and lower heating zones of the tempering furnace, simulating the complete transport process of the glass workpiece from entry to exit. The temperature values ​​of all nodes on the glass surface are calculated moment-by-moment, and key indicators such as the maximum temperature difference, peak center temperature, and edge temperature are output. In the example, when the set temperature for the upper zone of the tempering furnace is 700 degrees Celsius, the middle zone is 690 degrees Celsius, and the lower zone is 705 degrees Celsius, the simulation shows a maximum temperature difference of 12 degrees Celsius on the glass surface and a peak center temperature of 625 degrees Celsius. The temperature uniformity meets the basic requirements of the tempering process.

[0041] The pressure change curve simulation of the autoclave simulates the entire high-pressure curing cycle of laminated glass. The pressure curve is divided into three stages: pressure increase, constant pressure, and pressure decrease. The simulation process synchronously couples the effect of temperature changes inside the autoclave on the pressure, outputting the real-time pressure value inside the autoclave second by second, forming a continuous and smooth pressure-time change curve. In the example, the pressure increase stage is set to 25 minutes, and the pressure inside the autoclave gradually increases from atmospheric pressure to 1.2 MPa, with the pressure increase rate remaining uniform and without sudden changes; the constant pressure stage is set to 50 minutes, and the pressure is maintained stably at 1.2 MPa, with fluctuations controlled within ±0.02 MPa; the pressure decrease stage is set to 30 minutes, and the pressure decreases uniformly from 1.2 MPa to atmospheric pressure to avoid glass cracking or film delamination defects caused by sudden pressure changes.

[0042] The cycle time correlation simulation is based on the target operating rate of each process, simulating the flow of glass workpieces between the tempering section, the lamination chamber, and the autoclave. It calculates the work-in-process quantity, buffer waiting quantity, and equipment idling time for each process at each time step, identifying process cycle time bottlenecks and material accumulation risks. In the example, the tempering section operates at a rate of 12 pieces per minute, the lamination chamber at 11 pieces per minute, and the autoclave at 11.5 pieces per minute. The simulation shows an average accumulation of 3 pieces in the buffer before the lamination chamber, with a maximum accumulation not exceeding 6 pieces, eliminating the risk of prolonged material blockage, and controlling the cycle time correlation deviation between preceding and following processes within 5%. Temperature field distribution data, pressure change curve data, and cycle time correlation simulation data are integrated according to process dimensions and time axis to form a multi-physics simulation result set covering thermal, mechanical, and temporal dimensions. This result set contains complete simulation outputs corresponding to multiple sets of different parameter combinations, providing ample sample space for subsequent parameter optimization.

[0043] With product quality consistency indicators including at least the bonding strength of laminated glass, bubble residual rate and optical distortion rate as constraints, a Bayesian optimization algorithm is used to search for the optimal combination of heating temperature distribution, pressure curve parameters and rolling time in the multiphysics simulation result set to generate a candidate process parameter combination set. The core of this step is to use hard product quality indicators as the entry constraint, and to use Bayesian optimization algorithms to efficiently perform optimization calculations in a high-dimensional parameter space. This allows for the selection of all process parameter schemes that meet the quality requirements with fewer simulation iterations, avoiding the huge computational overhead of traversing the entire space. The specific implementation method is as follows: Product quality consistency indicators are rigid constraints for judging the qualification of process parameter combinations. Each of the three core indicators has a clear physical meaning and a qualification threshold; any parameter scheme exceeding the threshold will be directly rejected. Bond strength represents the degree of interfacial bonding between the glass substrate and the PVB film, measured in megapascals (MPa). The qualification threshold is set at no less than 0.8 MPa; a higher value indicates a more reliable bonding interface and less likelihood of delamination under external impact. Bubble residue rate represents the percentage of residual bubbles inside the interlayer after curing, measured as a percentage. The qualification threshold is set at no more than 0.1%. Excessive residual bubbles not only affect the optical appearance of the glass but also reduce local bonding strength. Optical distortion rate represents the degree of optical deformation of the finished glass, measured in angle minutes. The qualification threshold is set at no more than 2 minutes; an excessively high distortion rate will affect the visual appearance and optical performance of the glass.

[0044] Bayesian optimization is a highly efficient optimization algorithm adapted to expensive simulation evaluation scenarios. Its core logic involves constructing a Gaussian process surrogate model based on existing simulation samples. By balancing global exploration and local development through a data acquisition function, it finds the optimal parameter combination with the fewest simulation evaluations, significantly reducing the computational cost of high-dimensional parameter optimization. The parameter search space in this study includes seven adjustable dimensions: the heating temperature of the three zones of the tempering furnace, the pressurization rate and constant pressure of the autoclave, and the lamination rolling time. The value range of each dimension strictly adheres to the boundaries defined by the global optimization strategy. The algorithm's objective function is the overall product quality score, and the constraints are the pass / fail thresholds of three quality indicators. Only parameter combinations that simultaneously satisfy all constraints are included in the candidate set.

[0045] In the example, the Bayesian optimization algorithm first randomly selects 15 parameter combinations as initial samples, substitutes them into the virtual process simulation environment to obtain the corresponding quality index results, and then trains a Gaussian process surrogate model based on the initial samples to fit the mapping relationship between process parameters and quality indicators. Then, it uses the expectation-lift acquisition function to select the next parameter combination with the greatest optimization potential for simulation evaluation. The surrogate model is updated after each evaluation, and the algorithm continuously iterates and optimizes. After 32 iterations, the algorithm converges, selecting 8 parameter schemes that simultaneously satisfy the three quality constraints, forming a candidate set of process parameter combinations. Each candidate combination includes complete set values ​​for the three zones of the tempering furnace, the pressurization time of the autoclave, the constant pressure value, the lamination rolling time, and the corresponding simulation quality index results. All combinations have a bonding strength above 0.85 MPa, a bubble residue rate below 0.08%, and an optical distortion rate below 1.7 points, all meeting the product quality consistency constraints.

[0046] The product quality index is scored for each combination in the candidate process parameter combination set, and the combination with the highest score is selected as the optimal solution, thus generating the optimized process parameter combination.

[0047] The core of this step is to comprehensively rank all qualified candidate solutions through weighted quantitative scoring, select the process parameter solution with the best quality performance and overall performance, and compile it into standardized process parameters that can be directly deployed to the physical production line for execution. The specific implementation method is as follows: The comprehensive score for product quality indicators is calculated using a weighted summation method. The corresponding scoring formula is S = w_1 × S_1 + w_2 × S_2 + w_3 × S_3, where S represents the comprehensive quality score, ranging from 0 to 100. A higher score indicates a better overall quality performance of the process solution. S_1 is the score for bonding strength, S_2 is the score for bubble residue rate, and S_3 is the score for optical distortion rate. w_1, w_2, and w_3 are the weight coefficients for the three indicators. Considering the product performance priority and application requirements of laminated glass, w_1 is set to 0.4, w_2 to 0.35, and w_3 to 0.25, with the sum of the three weights being 1. The individual scores are converted according to the relative performance of the indicator value and the pass threshold. Higher bonding strength corresponds to a higher individual score, and lower bubble residue rate and optical distortion rate correspond to higher individual scores. All individual scores are mapped to a score range of 0 to 100.

[0048] In the example, the comprehensive score is calculated for each of the eight candidate combinations. The first combination has a bond strength of 0.92 MPa (94 points), a bubble residue rate of 0.06% (96 points), and an optical distortion rate of 1.5% (92 points). Substituting these values ​​into the scoring formula, the comprehensive score is 94 × 0.4 + 96 × 0.35 + 92 × 0.25 = 93.8 points. The second combination has a bond strength of 0.88 MPa (89 points), a bubble residue rate of 0.05% (97 points), and an optical distortion rate of 1.3% (94 points). The comprehensive score is 89 × 0.4 + 97 × 0.35 + 94 × 0.25 = 93.05 points. After calculating the scores for all candidate combinations, the comprehensive scores are compared, and the combination with the highest score is selected as the final optimal solution. The highest-scoring solution corresponds to a tempering furnace temperature of 698°C in the upper zone, 689°C in the middle zone, and 703°C in the lower zone; a high-pressure autoclave pressurization time of 24 minutes; a constant pressure of 1.22 MPa; and a lamination roller pressing time of 12 seconds. This solution achieved a comprehensive quality score of 94.2, with all three quality indicators falling within the excellent range. Furthermore, the process cycle time was smooth, and the energy consumption level met the overall optimization requirements. All process parameters of this solution were categorized and organized according to the process steps, with each parameter's meaning, value, and corresponding process step clearly labeled. This resulted in a complete and standardized optimized combination of process parameters that can be directly implemented on the physical production line. This combination balances product quality consistency, production cycle efficiency, and energy consumption control, fully adapting to the intelligent operation requirements of laminated glass production lines.

[0049] S204, the optimized process parameter combination is subjected to execution effect deviation detection. At the same time, the digital twin model is combined with the pre-built equipment health assessment model to perform trend analysis on the operating status data of key equipment components, predict potential equipment failure hazards, and generate early warning and diagnostic information containing failure types. Specifically, the optimized combination of process parameters can be sent to the digital twin model for virtual execution. The predicted product quality and equipment response values ​​after virtual execution are collected, and the deviation is calculated with the theoretical target value of the combination of process parameters to generate the execution effect deviation detection results. The core of this step is to leverage the virtual simulation capabilities of digital twin models to complete preliminary verification before implementing process parameters on the physical production line. By quantifying the deviation between the actual effect of the parameters and the theoretical target through virtual execution, the risks of material loss and production line downtime caused by direct trial production are avoided, ensuring the stability and consistency of the process parameters after implementation. The specific implementation method is as follows: The optimized process parameter combination includes core process values ​​such as the set temperatures of each temperature zone of the tempering furnace, the pressure rise and holding curve parameters of the autoclave, and the lamination rolling time. It is the theoretically optimal solution obtained through global optimization strategy and multiphysics simulation screening, but it has not yet been verified by actual production line operation. Directly issuing it may lead to uncontrollable problems such as equipment response lag and multi-process coupling deviation. The parameter combination is completely issued to the virtual process simulation environment of the digital twin model, driving the virtual production line to execute a complete round of laminated glass processing simulation according to the complete production process. The virtual execution process completely replicates the equipment response characteristics, material transfer delays and process connection logic of the physical production line, and the consistency of the output results with the physical production line can reach more than 95%.

[0050] After virtual execution, two types of core output data are collected simultaneously. The first type is the product quality prediction value, covering three core quality indicators: bonding strength, bubble residue rate, and optical distortion rate. Bonding strength represents the firmness of the bond between the film and the glass substrate, measured in megapascals (MPA); a higher value indicates stronger bonding reliability. Bubble residue rate represents the area ratio of residual bubbles inside the laminate after high-pressure lamination; a lower value indicates better optical quality. Optical distortion rate represents the degree of optical deformation of the finished glass; a lower value indicates a more uniform visual effect. The second type is the equipment response value, including three parameters: the actual temperature rise rate of the tempering furnace, the pressure response lag time of the autoclave, and the actual load rate of the roller pressing mechanism. These parameters reflect the difference between the actual operating state of the equipment under given parameters and the theoretical settings.

[0051] Deviation calculation is performed index-by-index using a relative deviation formula: δ=|(X_p-X_t)| / X_t×100%, where δ represents the relative deviation of a single index in percentage form; X_p represents the actual predicted value obtained from the virtual execution; and X_t represents the theoretical target value corresponding to the process parameters. In the example, the optimized process parameter combination sets the temperatures of the three zones of the tempering furnace to 720°C, 710°C, and 690°C, the autoclave pressurization rate to 0.08 MPa / s, and the rolling time to 12 seconds. The corresponding theoretical target values ​​are a bond strength of 8.5 MPa, a bubble residue rate of 0.3%, and an optical distortion rate of 0.18%. The predicted quality values ​​obtained after virtual execution are a bond strength of 8.2 MPa, a bubble residue rate of 0.32%, and an optical distortion rate of 0.19%. Substituting these values ​​into the formula, the relative deviations for the three indices are calculated to be 3.5%, 6.7%, and 5.6%, respectively. Regarding equipment response values, the theoretical target for the temperature rise rate of the tempering furnace is 15 degrees Celsius per minute, while the virtual measured value is 14.2 degrees Celsius per minute, with a relative deviation of 5.3%. The theoretical pressure response lag of the autoclave is 2 seconds, while the virtual measured value is 2.7 seconds, with a relative deviation of 35%.

[0052] The preset acceptable deviation thresholds for quality indicators are 5% and for equipment response indicators are 10%. Comparative calculations show that the bonding strength deviation meets requirements, but the bubble residue rate and optical distortion rate exceed the quality thresholds, and the pressure response hysteresis deviation exceeds the equipment thresholds. This indicates an execution deviation in these parameters, requiring fine-tuning. A preliminary analysis and integration of all indicator deviation values, acceptance / acceptance conclusions, and causes of deviation are conducted to generate the final execution effect deviation detection results, providing a quantitative basis for subsequent secondary parameter optimization.

[0053] Extract time-series data of the operating status of key equipment components from the digital twin model, including bearing temperature sequence, motor current RMS value sequence and vibration characteristic frequency sequence, to generate a dataset of the operating status of key components; The core of this step is to extract full-dimensional time-series operational data of core vulnerable components of the production line from the real-time synchronized digital twin model, complete data cleaning and structuring, and construct a status dataset covering three types of characteristics: thermal, electrical, and vibration. This provides standardized data input for subsequent equipment health assessment and fault early warning. The specific implementation method is as follows: Key equipment components in laminated glass production lines mainly include tempering furnace drive roller bearings, laminator drive motors, autoclave circulating pump bearings, and cleaning roller drive assemblies. These components operate under high temperature, high load, and continuous conditions for extended periods, making them critical points for production line failures and key targets for health monitoring. A digital twin model synchronizes with the physical production line in real time. All operational data collected by sensors is synchronously mapped to the corresponding components of the virtual equipment. Therefore, historical time-series data of the corresponding components can be directly retrieved from the digital twin model without additional interface to the physical equipment, resulting in higher data retrieval efficiency and stronger time-series consistency.

[0054] This study extracts three types of core time-series data, each with a unified timestamp and a uniform sampling interval of 30 seconds to ensure that the data time granularity matches the accuracy requirements of the health assessment. The first type is the bearing temperature sequence, collected by a temperature sensor mounted on the bearing housing, measured in degrees Celsius. Temperature changes directly reflect the bearing's friction and wear degree and lubrication status. During normal operation, the temperature remains stable within a reasonable range, but as wear intensifies, the temperature will continue to rise. The second type is the motor current RMS value sequence, collecting the RMS values ​​of the three-phase AC current of the drive motor, measured in amperes. The current magnitude directly corresponds to the motor's load condition; faults such as bearing jamming and rotor imbalance will cause increased current fluctuations and higher peak values. The third type is the vibration characteristic frequency sequence, obtained by collecting the raw vibration signal from a vibration acceleration sensor and performing a fast Fourier transform. The amplitude corresponding to characteristic frequencies such as the bearing inner ring fault frequency, outer ring fault frequency, cage fault frequency, and motor rotation frequency is extracted, measured in gravitational acceleration g. Different fault types will correspond to an increase in the amplitude of specific frequencies, which is the core basis for identifying fault types.

[0055] The data extraction timeframe was set to the most recent 7 consecutive days, covering the complete operating cycle and load fluctuation range of the equipment, totaling 20,160 sampling points, which can fully reflect the operating trends and fluctuation patterns of the components. After extraction, basic preprocessing was performed on the raw time-series data. For a small number of missing values ​​and abnormal jump points caused by network packet loss and instantaneous sensor interference, linear interpolation was used to complete and correct them, ensuring that the time-series data is continuous and uninterrupted, and will not affect the accuracy of subsequent health assessments.

[0056] In the example, data for the drive end bearing of the No. 1 transmission roller conveyor of the tempering furnace was extracted. The average bearing temperature sequence was 48.5 degrees Celsius, with a fluctuation range of 42 to 55 degrees Celsius. The temperature was higher during daytime production and decreased during nighttime standby. The average effective current sequence of the matching drive motor was 12.6 amps, with a peak of 28.3 amps at equipment startup, and the current fluctuation amplitude was less than 0.5 amps during stable operation. In the vibration characteristic frequency sequence, the bearing outer ring fault frequency amplitude was 0.08g, the inner ring fault frequency amplitude was 0.12g, and the rotational frequency amplitude was 0.05g, all within the normal fluctuation range. The three types of time-series data of all key components were structured and integrated according to the order of equipment number, component name, and timestamp to generate a key component operating status dataset. The dataset has a unified data format, continuous time sequence, and complete features, and can be directly input into the health assessment model for analysis.

[0057] The key component operating status dataset is input into a pre-built equipment health assessment model. This model uses the isolated forest algorithm to detect anomalies in the status data, calculates the health score of each component, and generates component health trend analysis results. The core of this step is to rely on a pre-trained equipment health assessment model to identify abnormal features in the operating status of components through an unsupervised anomaly detection algorithm, quantify the component health score, and analyze the time-series change trend of the health score. This intuitively presents the component's degradation rate and health trajectory, providing a quantitative basis for subsequent fault early warning. The specific implementation method is as follows: The equipment health assessment model is a dedicated assessment model pre-trained based on historical production line operating data and fault samples. Its core function is to transform multi-dimensional operating status data into intuitive health scores. The core algorithm of the model is the Isolation Forest algorithm, an unsupervised anomaly detection algorithm that can identify anomalies without requiring a large number of labeled fault samples, making it highly suitable for the characteristics of industrial scenarios where fault samples are scarce and normal samples are abundant. The core principle of the Isolation Forest algorithm is to construct multiple isolated trees by randomly selecting features and randomizing the partitioning threshold. Each sample is isolated, and abnormal samples, because their features deviate from the normal group, will be isolated in fewer partitions, resulting in a shorter average path length. This is used to calculate the anomaly score of the sample, distinguishing between normal and abnormal states.

[0058] After anomaly detection, the anomaly score is mapped to a health score ranging from 0 to 100, using the formula H = 100 × (1 - s(x)). In this formula, H represents the component health score, which is dimensionless; a higher score indicates a better component health, with 100 points corresponding to a brand-new, intact state and 0 points corresponding to a completely failed state. s(x) represents the sample anomaly score output by the isolated forest, ranging from 0 to 1; a score closer to 1 indicates a higher degree of anomaly. After health calculation, a 24-hour sliding window is used to fit the historical health data to calculate the average daily rate of decline in health, expressed in minutes per day. This rate quantifies the speed of component degradation; a faster decline rate indicates a faster failure progression and a shorter remaining reliable operating time.

[0059] In the example, after inputting the key component operating status dataset into the model, the Isolation Forest algorithm calculates the anomaly score and health score for each component. The anomaly score for the No. 1 transmission roller bearing is 0.18, resulting in a health score of 82. The anomaly score for the matching drive motor is 0.07, corresponding to a health score of 93. The anomaly score for the autoclave circulating pump bearing is 0.11, corresponding to a health score of 89. Trend analysis results show that the health score of the No. 1 roller bearing has continuously decreased from 91 to 82 over the past 7 days, with an average daily decrease rate of 1.29 points, indicating a gradually accelerating deterioration trend. The health score of the drive motor has remained stable between 92 and 94, with an average daily fluctuation of less than 0.5 points, showing no significant deterioration trend. The health score of the circulating pump bearing has slowly decreased from 92 to 89, with an average daily decrease rate of 0.43 points, indicating a gradual deterioration rate.

[0060] The current health score, health change curve over the past 7 days, average daily rate of decline, and degradation trend level of all components are integrated and analyzed. Based on the health score, components are categorized into four levels: Excellent, Good, Warning, and Severe. A score of 85 or above is considered Good, 70-85 is Warning, and below 70 is Severe. The final component health trend analysis results clearly present the current health status and future evolution trend of each key component, intuitively distinguishing between high-risk and low-risk components, providing a core basis for subsequent warning triggering and fault identification.

[0061] Based on the component health trend analysis results, an early warning is triggered when the health score is lower than the preset threshold. A fault mode matching algorithm is used to identify potential fault types, and finally, early warning diagnostic information containing fault types is generated.

[0062] The core of this step is to trigger early warning levels based on health thresholds, match fault characteristics to locate specific fault types, and output structured early warning diagnostic conclusions. This enables early identification and accurate classification of equipment faults, supporting subsequent fault location and maintenance plan generation. The specific implementation method is as follows: Two preset health warning thresholds are implemented. The first level, a yellow threshold, is set at 75 points. When a component's health falls below this threshold, it indicates significant degradation and a potential risk of failure, requiring planned maintenance. The second level, a red threshold, is set at 60 points. When the health falls below this threshold, the component's risk of failure is extremely high, and a shutdown is imminent, requiring immediate repair. In addition to the absolute health threshold, a degradation rate threshold is also set. If the average daily rate of decline exceeds 1 point, a warning will be triggered even if the health has not yet fallen below the warning threshold, preventing rapidly deteriorating components from suddenly failing before the threshold is triggered.

[0063] The fault mode matching algorithm employs a feature similarity matching mechanism. It pre-constructs a fault feature template library containing various typical faults. Each fault type corresponds to a standardized feature vector, with features including core indicators such as temperature rise rate, current fluctuation coefficient, and the proportion of amplitudes at different vibration frequencies. When a component triggers an early warning, the current state feature vector of that component is extracted and its cosine similarity is calculated against all fault templates in the library. The similarity calculation formula is sim(A,B)=(A...). The formula is sim(A,B) / (|A|×|B|), where sim(A,B) represents the cosine similarity between two feature vectors, ranging from 0 to 1, with higher values ​​indicating higher matching. A is the feature vector of the component to be diagnosed, and B is the feature vector of the fault template. The fault type with the highest similarity is selected as the potential fault type of the component. When the matching similarity is lower than 0.7, the fault type is considered unclear and marked as an abnormal state to be investigated. Common fault types include bearing outer ring wear, poor bearing lubrication, motor rotor imbalance, aging seals, and abnormal noise from transmission components, covering more than 90% of common faults of key components on the production line.

[0064] In the example, the current health score of the No. 1 transmission roller bearing is 82 points, which has not yet reached the yellow warning threshold of 75 points. However, its daily decline rate is 1.29 points per day, exceeding the deterioration rate threshold of 1 point per day, thus triggering a yellow warning. The feature vector of this component is extracted and matched with the fault template library. The calculated cosine similarity with the bearing outer ring wear template is 0.89, the highest matching degree. Therefore, the potential fault type is determined to be outer ring wear of the transmission roller bearing. Based on the deterioration rate, the remaining reliable operating time of this component is estimated to be approximately 12 days. If maintenance is not performed in time, significant failure will occur within two weeks.

[0065] By integrating information such as warning level, associated equipment and components, current health status, degradation rate, potential fault type, and estimated remaining runtime, a warning and diagnostic message containing the fault type is generated. This diagnostic message has a clear structure and specific direction, including both qualitative judgment of risk level and precise identification of fault type. It can directly guide subsequent fault tracing and repair plan matching, effectively improving the pertinence and efficiency of equipment operation and maintenance, and reducing the risk of unplanned downtime.

[0066] S205, based on the aforementioned early warning and diagnostic information, a digital twin model is used to trace the fault back and locate it spatially, and a repair plan corresponding to the fault type is matched with a pre-set expert knowledge base to generate a fault solution that includes fault location identification and maintenance suggestions.

[0067] Specifically, it can parse the fault type identifier and fault occurrence timestamp in the early warning and diagnosis information, replay the time series data of equipment operation status before and after the fault in the digital twin model, trace the key nodes of fault evolution along the time axis, and generate a fault backtracking trajectory diagram. The core of this step is to leverage the full-time-series data storage capabilities of the digital twin model to trace the entire evolution of the fault backward from the early warning trigger point, reconstructing the complete time chain from the fault's inception and development to triggering the early warning, identifying the key inflection points in the fault's evolution, and providing a temporal basis for subsequent spatial positioning and solution matching. The specific implementation method is as follows: The early warning diagnostic information is a structured early warning message output by the equipment health assessment model. It includes four core fields: fault type identifier, early warning trigger timestamp, health score, and abnormal characteristic parameters. The fault type identifier is used to distinguish different fault modes. The example includes more than ten typical faults such as bearing fatigue wear, heating wire open circuit, roller rubber layer aging, and transmission chain loosening. Each fault corresponds to a unique numerical code, which facilitates automatic identification and retrieval by the system. The fault occurrence timestamp adopts a unified format of year, month, day, hour, minute, second, and millisecond, with an accuracy controlled within 10 milliseconds, accurately marking the moment when the early warning signal is first triggered. In the example, the trigger timestamp for a fault in the roller bearing of a tempering furnace is 20240615142830120, corresponding to 14:28:30.120 on June 15, 2024.

[0068] After parsing the two types of core identifiers, the time-series playback function of the digital twin model is invoked. The data layer of the digital twin model continuously stores the time-series data of the operating status of all equipment on the entire production line. The data sampling interval is divided into three levels according to the importance of the equipment: 1 second, 10 seconds, and 60 seconds. The core parameters of critical equipment are sampled at a high frequency of 1 second to ensure that no details of the fault evolution are lost. The playback interval is set by default to look back 30 minutes before the warning trigger time and extend forward 5 minutes, fully covering the entire process of the fault gradually deteriorating from the normal state to triggering the warning. During the playback, the operating parameters, load rate, vibration characteristics, temperature curves, and other multi-dimensional data of the equipment are displayed simultaneously. All data are linked in real time with the corresponding components in the virtual 3D model. The component color will gradually change from normal green to yellow and then red according to the degree of parameter anomaly, intuitively presenting the development process of the fault.

[0069] When tracing key nodes of fault evolution along the timeline, a sliding window extreme value detection algorithm is used to compare parameter fluctuation amplitudes frame by frame within the sliding window. This identifies the moment when the abnormal state first exceeds the normal threshold, i.e., the fault initiation point. The inflection point where the parameter degradation rate suddenly accelerates, and the trigger point where the parameter reaches the warning threshold, are then located. These three nodes together constitute the key time nodes of fault evolution. In the example, the bearing fault initiation point is 22 minutes before the warning is triggered, at which point the vibration characteristic frequency amplitude first exceeds the normal upper limit by 0.02g; the degradation inflection point is 8 minutes before the warning is triggered, at which point the bearing temperature rise rate increases from 0.3℃ per 10 minutes to 1.2℃ per 10 minutes; the trigger point is the warning timestamp. The times, corresponding parameter values, and parameter change rates of the three key nodes are concatenated in chronological order and superimposed on the timeline of the digital twin to generate a complete fault backtracking trajectory diagram. The trajectory diagram uses time as the horizontal axis and health score as the vertical axis, marking the positions and corresponding characteristics of the three key nodes, intuitively presenting the fault evolution path and development speed, and providing a temporal basis for subsequent fault location and maintenance priority determination.

[0070] Based on the spatial location of the first occurrence of the abnormal state in the fault backtracking trajectory diagram, and combined with the equipment spatial coordinate mapping relationship in the digital twin model, the specific equipment component and spatial coordinates where the fault occurred are located, and the fault spatial location result is generated. The core of this step is to map the anomaly features at the time-series level onto specific physical components in a three-dimensional virtual space, thereby achieving precise spatial location of the fault, identifying the faulty equipment, component, and its precise coordinates, avoiding the time wasted on extensive troubleshooting, and improving fault handling efficiency. The specific implementation method is as follows: The fault backtracking trajectory diagram records the corresponding data acquisition sensor number and installation location when the abnormal state first occurs. Each sensor has a unique spatial coordinate mapping relationship in the digital twin model. The installation location of the sensor in physical space corresponds one-to-one with the coordinate point in virtual space, and the positioning accuracy can reach the centimeter level, ensuring that the spatial mapping is error-free. The virtual layer of the digital twin model constructs a 1:1 three-dimensional geometric model of the production line. All equipment, components, and sensors are modeled according to their actual physical dimensions and spatial positions, using a unified workshop coordinate system. The origin of the coordinate system is set at the ground reference point at the feed end of the production line. The x-axis extends along the material flow direction of the production line, the y-axis extends along the width direction of the equipment, and the z-axis extends along the direction perpendicular to the ground. The coordinate unit is uniformly set in meters, and the spatial coordinates of all components are defined based on this coordinate system.

[0071] Based on the sensor number corresponding to the abnormal data, the installation location of the corresponding sensor is located in the digital twin model. Then, combined with the fault occurrence pattern of the fault type, the fault is traced along the equipment transmission chain, heating circuit, and other related structures to pinpoint the specific equipment component where the fault occurred. In the example, the abnormal vibration data comes from the vibration sensor at the drive end of roller conveyor No. 5 in the heating zone 3 of the tempering furnace. Combining this with the fault type of bearing fatigue wear, the faulty component can be directly located as the deep groove ball bearing at the drive end of roller conveyor No. 5 in the heating zone 3 of the tempering furnace. Reading the spatial coordinates of this component in the 3D model, the precise coordinates of the fault point are obtained as x=12.84m, y=4.21m, z=1.62m, where the x-coordinate corresponds to the position along the production line direction, the y-coordinate corresponds to the lateral position of the equipment, and the z-coordinate corresponds to the height above the ground.

[0072] After location is determined, the impact range of the fault is further calculated. Combining the equipment's associated structure and production cycle time, the potential impact on the process range and product batches is assessed. In the example, the roller bearing failure affects the smoothness of glass conveying in heating zone 3, potentially causing scratches and uneven temperature on the glass surface. The impact covers the subsequent two roller conveyor stations and the three currently being manufactured glass products. Information such as the faulty equipment number, component name, precise spatial coordinates, impact range, and associated sensor numbers are integrated to generate the fault spatial location result. This result can directly highlight the faulty component in the digital twin 3D model. Maintenance personnel can intuitively see the specific location of the fault and its surrounding structure through a virtual interface, eliminating the need for on-site inspection and significantly reducing the time required for fault location.

[0073] The fault type identifier and the fault spatial location result are input into a preset expert knowledge base for matching and retrieval. The knowledge base stores a structured correspondence between fault types and maintenance solutions. After matching, the corresponding maintenance operation steps and replacement parts list are output, and a set of candidate repair solutions is generated. The core of this step is to rely on a structured expert knowledge base to automatically match fault characteristics with historically accumulated maintenance experience, quickly outputting suitable maintenance solutions and spare parts requirements. This avoids deviations caused by relying on human experience and ensures the standardization and effectiveness of the maintenance solutions. The specific implementation method is as follows: The pre-built expert knowledge base is a structured knowledge system compiled from years of operation and maintenance experience in laminated glass production lines. The knowledge base adopts a fault mode-maintenance solution mapping storage structure. Each fault type corresponds to multiple maintenance solutions with different handling levels. Each solution includes five core information categories: maintenance operation steps, required spare parts list, estimated maintenance time, required personnel configuration, and post-maintenance performance recovery level. The knowledge base is regularly updated and iterated based on newly added fault handling cases, continuously supplementing new fault modes and optimized maintenance solutions to ensure the applicability and advancement of the solutions.

[0074] The fault type identifier is used as the primary search keyword, and the fault spatial location result is used as the auxiliary search condition. These are input into the knowledge base for precise matching retrieval. The search logic adopts a rule of prioritizing precise matching and supplementing with similar matching. First, entries that are completely identical to the fault type identifier are retrieved. If a perfect match exists, the corresponding solution is directly output. If no perfect match exists, the three fault solutions with the highest similarity are selected as references through fault feature similarity calculation. In the example, the fault type identifier is bearing fatigue wear, the faulty component is the bearing at the drive end of the tempering furnace roller conveyor, and the knowledge base contains three different levels of maintenance solutions for this fault.

[0075] The first option is a complete replacement plan. The steps are as follows: shut down and disconnect power, remove the roller conveyor end cover, remove the failed bearing, clean the journal and bearing housing, press in a new bearing of the same model, add grease, reinstall the end cover, and power on for a test run. Required spare parts include one set of the corresponding model deep groove ball bearing, one tube of high-temperature grease, and two sealing gaskets. The estimated repair time is 120 minutes, requiring two maintenance personnel. After repair, the equipment performance can be fully restored to normal levels. The second option is a temporary reinforcement plan. The steps are as follows: reduce equipment speed, add high-temperature grease, adjust bearing preload, and monitor operating parameters. No spare parts need to be replaced. The estimated operation time is 15 minutes, requiring only one maintenance personnel. After repair, the equipment can maintain reduced speed operation for 8 to 12 hours. A complete replacement must be performed during a subsequent shutdown window. The third option is a preventative maintenance plan, suitable for the initial stage of a fault. The steps are: add special grease, optimize the operating load, and shorten the inspection cycle. The estimated time is 10 minutes. This plan can delay the development of the fault and extend the service life of components by more than 30%.

[0076] The three solutions are adapted to different production scenarios and fault severity levels. They are organized from highest to lowest level of repair thoroughness. Each solution retains complete information such as operation steps, spare parts list, time duration, and personnel requirements, forming a candidate repair solution set. The solutions in this set cover different needs, from emergency production maintenance to complete repair, providing ample options for selecting the optimal solution based on actual production conditions.

[0077] Based on the equipment accessibility and current production schedule in the fault spatial location results, the feasibility of each scheme in the candidate repair scheme set is ranked, the optimal scheme is selected and a fault location identifier is attached, and finally a fault solution containing fault location identifier and maintenance and handling suggestions is generated.

[0078] The core of this step is to combine the actual on-site production situation and select the maintenance plan with the best overall benefits from multiple candidate solutions. This minimizes the impact on production schedule while ensuring equipment safety, balancing maintenance effectiveness with production losses, and outputting a complete fault handling plan that can be directly implemented. The specific implementation method is as follows: Feasibility assessment comprises two core dimensions: equipment accessibility and production scheduling adaptability. Each dimension has a weighted coefficient, and a weighted sum is used to calculate the comprehensive feasibility score for each solution. A higher score indicates better overall benefits. The equipment accessibility dimension primarily assesses the operating space of the faulty component, whether a complete machine shutdown is required, whether disassembly of surrounding components is necessary, and the safety risk level of the maintenance operation. Better accessibility means more convenient maintenance operations and lower downtime losses. The weighted coefficient for this dimension is set to 0.4, with a score range of 0 to 100; higher scores indicate better accessibility. The production scheduling adaptability dimension primarily assesses the urgency of the current production order, the time remaining until the next planned downtime window, the impact of the fault on product quality, and whether the maintenance time can be incorporated into production gaps. Better adaptability means less impact of maintenance on production schedule. The weighted coefficient for this dimension is set to 0.6, with a score range of 0 to 100; higher scores indicate better scheduling adaptability.

[0079] The comprehensive feasibility score is calculated using the formula S = w_1 * S_1 + w_2 * S_2, where S represents the comprehensive score, w_1 is the equipment accessibility weight, S_1 is the equipment accessibility score, w_2 is the production schedule adaptability weight, and S_2 is the production schedule adaptability score. In the example, the current production line is executing an urgent order, with 42 minutes remaining before the next planned changeover downtime window, which is estimated to last 18 minutes. The fault is currently in the middle stage of deterioration and has not yet caused product quality defects. Scores are calculated for each of the three candidate solutions. A complete replacement solution requires a 120-minute downtime, which would interrupt the production of the urgent order. Its production schedule adaptability score is only 30 points, its equipment accessibility score is 75 points, and its comprehensive score is 0.4. 75+0.630 = 48 points. The temporary reinforcement solution only requires a brief slowdown of the equipment, without stopping the machine, and can be completed during the current production break. The scheduling adaptability score is 92 points, the equipment accessibility score is 85 points, and the overall score is 0.4. 85+0.6 92 = 89.2 points. The adaptability score of the preventive maintenance plan is 88 points, the accessibility score is 90 points, and the overall score is 88.8 points.

[0080] Based on the comprehensive scores, the temporary reinforcement solution received the highest score and was selected as the optimal solution. The location identifier from the fault spatial localization results was then added to the solution. This identifier includes the equipment name, component number, and precise coordinates, facilitating rapid fault location by maintenance personnel. The final fault solution includes six core components: fault location identifier, fault type description, recommended repair plan, detailed operating steps, spare parts requirements, estimated operation time, and follow-up suggestions. In the example, the solution clearly identifies the fault location as the bearing at the drive end of roller conveyor No. 5 in heating zone 3 of the tempering furnace. It recommends immediate temporary lubrication and reinforcement, replenishing high-temperature grease and adjusting the preload. The bearing should be completely replaced within the 42-minute changeover shutdown window. Simultaneously, the frequency of inspections should be increased within the following two hours to monitor bearing temperature and vibration data. This comprehensive solution ensures continuous production of the urgent order while allowing for a complete repair within the planned timeframe, balancing production efficiency and equipment safety, and demonstrating strong feasibility.

[0081] Another embodiment of the present invention provides an intelligent interactive system for production lines based on the integration of digital twins and the Internet of Things, see [link to relevant documentation]. Figure 3 The system may include: The acquisition module 301 is used to collect the operating status and process cycle data in real time through multi-source sensors deployed at each process node of the laminated glass production line, so as to construct a digital twin framework including a physical layer, a virtual layer, a data layer and an interactive control service layer, and to map and update the equipment geometry, operating status and process flow of the physical production line in the virtual space in real time to generate a digital twin model of virtual and real mapping. The simulation module 302 is used to train a global optimization strategy for the production process based on the digital twin model using a reinforcement learning algorithm, dynamically simulate the production line operation status using the digital twin model as an environmental interaction platform, and output the global optimization strategy after reinforcement learning iteration optimization. The search module 303 is used to simulate and deduce the temperature field distribution, pressure change curve and cycle connection relationship of each process in the production process according to the global optimization strategy. With product quality consistency as a constraint, it searches for the optimal combination of heating temperature, pressure parameters and rolling time to generate an optimized process parameter combination. The prediction module 304 is used to detect the deviation of the execution effect of the optimized process parameter combination, and at the same time, it uses the digital twin model combined with the pre-built equipment health assessment model to perform trend analysis on the operating status data of key equipment components, predict potential equipment failure hazards, and generate early warning and diagnostic information containing failure types. The generation module 305 is used to perform fault backtracking and spatial positioning using a digital twin model based on the early warning and diagnostic information, and to combine a preset expert knowledge base to match the repair scheme corresponding to the fault type, thereby generating a fault solution that includes fault location identification and maintenance suggestions.

[0082] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0083] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0084] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0085] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for intelligent interaction in production lines based on the integration of digital twins and the Internet of Things, characterized in that, The method includes: By deploying multi-source sensors at each process node of the laminated glass production line, the operating status and process cycle data of the production process are collected in real time to build a digital twin framework that includes a physical layer, a virtual layer, a data layer and an interactive control service layer. The equipment geometry, operating status and process flow of the physical production line are mapped and updated in real time in the virtual space to generate a digital twin model that maps the virtual and real. Based on the digital twin model, a reinforcement learning algorithm is used to train a global optimization strategy for the production process. The digital twin model is used as an environment interaction platform to dynamically simulate the production line operation status and output the global optimization strategy after reinforcement learning iteration. Based on the global optimization strategy, the temperature field distribution, pressure change curve and cycle connection relationship of each process in the production process are simulated and deduced. With product quality consistency as a constraint, the optimal combination of heating temperature, pressure parameters and rolling time is searched to generate the optimized process parameter combination. The optimized process parameter combination is subjected to performance deviation detection. At the same time, the digital twin model is combined with the pre-built equipment health assessment model to perform trend analysis on the operating status data of key equipment components, predict potential equipment failure hazards, and generate early warning and diagnostic information containing failure types. Based on the early warning and diagnostic information, a digital twin model is used for fault tracing and spatial positioning. Combined with a pre-set expert knowledge base, a repair plan corresponding to the fault type is matched to generate a fault solution that includes fault location identification and maintenance suggestions.

2. The method according to claim 1, characterized in that, The process utilizes multi-source sensors deployed at various process nodes of the laminated glass production line to collect real-time data on the operating status and cycle time of the production process. This data is used to construct a digital twin framework comprising a physical layer, a virtual layer, a data layer, and an interactive control service layer. The physical production line's equipment geometry, operating status, and technological flow are then mapped and synchronously updated in virtual space in real time, generating a digital twin model that maps the physical and virtual components. By deploying various types of sensing devices at each process node of the production line, real-time data collection of equipment operating parameters, product position signals and process completion status signals is generated to produce multi-source sensor raw datasets. The edge-side preprocessing of the original dataset from multi-source sensing is performed, the filtering algorithm is used to remove signal noise, and the signals from different sensing devices are uniformly converted into standard format data frames through protocol parsing to generate a standardized sensing data frame sequence. Based on standardized sensor data frame sequences, the operating status features of each process equipment, including start / stop status and load rate, are extracted, and the process cycle time features, including product processing time and waiting time, are extracted to generate a set of equipment status and process cycle time features. Based on the equipment status and process cycle characteristics, a digital twin framework including a physical layer, a virtual layer, a data layer, and an interactive control service layer is constructed in the virtual space. The equipment geometry, operating status, and process flow of the physical production line are mapped and updated synchronously in the virtual space in real time, generating a digital twin model that maps the virtual and real worlds.

3. The method according to claim 2, characterized in that, Based on the digital twin model, a reinforcement learning algorithm is used to train a global optimization strategy for the production process. The digital twin model serves as an interactive platform for dynamically simulating the production line's operational status, and the resulting global optimization strategy, iteratively optimized through reinforcement learning, is output, including: The current production line operating status is extracted from the digital twin model as the initial observation. The state space of reinforcement learning is defined as the load rate, cycle time deviation and buffer accumulation of each process equipment, and the action space is the adjustment of the operating speed of each equipment. Reinforcement learning environment configuration parameters are generated. Using a digital twin model as an environment interaction platform for reinforcement learning, the current observation state is input into the policy network, the policy network outputs action decisions and applies them to the virtual devices in the digital twin model, driving the digital twin model to perform dynamic inference, and generating the next state and production line output results after the inference. The comprehensive reward value is calculated based on the production line output results after simulation. The reward function includes a weighted sum of three dimensions: product output, equipment utilization rate, and energy consumption. The reward value is then used to update the weight parameters of the policy network through the near-end policy optimization algorithm to generate the updated policy network. Repeat the iterative process of state observation, action decision-making, deduction, reward calculation and network update until the reward value converges or reaches the preset number of iterations, and finally output the global optimization strategy after reinforcement learning iteration optimization.

4. The method according to claim 3, characterized in that, The process involves simulating and extrapolating the temperature field distribution, pressure change curves, and cycle time relationships of each step in the production process based on the global optimization strategy. Using product quality consistency as a constraint, the optimal combination of heating temperature, pressure parameters, and rolling time is searched to generate an optimized combination of process parameters, including: The target operating rate and temperature setting range of each process are extracted from the global optimization strategy, and a virtual process simulation environment including the heating section of the tempering furnace, the lamination chamber and the autoclave are constructed in the digital twin model. In a virtual process simulation environment, the temperature field distribution of each heating zone of the tempering furnace is simulated and calculated, the pressure change curve of the autoclave is dynamically simulated, and the timing relationship between the laminator and the preceding and following processes is deduced to generate a multiphysics simulation result set. With product quality consistency indicators including at least the bonding strength of laminated glass, bubble residual rate and optical distortion rate as constraints, a Bayesian optimization algorithm is used to search for the optimal combination of heating temperature distribution, pressure curve parameters and rolling time in the multiphysics simulation result set to generate a candidate process parameter combination set. The product quality index is scored for each combination in the candidate process parameter combination set, and the combination with the highest score is selected as the optimal solution, thus generating the optimized process parameter combination.

5. The method according to claim 4, characterized in that, The optimized process parameter combination is subjected to performance deviation detection. Simultaneously, the digital twin model, combined with a pre-built equipment health assessment model, is used to perform trend analysis on the operating status data of key equipment components, predict potential equipment failures, and generate early warning and diagnostic information containing failure types, including: The optimized combination of process parameters is sent to the digital twin model for virtual execution. The predicted product quality and equipment response values ​​after virtual execution are collected and the deviation is calculated with the theoretical target value of the combination of process parameters to generate the execution effect deviation detection results. Extract time-series data of the operating status of key equipment components from the digital twin model, including bearing temperature sequence, motor current RMS value sequence and vibration characteristic frequency sequence, to generate a dataset of the operating status of key components; The key component operating status dataset is input into a pre-built equipment health assessment model. This model uses the isolated forest algorithm to detect anomalies in the status data, calculates the health score of each component, and generates component health trend analysis results. Based on the component health trend analysis results, an early warning is triggered when the health score is lower than the preset threshold. A fault mode matching algorithm is used to identify potential fault types, and finally, early warning diagnostic information containing fault types is generated.

6. The method according to claim 5, characterized in that, Based on the early warning and diagnostic information, a digital twin model is used for fault tracing and spatial positioning. Combined with a pre-set expert knowledge base, a repair solution corresponding to the fault type is matched to generate a fault solution containing fault location identification and maintenance suggestions, including: The fault type identifier and fault occurrence timestamp in the early warning and diagnosis information are analyzed. The time series data of equipment operation status before and after the fault occurs are replayed in the digital twin model. The key nodes of the fault evolution are traced along the time axis to generate a fault backtracking trajectory diagram. Based on the spatial location of the first occurrence of the abnormal state in the fault backtracking trajectory diagram, and combined with the equipment spatial coordinate mapping relationship in the digital twin model, the specific equipment component and spatial coordinates where the fault occurred are located, and the fault spatial location result is generated. The fault type identifier and the fault spatial location result are input into a preset expert knowledge base for matching and retrieval. The knowledge base stores a structured correspondence between fault types and repair solutions. After matching, the corresponding repair operation steps and replacement parts list are output, and a set of candidate repair solutions is generated. Based on the equipment accessibility and current production schedule in the fault spatial location results, the feasibility of each scheme in the candidate repair scheme set is ranked, the optimal scheme is selected and a fault location identifier is attached, and finally a fault solution containing fault location identifier and maintenance and handling suggestions is generated.

7. An intelligent interactive system for production lines based on the integration of digital twins and the Internet of Things, characterized in that, The system includes: The data acquisition module is used to collect real-time data on the operating status and cycle time of the production process through multi-source sensors deployed at each process node of the laminated glass production line. This is to build a digital twin framework that includes a physical layer, a virtual layer, a data layer, and an interactive control service layer. The module also maps and updates the equipment geometry, operating status, and process flow of the physical production line in the virtual space in real time, generating a digital twin model that maps the physical and virtual components. The simulation module is used to train a global optimization strategy for the production process based on the digital twin model using a reinforcement learning algorithm, dynamically simulate the production line operation status using the digital twin model as an environmental interaction platform, and output the global optimization strategy after reinforcement learning iteration optimization. The search module is used to simulate and deduce the temperature field distribution, pressure change curve and cycle connection relationship of each process in the production process according to the global optimization strategy. With product quality consistency as a constraint, it searches for the optimal combination of heating temperature, pressure parameters and rolling time to generate the optimized process parameter combination. The prediction module is used to detect the deviation of the execution effect of the optimized process parameter combination. At the same time, it uses the digital twin model combined with the pre-built equipment health assessment model to perform trend analysis on the operating status data of key equipment components, predict potential equipment failure hazards, and generate early warning and diagnostic information containing failure types. The generation module is used to perform fault backtracking and spatial positioning using a digital twin model based on the early warning and diagnostic information, and to combine a pre-set expert knowledge base to match the repair scheme corresponding to the fault type, thereby generating a fault solution that includes fault location identification and maintenance suggestions.

8. The system according to claim 7, characterized in that, The acquisition module is specifically used for: By deploying various types of sensing devices at each process node of the production line, real-time data collection of equipment operating parameters, product position signals and process completion status signals is generated to produce multi-source sensor raw datasets. The edge-side preprocessing of the original dataset from multi-source sensing is performed, the filtering algorithm is used to remove signal noise, and the signals from different sensing devices are uniformly converted into standard format data frames through protocol parsing to generate a standardized sensing data frame sequence. Based on standardized sensor data frame sequences, the operating status features of each process equipment, including start / stop status and load rate, are extracted, and the process cycle time features, including product processing time and waiting time, are extracted to generate a set of equipment status and process cycle time features. Based on the equipment status and process cycle characteristics, a digital twin framework including a physical layer, a virtual layer, a data layer, and an interactive control service layer is constructed in the virtual space. The equipment geometry, operating status, and process flow of the physical production line are mapped and updated synchronously in the virtual space in real time, generating a digital twin model that maps the virtual and real worlds.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-6.

Images