Intelligent regulation and control system for injection molding process of industrial control system

Abstract

The invention belongs to the field of artificial intelligence, particularly relates to an intelligent regulation and control system for an injection molding process of an industrial control system, and aims to solve the problem that high-precision cooperative regulation and control are difficult under material batch fluctuation, mold state change and environmental disturbance. The system comprises a multi-source sensing module, a dynamic modeling module, a self-adaptive decision-making module, an execution feedback module and a knowledge evolution module, and high-stability and high-adaptability intelligent regulation and control of the injection molding process are achieved through a mixed digital twin model integrating a physical mechanism and data driving, confidence-guided multi-objective optimization and continuous evolution of a process knowledge graph.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence, and particularly relates to an intelligent control system for an injection molding process of an industrial control system. BACKGROUND

[0002] In the field of industrial automation, injection molding is a core process for high-efficiency and mass production of plastic products, and the process control precision and stability directly determine the product quality and production cost. With the deepening of intelligent manufacturing and Industry 4.0, the industrial control system puts forward higher requirements for the real-time perception, dynamic response and adaptive adjustment capability of the injection molding process. The process involves temperature, pressure, speed, time and other multi-physical field coupling parameters, and its control effect is highly dependent on the comprehensive judgment and collaborative control capability of the system on complex working conditions.

[0003] Among them, the intelligent control of the injection molding process focuses on realizing the closed-loop optimization of key parameters in the molding cycle through data-driven and model fusion means. Its core goal is to improve equipment efficiency, shorten molding cycle and reduce scrap rate while ensuring consistent product quality. This direction usually relies on sensor networks to collect real-time data such as melt temperature, mold cavity pressure and screw position, and combines process knowledge or learning models for decision intervention.

[0004] The existing technology still has significant defects in the injection molding control: multi-source sensing data lacks effective fusion mechanism, leading to fragmented state perception; traditional PID or rule-based controllers are difficult to cope with material batch differences, mold wear and environmental disturbances and other non-steady-state factors; the control strategy is mostly static preset, which cannot be iteratively optimized online according to product quality feedback; at the same time, the system lacks early identification and active intervention capability for abnormal working conditions (such as short shots and flash). These problems are particularly prominent in high-precision and high-value product production scenarios, severely restricting the intelligent level and flexible manufacturing capability of the injection molding process. Therefore, an intelligent control system that can realize deep fusion of multi-modal data, dynamic environmental adaptive response and closed-loop quality optimization is urgently needed. SUMMARY

[0005] The purpose of the present application is to provide an intelligent control system for an injection molding process of an industrial control system to solve the technical problems that the existing injection molding control system is difficult to realize high-precision, high-stability and high-adaptability collaborative control in the face of material batch fluctuations, mold state changes and environmental disturbances and other multiple uncertain factors.

[0006] The technical scheme of the present application is an intelligent regulation and control system for an injection molding process of an industrial control system, comprising a multi-source perception module, a dynamic modeling module, an adaptive decision module, an execution feedback module, and a knowledge evolution module; the multi-source perception module is used to collect injection molding machine operating parameters, melt rheological properties, mold temperature field distribution, environmental temperature and humidity, and raw material batch identification information in real time; the dynamic modeling module is used to construct a hybrid digital twin model that combines physical mechanisms and data-driven models based on the data collected by the multi-source perception module, and to update the model parameters in real time; the adaptive decision module is used to generate optimal process parameter adjustment instructions based on the current working condition state prediction results output by the dynamic modeling module, combined with the preset process target constraints and multi-objective optimization criteria; the execution feedback module is used to issue the adjustment instructions generated by the adaptive decision module to the injection molding machine controller, and to synchronously collect the actual process response data after execution for closed-loop verification and error compensation; the knowledge evolution module is used to structure the storage and causal reasoning of successful cases and failure experiences in the historical regulation and control process, and to continuously optimize the model structure of the dynamic modeling module and the decision strategy of the adaptive decision module.

[0007] Further, the multi-source perception module includes a melt pressure-temperature composite sensor array installed at the front end of the injection molding machine cylinder, a distributed optical fiber temperature sensing unit embedded in the mold cooling water channel, an environmental monitoring node deployed in the injection molding workshop, and a batch information reading interface connected with the raw material storage system; the melt pressure-temperature composite sensor array synchronously acquires the pressure gradient and temperature distribution of the melt in the injection stage at a sampling frequency of not less than 100 times per second; the distributed optical fiber temperature sensing unit is arranged along the mold main flow channel and the cavity cooling circuit, with a spatial resolution of 5 millimeters and a temperature measurement accuracy of ±0.5℃; the environmental monitoring node acquires the workshop environmental temperature, relative humidity, and dust concentration in real time; the batch information reading interface automatically acquires the supplier code, production date, drying treatment record, and material physical property parameter nominal value of the current raw material through the industrial Internet of Things protocol.

[0008] Further, the dynamic modeling module adopts a hierarchical hybrid modeling architecture, the bottom layer is a physical mechanism sub-model based on mass conservation, momentum conservation and energy conservation equation, the middle layer is a time series bias compensation sub-model based on long short-term memory network, and the top layer is an uncertainty quantification sub-model based on Bayesian inference; the physical mechanism sub-model automatically switches the control equation according to the injection molding process stage, adopts the non-Newtonian fluid Navier-Stokes equation in the injection stage, introduces the volume shrinkage compensation term in the holding stage, and couples the heat conduction and phase change latent heat model in the cooling stage; the time series bias compensation sub-model receives the output of the physical mechanism sub-model and the actual observation value of the multi-source perception module, trains a residual mapping function to correct the systematic bias of the mechanism model caused by the simplification assumption; the uncertainty quantification sub-model dynamically evaluates the confidence interval of the model prediction result based on the current raw material batch information and the environmental disturbance intensity, and transmits the confidence index to the adaptive decision module.

[0009] Further, the adaptive decision module includes a target analysis unit, a constraint checking unit, a multi-objective optimization solver and a safety boundary checking unit; the target analysis unit converts the product quality indicators set by the user into quantifiable process objective functions, including size tolerance band center offset, warping deformation, weld line strength and surface gloss; the constraint checking unit verifies whether the candidate process parameters meet the equipment capacity boundary, the mold safety temperature upper limit and the material thermal degradation threshold in real time; the multi-objective optimization solver adopts an improved non-dominated sorting genetic algorithm to search for a Pareto optimal solution set under the condition of meeting the constraints, and dynamically adjusts the weights of each target according to the current model confidence; the safety boundary checking unit verifies the robustness of the optimization results to ensure that the product quality remains within the qualified range under ±5% parameter disturbance.

[0010] Further, the execution feedback module includes an instruction converter, an execution monitor and an error compensator; the instruction converter converts the process parameter adjustment instructions output by the adaptive decision module into a control signal format recognizable by the injection molding machine PLC; the execution monitor compares the instruction value with the actual execution value fed back by the injection molding machine in real time, and triggers an execution exception alarm when the deviation exceeds the preset threshold; the error compensator generates a feedforward compensation amount based on the correlation analysis of the execution deviation and the product quality detection results, and superimposes it into the next round of decision instructions.

[0011] Further, the knowledge evolution module is constructed with a process knowledge graph, the nodes include raw material types, mold numbers, equipment models, process parameter combinations and corresponding quality evaluation labels, and the edges represent causal relationships or similarity relationships; the knowledge evolution module automatically injects the input, decision, execution and result data of this time's regulation process into the knowledge graph and uses a graph neural network to perform relationship reasoning through an online learning mechanism; when the system identifies that the similarity between a new working condition and a historical high-confidence case exceeds 90%, the historical optimal strategy is directly called as the initial decision, greatly shortening the regulation convergence time; at the same time, the knowledge evolution module triggers an offline retraining process regularly to update the network weights of the dynamic modeling module and the time series bias compensation sub-model.

[0012] Further, the system runs in an edge-cloud collaborative computing architecture; the edge computing node is deployed locally in the injection molding workshop and is responsible for real-time processing of multi-source sensing data, millisecond-level response of dynamic modeling and adaptive decision-making; the cloud platform is responsible for global maintenance of the knowledge graph, multi-factory data aggregation analysis and model version management; the edge node and the cloud are synchronized through an industrial security gateway, and the synchronization period is every 100 production cycles or every 24 hours, whichever comes first.

[0013] Compared with the prior art, the advantages and positive effects of the present application are that: The present application effectively overcomes the dual defects of insufficient description of complex nonlinear processes by pure mechanism models and weak generalization ability of pure data models by constructing a hybrid digital twin model that combines physical mechanisms and data-driven models, significantly improving the prediction accuracy of the dynamic behavior of the injection molding process; the adaptive decision-making module introduces a multi-objective optimization mechanism based on confidence, enabling the system to automatically reduce the degree of decision-making aggressiveness in working conditions with high model uncertainty, ensuring production stability, and fully exploiting process potential in high-confidence scenarios to achieve coordinated improvement of quality and efficiency; the knowledge evolution module converts single regulation experience into reusable and transferable process knowledge assets through structured storage and causal reasoning, enabling the system to have continuous learning and self-evolution capabilities, fundamentally solving the rigid problem of traditional injection molding control systems that are "tuned once and run for a long time"; the edge-cloud collaborative architecture not only meets the stringent requirements of industrial sites for low-latency control, but also realizes knowledge sharing and model iteration across devices and factories, providing an extensible and evolvable system-level solution for injection molding intelligent manufacturing. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is the overall technical scheme architecture schematic diagram of the injection molding process intelligent regulation system of the industrial control system proposed by the present application; Figure 2 is the core principle framework schematic diagram of the hybrid digital twin model that combines physical mechanisms and data-driven models in the present application. DETAILED DESCRIPTION

[0015] Embodiment 1 The present application provides an intelligent control system for injection molding process of industrial control system, aiming to solve the technical problems that the existing injection molding control system is difficult to realize high precision, high stability and high adaptability collaborative control in the face of multiple uncertain factors such as material batch fluctuation, mold state change and environmental disturbance. The system realizes comprehensive real-time monitoring, accurate prediction, intelligent decision and continuous optimization of the injection molding process by integrating multiple source perception, dynamic modeling, adaptive decision, execution feedback and knowledge evolution and other core modules.

[0016] The working process of the system is as follows: first, the multi-source perception module is responsible for collecting all key operation data related to the injection process; then, the dynamic modeling module builds and updates a hybrid digital twin model combining physical mechanism and data-driven based on these real-time data to predict the current working condition; next, the adaptive decision module generates the optimal process parameter adjustment instruction according to the prediction result, the preset process target and the multi-objective optimization criterion; the generated instruction is issued to the injection molding machine through the execution feedback module, and the actual execution effect is collected synchronously to verify; finally, the knowledge evolution module learns and accumulates the experience of the whole control process, continuously optimizing the aforementioned modeling and decision strategy.

[0017] The multi-source perception module aims to provide comprehensive, accurate and real-time underlying data support for subsequent intelligent decision and model updating. The module is a highly integrated distributed sensing network and data interface platform, and its core responsibility is to collect injection molding machine operating parameters, melt rheological properties, mold temperature field distribution, environmental temperature and humidity, and raw material batch identification information in real time. Specifically, the multi-source perception module includes the following core components: Melt pressure-temperature composite sensor array: This array is installed at the front end of the injection molding machine barrel, close to the injection nozzle area, for direct measurement of the physical properties of the melt during the injection stage. The array is composed of multiple micro piezoresistive pressure sensors and thermocouples or resistance temperature detectors, arranged in an array to obtain the pressure gradient and temperature distribution of the melt flowing through this area. The sensors synchronously acquire these data at a sampling frequency of not less than 100 times per second, ensuring the capture of rapid dynamic changes during the injection process. Data transmission uses industrial Ethernet protocol to ensure low delay and high reliability. The collected data includes but is not limited to: injection pressure curve, holding pressure curve, melt temperature distribution along the length of the barrel, injection speed, screw position, back pressure. After preprocessing such as denoising, smoothing and unit conversion, these data form a unified data package and are transmitted to the dynamic modeling module.

[0018] Distributed Fiber-optic Temperature Sensing Unit: This unit is embedded inside the mold cooling water channels and key areas of the cavity. Fiber-optic temperature sensing technology has the advantages of strong anti-electromagnetic interference, fast response speed, and high spatial resolution. The sensing optical fiber is precisely laid along the main flow channel of the mold and the cooling circuit of the cavity, with a spatial resolution of 5 millimeters, meaning that each 5 millimeter length of optical fiber can provide an independent temperature measurement point, thus constructing a fine temperature field distribution inside the mold. The temperature measurement accuracy is ±0.5°C, meeting the stringent requirements of temperature control for injection molding. By monitoring the inlet and outlet temperatures of the cooling liquid, the mold wall surface temperature, and the core area temperature of the cavity, the thermal state of the mold can be assessed in real time, which is crucial for controlling the cooling speed of the product and preventing warping deformation.

[0019] Environmental Monitoring Nodes: These nodes are deployed at multiple representative locations within the injection molding workshop to obtain real-time environmental temperature, relative humidity, and dust concentration. Environmental temperature and humidity have a direct impact on the pre-drying effect of injection molding materials, mold dew, and product performance stability. Dust concentration indirectly reflects the cleanliness of the workshop and has some impact on the surface quality of precision injection molded products. Each node integrates high-precision temperature and humidity sensors and optical dust sensors, with a data collection period of 5 seconds, and data is aggregated through a wireless sensor network.

[0020] Batch Information Reading Interface: This interface is deeply connected with the raw material storage system, and realizes automatic data exchange through industrial Internet of Things protocols. When raw materials are fed into the hopper of the injection molding machine, the interface automatically obtains the supplier code, production date, drying treatment record, and nominal value of material physical property parameters of the current feeding raw materials. These physical property parameters, including melt flow rate, density, heat distortion temperature, and shrinkage, are key factors affecting the injection molding process and product quality. The interface uses radio frequency identification or barcode scanning technology to ensure the accuracy and automation of batch information identification. The obtained batch information is stored in a structured manner and used as a key input to the uncertainty quantification sub-model of the dynamic modeling module.

[0021] Through the coordinated work of the above components, the multi-source perception module realizes real-time and high-precision perception of all elements of the injection molding process, including "man, machine, material, method, and environment". After time stamp alignment, data cleaning, and normalization processing, all collected data form a multi-dimensional time series data set, laying a solid data foundation for subsequent intelligent analysis and decision-making. The storage of data uses a distributed time series database to ensure the storage efficiency and query response speed of massive data.

[0022] A dynamic modeling module aims to build a hybrid digital twin model that combines physical mechanisms and data-driven methods based on the data collected by the multi-source perception module, and update the model parameters in real time. The core of this module is to combine the internal mechanism of the complex physical process of injection molding with the empirical knowledge learned from data, thereby providing more accurate and robust prediction capabilities than single modeling methods. The module adopts a hierarchical hybrid modeling architecture, which specifically includes: Physical mechanism sub-model: This sub-model is at the bottom layer of the hierarchical architecture and is responsible for automatically switching the corresponding control equations according to the injection process stage to simulate the flow, heat transfer, and solidification process of the melt in the mold.

[0023] During the injection phase, the sub-model uses the Navier-Stokes equation set for non-Newtonian fluids, supplemented by the continuity equation and energy equation, to accurately describe the flow behavior of the melt under high shear rates. The viscosity of the melt changes with shear rate, temperature, and pressure, which is described by the Cross-WLF model. The equation set is solved using the finite element method, which discretizes the mold cavity into a large number of grid elements, and iteratively calculates the pressure, velocity, and temperature distribution of each grid point at each time step.

[0024] During the holding phase, as the injection is completed, the screw continues to apply pressure to compensate for material cooling shrinkage. The physical mechanism sub-model at this time introduces a volume shrinkage compensation term, which quantitatively calculates the volume change of the material during cooling by considering parameters such as cooling liquid temperature, mold temperature, mold cavity pressure, polymer material type, etc. through the P-V-T specific volume state equation, accurately simulates the volume change of the material during cooling, and predicts the shrinkage behavior of the product to avoid shrinkage and internal stress.

[0025] During the cooling phase, the sub-model couples the heat conduction and phase change latent heat models. The heat conduction equation describes the heat transfer process from the melt to the mold, while the phase change latent heat model considers the influence of the latent heat released by the semi-crystalline polymer during the crystallization process on the cooling speed. Accurate prediction of the cooling time is crucial to the dimensional accuracy and cycle time of the product.

[0026] Temporal bias compensation sub-model: This sub-model is located in the middle layer of the hierarchical architecture. It receives the output of the physical mechanism sub-model and the actual observations provided by the multi-source perception module. Since the physical mechanism model is usually based on simplifying assumptions, there may be systematic biases. The temporal bias compensation sub-model uses a long short-term memory network model to train a residual mapping function to correct these biases. The long short-term memory network can effectively handle time series data and capture the nonlinear dynamic relationship between the predicted values of the physical mechanism model and the actual observations. The model input includes the historical time of the physical mechanism model residual, the actual observation value and the current working condition parameter, and the output is the residual correction amount at the current time. Through continuous online learning, the sub-model can dynamically adapt to the systematic bias caused by changes in the production environment, ensuring the overall prediction accuracy of the digital twin model. For example, when there is a persistent bias between the actual pressure curve observed by the melt pressure-temperature composite sensor array and the pressure curve predicted by the physical mechanism sub-model, the temporal bias compensation sub-model will learn this bias pattern and compensate in the next round of prediction, so that the output of the digital twin model is closer to the true value.

[0027] Uncertainty quantification sub-model: This sub-model is located at the top layer of the hierarchical architecture. It dynamically evaluates the confidence interval of the digital twin model prediction results based on the current raw material batch information, environmental disturbance intensity and the inherent uncertainty of the model itself. The sub-model uses Bayesian inference method to combine historical data and prior knowledge to quantify the uncertainty of the model prediction. For example, when a new raw material batch is used (its physical property parameters may differ slightly from historical batches) or there is a significant change in the workshop environment, the uncertainty quantification sub-model will calculate a wider confidence interval, indicating that the uncertainty of the prediction result increases. Conversely, under stable working conditions, the confidence interval will narrow. The sub-model passes the confidence index (such as the probability of the prediction result falling within the confidence interval) to the adaptive decision-making module, so that the decision-making module can adopt a more conservative strategy in situations with high uncertainty.

[0028] The dynamic modeling module builds a hybrid digital twin model that has both physical interpretability and data-driven adaptability through the collaborative work of the above three sub-models. The model can accurately predict key state variables in the injection molding process, such as cavity pressure, melt temperature, product shrinkage and cooling time, and quantify the uncertainty of the prediction, providing a high-confidence basis for subsequent intelligent decision-making. The real-time updating mechanism of the model parameters ensures that the digital twin model always maintains high consistency with the physical injection molding machine and can effectively respond to various dynamic changes in the production process.

[0029] An adaptive decision module aims to generate optimal process parameter adjustment instructions based on the current operating condition prediction results output by the dynamic modeling module, combined with pre-set process target constraints and multi-objective optimization criteria. This module is the "brain" of the entire intelligent control system, responsible for converting complex process targets into executable parameter adjustment schemes. The module includes the following core units: Target analysis unit: This unit is responsible for converting high-level product quality indicators set by users or managers, such as "reduce warping" and "improve surface glossiness," into quantifiable process objective functions that can be used as optimization targets. These objective functions include: Dimensional tolerance band center offset: Quantifies the degree of deviation between the key dimensions of the product and the design nominal value.

[0030] Warping deformation amount: Obtain product topography data through three-dimensional scanning or optical measurement, and calculate the degree of deviation from the ideal geometric shape, which is used as an optimization target.

[0031] Weld line strength: Evaluated through mechanical test data, aiming to optimize injection and holding parameters to improve the bonding strength at the melt convergence point.

[0032] Surface glossiness: Quantified by optical measurement equipment to adjust mold temperature, injection speed, and other parameters to improve product surface quality.

[0033] These quantified objective functions convert user-friendly indicators into mathematical form through pre-set mapping rules for multi-objective optimization solvers to process.

[0034] Constraint checking unit: This unit verifies whether the candidate process parameters meet a series of pre-set hard constraints in real time. These constraints ensure the safe and stable operation of the injection molding process and the protection of equipment and materials. Constraint conditions include: Device capability boundaries: For example, the maximum injection pressure, maximum injection speed, maximum screw speed, and upper limit of clamping force of the injection molding machine.

[0035] Upper limit of mold safety temperature: Avoids damage to the mold due to overheating, affecting the service life of the mold and the quality of the product.

[0036] Material thermal degradation threshold: Prevents degradation of polymer materials due to long-term exposure to high temperatures, leading to decreased product performance or even scrap.

[0037] When the multi-objective optimization solver generates candidate parameter combinations, the constraint checking unit performs real-time verification to ensure that any parameter combination that does not meet safety or operational specifications is excluded.

[0038] Multi-objective optimization solver: This solver is the core algorithmic part of the decision module. It employs an improved non-dominated sorting genetic algorithm to search for a Pareto optimal solution set under the constraints. The non-dominated sorting genetic algorithm is an evolutionary algorithm suitable for solving optimization problems with multiple conflicting objective functions. The improvements include: introducing adaptive crossover and mutation operators to improve the convergence speed and diversity of the algorithm at different search stages; and combining the model confidence output by the dynamic modeling module to dynamically adjust the weights of each objective. For example, when the model confidence is high, the system can more aggressively pursue to improve the production efficiency or quality upper limit; while when the model confidence is low, it will tend to select conservative parameter combinations that can guarantee production stability and reduce scrap rate, for example, giving higher weights to product quality and stability objectives, and lower weights to production efficiency objective. The solver outputs a Pareto frontier, containing multiple optimal parameter combinations that trade off between different objectives.

[0039] Safety boundary checking unit: This unit further verifies the robustness of the Pareto optimal solution set generated by the multi-objective optimization solver. The purpose is to ensure that even in the presence of small perturbations in process parameters, the product quality can still be maintained within the acceptable range. The verification process is a Monte Carlo simulation of the selected optimal parameter combination: a random perturbation (e.g. ±5% parameter perturbation) within a pre-set range is introduced on each parameter (e.g. injection speed, holding pressure, mold temperature), and then the impact on product quality indicators is simulated through the digital twin model. If the simulation results show that the product quality indicators (e.g. size tolerance band, warping deformation amount) remain within the acceptable range under these perturbations, then the parameter combination is considered to have sufficient robustness. Through this verification, the system can select not only optimal but also sufficiently stable process parameter adjustment instructions.

[0040] The adaptive decision module realizes intelligent mapping and optimization from high-level quality objectives to specific process parameters through the cooperative work of the above units. The module can dynamically adapt to changes in working conditions and model uncertainty, providing optimal decision-making solutions that meet production requirements while considering robustness.

[0041] The execution feedback module aims to issue the adjustment instructions generated by the adaptive decision module to the injection molding machine controller, and simultaneously collect the actual process response data after execution, which is used for closed-loop verification and error compensation. This module is the bridge between the intelligent control system and the physical injection molding machine, ensuring the effective landing of decisions and the system's perception of execution results. The module includes the following core components: Instruction converter: This converter is responsible for converting the abstract process parameter adjustment instructions output by the adaptive decision module, such as "increase injection pressure by 10 MPa" or "raise mold temperature by 5°C", into specific control signal formats recognizable by the injection molding machine. Different injection molding machine manufacturers may use different programmable logic controller protocols or data interface standards, such as OPC UA, Modbus TCP, or Profinet. The instruction converter has multiple protocol adapters and data format mapping tables built-in, which can automatically select and generate corresponding binary instructions or register write sequences according to the connected injection molding machine model, ensuring that the instructions can be accurately parsed and executed by the injection molding machine controller. The converter is also responsible for performing safety checks on the instructions to prevent the generation of control signals that exceed the safety range of the device.

[0042] Execution monitor: This monitor compares the instruction values sent by the instruction converter with the actual execution values fed back by the injection molding machine in real time. The injection molding machine controller will feed back the actual executed injection pressure, mold temperature, screw position, etc. data to the execution feedback module through its own sensor system, such as pressure sensor, temperature sensor, position encoder. The execution monitor compares the instruction value with the feedback value at a high frequency (such as every 20 milliseconds). When the deviation between the two exceeds the preset fault tolerance threshold (such as pressure deviation exceeding 0.5 MPa, temperature deviation exceeding 0.2°C), the execution monitor will immediately trigger an execution exception alarm. The alarm information will be sent to the operator interface and recorded as a potential fault or abnormal event by the knowledge evolution module for subsequent fault diagnosis and experience learning.

[0043] Error compensator: Based on the execution deviation detected by the execution monitor and the product quality detection results obtained by the online quality detection system (such as size measurement, optical detection), the error compensator performs correlation analysis, generates feedforward compensation and superimposes it into the next round of decision instructions. For example, if the instruction requires an injection pressure of 80 MPa, but the actual feedback is only 78 MPa, and the subsequent product detection finds that the product size is too small, the error compensator will calculate a positive pressure compensation (such as 2 MPa), which will be superimposed on the injection pressure instruction generated by the adaptive decision module in the next round to offset the deficiency of the previous execution. The compensator uses reinforcement learning or adaptive filtering algorithm based on the causal relationship between historical execution deviation and product quality to learn and adjust. This feedforward compensation mechanism enables the system to correct known systematic execution errors in advance, improving the real-time and effectiveness of decision-making.

[0044] The execution feedback module forms a closed-loop control link of the intelligent injection molding control system through the close cooperation of the above components. The module not only ensures the accurate landing of intelligent decision instructions, but also continuously optimizes the control accuracy and response speed of the system through real-time monitoring and error compensation mechanism, effectively compensates for the uncertainty of the physical execution level, and makes the entire system more robust and efficient.

[0045] The knowledge evolution module aims to structure the storage and causal reasoning of successful cases and failure experiences in the historical control process, and continuously optimize the model structure of the dynamic modeling module and the decision strategy of the adaptive decision module. The module is the core of the system to realize self-learning and self-evolution, enabling the system to transform individual control experience into reusable and transferable process knowledge assets. The module is built with: Process knowledge graph: This graph is a large-scale semantic network used to structure the storage of various knowledge entities related to the injection molding process and their mutual relationships. The nodes in the graph include: raw material types (such as ABS, PC, PP), mold numbers, equipment models (such as a certain brand of injection molding machine), process parameter combinations (such as injection pressure, holding pressure, mold temperature, cooling time), and corresponding quality evaluation labels (such as size compliance, warpage out-of-tolerance, surface glossiness excellent). The edges in the graph represent the causal relationships or similarity relationships between these nodes. For example, "low injection pressure leads to short shot" is a causal relationship, and "mold A is similar in structure to mold B" is a similarity relationship. The construction of the knowledge graph uses ontology and graph database technology, which can efficiently store and query complex knowledge structures.

[0046] Online learning mechanism: The knowledge evolution module uses an online learning mechanism to automatically inject the input (multi-source perception data), decision (output of the adaptive decision module), execution (actual execution data of the execution feedback module), and result (product quality detection data) of this control process into the knowledge graph after each production cycle is completed. These new data are parsed as new nodes and edges, enriching the content of the graph. The module uses graph neural networks for relationship reasoning, such as predicting potential product quality problems under a specific process parameter combination, or discovering the implicit influence of different raw material types and mold structures on process parameters. Graph neural networks can capture complex patterns in graph structure data, automatically learn the embedding representation of nodes and edges, and thus reveal deep knowledge.

[0047] Similarity matching and historical strategy invocation: When the system identifies that the new production condition (defined by multi-source perception data) is similar to a high-confidence successful case in the historical knowledge graph with a similarity exceeding 90%, the knowledge evolution module will directly invoke the historical optimal strategy as the initial scheme for the current decision. This case-based reasoning mechanism can significantly shorten the system's convergence time when facing known or similar conditions, reducing trial and error costs. Similarity calculation uses the cosine similarity or Euclidean distance of multi-dimensional feature vectors, which are composed of key parameters of the current condition (such as raw material batch, environmental temperature, product type).

[0048] Offline retraining process: The knowledge evolution module periodically triggers an offline retraining process, such as every 1000 production cycles or monthly. This process uses all historical data accumulated in the knowledge graph to update and optimize the network weights of the time series bias compensation sub-model (long short-term memory network) in the dynamic modeling module and the multi-objective optimization solver (fitness function or weight adjustment strategy in non-dominated sorting genetic algorithm) in the adaptive decision-making module. Offline retraining can ensure that the model continuously learns the latest process knowledge and trends, improving its long-term adaptability and prediction decision accuracy.

[0049] The knowledge evolution module enables the intelligent injection molding process control system to have the ability of continuous learning and self-evolution through the above mechanisms. The system is no longer a static control logic, but an intelligent entity that can continuously grow from experience and optimize its performance, fundamentally solving the rigid problem of traditional injection molding control systems that are "tuned once and run long-term."

[0050] The system runs in an edge-cloud collaborative computing architecture. This architecture combines the low latency of edge computing and the powerful capabilities of cloud computing, providing an ideal solution for industrial control.

[0051] Edge computing node: Deployed locally in the injection molding workshop, close to the injection molding equipment. Its core responsibilities include real-time processing of multi-source perception data, dynamic modeling (especially part of the calculation of physical mechanism sub-model and time series bias compensation sub-model), and millisecond-level response of adaptive decision-making. For example, sensor data is preliminarily filtered and normalized at the edge node; the digital twin model performs rapid inference based on real-time data at the edge to predict the current condition; the constraint check and multi-objective optimization solver of the adaptive decision-making module are also mainly run at the edge node to ensure that process parameter adjustment instructions are generated within 50 milliseconds, meeting the stringent real-time requirements of industrial control. The edge node usually uses a high-performance industrial PC or an embedded system.

[0052] Cloud platform: responsible for the global maintenance of the knowledge graph, multi-factory data aggregation analysis, and model version management. The cloud platform has strong computing and storage capabilities, capable of storing millions of historical production data and process knowledge. Online learning and offline retraining of the knowledge graph, complex relationship reasoning, and other computationally intensive tasks are mainly carried out on the cloud. In addition, if there are multiple injection molding workshops or factories, the cloud platform can aggregate data from different factories, conduct macro analysis and cross-factory knowledge sharing, such as discovering the common influence rules of a certain raw material on different equipment. Model version management ensures the iterative updating and traceability of the digital twin model and decision-making strategy.

[0053] The edge node and the cloud are synchronized through an industrial security gateway. The data transmission protocol uses TLS encryption to ensure the security and privacy of industrial data. The synchronization period is set to every 100 production cycles or every 24 hours, whichever comes first. This means that once 100 products have been produced or 24 hours have passed, the edge node will synchronize the latest production data, decision logs, and knowledge graph update data to the cloud. This mechanism balances real-time performance and data synchronization efficiency, avoiding excessive network bandwidth usage from frequent synchronization.

[0054] The implementation of the edge-cloud collaborative computing architecture not only meets the stringent requirements of low-latency control in industrial sites, but also enables knowledge sharing and model iteration across devices and factories, providing an extensible and evolving system-level solution for injection molding intelligent manufacturing. Through the collaborative work of multiple modules such as multi-source perception, dynamic modeling, adaptive decision-making, execution feedback, and knowledge evolution, the system builds a closed-loop intelligent control system. Instead of relying on static process parameters, the system can accurately perceive environmental changes, accurately predict process behavior, intelligently adjust process parameters, and learn from each production experience, achieving high precision, high stability, and high adaptability in the injection molding process.

[0055] Compared with existing technologies, the present application has significant advantages and positive effects. First, by constructing a hybrid digital twin model that combines physical mechanisms and data-driven methods, the system effectively overcomes the dual defects of pure mechanism models in describing complex nonlinear processes and pure data models in generalization ability. For example, in the injection phase, the physical mechanism sub-model can accurately simulate the flow characteristics of non-Newtonian fluids, while the time deviation compensation sub-model corrects the actual deviation of the mechanism model caused by idealized assumptions through data-driven methods, significantly improving the prediction accuracy of the injection molding process dynamic behavior, such as the accuracy of predicting the cavity pressure peak and the arrival time of the melt front, which can be improved by 15% to 20%.

[0056] Secondly, the adaptive decision module introduces a multi-objective optimization mechanism based on confidence, enabling the system to automatically reduce the degree of decision-making aggressiveness in conditions with high model uncertainty, such as when the raw material batch fluctuates greatly. In such cases, the system will prioritize parameter combinations that guarantee product quality rather than extreme efficiency, thereby ensuring production stability. In high-confidence scenarios, such as using stable raw materials and a well-maintained mold state, the system fully exploits the process potential, achieving a coordinated improvement in quality and efficiency, such as reducing the production cycle by 5% to 10% while maintaining product size tolerance within ±0.02 millimeters. This dynamic adjustment strategy significantly reduces waste and improves resource utilization.

[0057] Thirdly, the knowledge evolution module converts single-time control experience into reusable and transferable process knowledge assets through structured storage and causal reasoning. For example, when the system successfully solves a product warping problem caused by excessive local mold temperature, the relevant parameter adjustment scheme and product quality improvement results are recorded in the knowledge graph. When a similar problem occurs again, the system can quickly call the historical optimal strategy as the initial decision, significantly shortening the control convergence time, such as reducing the fault resolution time from hours to minutes. This fundamentally solves the rigid problem of traditional injection molding control systems, which require "one-time parameter adjustment and long-term fixed operation," enabling continuous learning and self-evolution of the system.

[0058] Finally, the edge-cloud collaborative architecture meets the stringent requirements of low-latency control in industrial sites, such as controlling the response time of the injection process to tens of milliseconds, while achieving knowledge sharing and model iteration across devices and factories. For example, the process parameter optimization experience of new materials can be aggregated on the cloud and then distributed to all edge nodes, enabling all injection molding machines to quickly adapt to the production needs of new materials. This architecture provides an extensible and evolvable system-level solution for injection molding intelligent manufacturing, improving overall production efficiency and flexibility.

[0059] Embodiment 2 This embodiment further details the specific working principle and training process of the time sequence bias compensation sub-model in the dynamic modeling module, as well as the Bayesian inference details of the uncertainty quantification sub-model, to further improve the accuracy and robustness of the model.

[0060] The core of the time-series bias compensation sub-model lies in using a Long Short-Term Memory (LSTM) network to train and generate a residual mapping function. LSTM is a special form of recurrent neural network (RNN) that effectively captures long-term dependencies in sequential data, avoiding the vanishing or exploding gradient problems that occur in traditional RNNs when processing long sequences. In this embodiment, the LSTM network specifically employs a stacked LSTM network structure with two hidden layers, each containing 128 neurons. Furthermore, to further enhance the model's predictive ability and capture complex nonlinear patterns, a batch normalization layer is added after each LSTM network layer to accelerate model training and improve generalization ability. The network output layer is a fully connected layer used to map the hidden states of the LSTM network to residual correction values.

[0061] The training data for the temporal deviation compensation sub-model comes from the deviation between the output of the physical mechanism sub-model and the actual observations of the multi-source sensing module during historical production processes. The training process is a continuous online learning process. Whenever new production data (including the predicted values ​​of the physical mechanism sub-model and the actual observations) arrives, the system calculates the prediction residual, which is the actual observation minus the predicted value of the physical mechanism sub-model. These residuals, along with the corresponding operating parameters (e.g., injection speed, mold temperature, melt pressure, raw material batch type, ambient humidity, etc.), are organized into a time series and input into the Long Short-Term Memory (LSTM) network for training. The training objective is to minimize the mean squared error between the correction amount output by the LTM network and the actual residuals. The optimizer uses the Adam optimizer with a learning rate set to 0.001. To avoid overfitting, an early stopping strategy is introduced during training; training stops when the loss on the validation set no longer decreases within 10 consecutive cycles. Through this continuous iterative training, the LTM network can dynamically learn and correct systematic deviations in the physical mechanism model caused by simplification assumptions, parameter drift, or environmental changes.

[0062] The core of the uncertainty quantification sub-model is the dynamic evaluation of the confidence interval of the model's prediction results based on Bayesian inference. This embodiment uses Gaussian process regression to construct the uncertainty quantification sub-model. Gaussian process regression is a non-parametric Bayesian method that provides a probability distribution for the prediction results, thus naturally quantifying uncertainty. The model's input features include current operating parameters (e.g., injection pressure, mold temperature, injection speed), raw material batch information (e.g., nominal and actual fluctuation range of melt flow rate), environmental disturbance intensity (e.g., standard deviation of ambient temperature and humidity), and the prediction outputs of the physical mechanism model and the time-series deviation compensation sub-model. The model's output is the mean and variance of the digital twin model's predictions.

[0063] The training data of the Gaussian process regression model also comes from historical production data. For each historical production cycle, the system records the actual observed product quality indicators (such as size, warping) and the corresponding working condition parameters, raw material batch information, and environmental data. Through these data, the Gaussian process regression model learns the statistical relationship between the input features and the prediction error, and how the variance of the prediction error changes with the change of the input features. During the training process, the system optimizes the kernel function parameters and noise variance of the Gaussian process to maximize the log-likelihood of the training data.

[0064] When the digital twin model makes predictions for the current working conditions, the uncertainty quantification sub-model outputs a mean value of the prediction and a prediction variance associated with it based on its training results. This variance directly quantifies the uncertainty of the prediction. For example, when the raw material batch fluctuates greatly or the sensor data noise is high, the Gaussian process regression model will output a larger prediction variance, indicating that the confidence of the prediction result is low. Conversely, in stable working conditions, the variance will be smaller. This prediction variance is then converted into a confidence index (for example, the inverse of the variance or converted to a value between 0 and 1 through a specific mapping function) and passed to the adaptive decision-making module. The adaptive decision-making module will dynamically adjust the weights of each objective function in the multi-objective optimization solver based on this confidence index, taking into account the uncertainty of the prediction to ensure the robustness of the decision-making.

[0065] Through the above specific description of the time series bias compensation sub-model and the uncertainty quantification sub-model, this embodiment further improves the overall performance of the dynamic modeling module. The long short-term memory network can effectively correct systematic bias, making the prediction of the digital twin model closer to the real physical process. The Gaussian process regression provides a mathematically rigorous way to quantify the uncertainty of the prediction, enabling the adaptive decision-making module to intelligently adjust the aggressiveness of the decision-making strategy based on this uncertainty information, thereby maximizing efficiency and economic benefits while ensuring product quality and production stability. This dual optimization mechanism collectively ensures the high performance of the entire intelligent control system in complex and variable industrial environments.

[0066] Embodiment 3 This embodiment further details the details of the improved non-dominated sorting genetic algorithm of the multi-objective optimization solver in the adaptive decision-making module, as well as the robustness verification method of the safety boundary checking unit.

[0067] The multi-objective optimization solver uses an improved non-dominated sorting genetic algorithm. The non-dominated sorting genetic algorithm is an evolutionary algorithm based on the Pareto optimal concept, used to solve optimization problems with multiple conflicting objective functions. The improvement in this embodiment aims to improve the convergence speed, uniformity of solutions, and adaptability to dynamic working conditions of the algorithm.

[0068] Specifically, the improved non-dominated sorting genetic algorithm introduces the following core optimization points: Adaptive crossover and mutation operators: Traditional genetic algorithms usually adopt fixed probabilities for crossover and mutation. This improved scheme dynamically adjusts the crossover probability and mutation probability according to the distribution of individuals in the target space in the current population. For example, when the population converges slowly or lacks diversity, increase the mutation probability to explore new search space; when the population is too concentrated, reduce the crossover probability to avoid premature convergence to local optimum. The adaptive adjustment mechanism uses a fuzzy logic controller, whose inputs are the diversity index and convergence index of the population.

[0069] Strengthening of elitist strategy: During the evolution of each generation, not only non-dominated solutions are preserved, but also clustering analysis is performed on the non-dominated solution set to ensure uniform distribution of solutions on the Pareto front and preferentially preserve non-dominated solutions located in sparse areas. This helps maintain the diversity of the population and prevents the algorithm from prematurely losing potential solutions.

[0070] Dynamic adjustment of target weight combined with dynamic model confidence: As before, the solver dynamically adjusts the weights of each target based on the model confidence output by the dynamic modeling module. During the fitness evaluation stage of the genetic algorithm, the fitness of an individual is calculated based on its performance on each objective function and the corresponding target weight. When the model confidence is low, the system automatically increases the weight of targets related to production stability and product qualification rate (such as size tolerance band, warping deformation amount), while reducing the weight of targets related to production efficiency and cost (such as cycle time). This dynamic adjustment mechanism allows the algorithm to prioritize lower-risk, more robust process parameter combinations when model prediction uncertainty is high, ensuring production stability. When the model confidence is high, the system allows more aggressive pursuit of efficiency or performance improvement.

[0071] The input of the multi-objective optimization solver is the prediction results of the current working condition (including predicted values and confidence), pre-set process objective functions (such as size tolerance band center offset, warping deformation, weld line strength, surface glossiness), and device capability boundaries and material thresholds verified by the constraint checking unit. The output is a solution set containing multiple Pareto optimal process parameter combinations, each of which represents an optimal trade-off between different objectives.

[0072] The safety boundary verification unit verifies the robustness of the optimization results to ensure that the product quality remains within the acceptable range under a ±5% parameter perturbation. The robustness verification process uses a Monte Carlo simulation method based on Latin hypercube sampling.

[0073] First, one or more representative process parameter combinations are selected from the Pareto optimal solution set output by the multi-objective optimization solver as the benchmark points to be verified.

[0074] Second, for each benchmark process parameter combination, the safety margin verification unit introduces a random perturbation within a range of ±5% on each key parameter (e.g. injection speed, holding pressure, mold temperature, cooling time). The sampling of parameter perturbation employs the Latin hypercube sampling technique. Latin hypercube sampling is a stratified random sampling method that can more uniformly cover the multi-dimensional parameter space, improving simulation efficiency and the reliability of the results. For example, if the injection pressure is 100 MPa, the perturbation range is 95 MPa to 105 MPa, and multiple pressure values are randomly generated within this range.

[0075] Third, for each perturbed parameter combination, the system calls the digital twin model of the dynamic modeling module to perform a complete injection molding process simulation. The simulation output includes the predicted values of product quality indicators such as dimensional tolerance band, warpage amount, weld line strength, and surface gloss.

[0076] Finally, the system statistically analyzes all simulation results. If the product quality indicators under all perturbed samples are within the pre-set acceptable range, the benchmark process parameter combination is considered to have sufficient robustness and can be recommended for execution. Conversely, if any perturbed sample causes the product quality to be out of tolerance, the parameter combination is marked as insufficiently robust, and the system will return to the multi-objective optimization solver or suggest the operator to select a more conservative parameter combination.

[0077] By introducing the improved non-dominated sorting genetic algorithm and Latin hypercube sampling-based Monte Carlo robustness verification, this embodiment greatly improves the performance of the adaptive decision-making module. The algorithm can efficiently search for multiple high-quality Pareto optimal solutions and ensure the adaptability of the decision through confidence dynamic adjustment. Robustness verification provides additional safety for the decision scheme, ensuring that even in the presence of inevitable parameter fluctuations in the production site, the product quality can be maintained within the acceptable range, thereby significantly improving the reliability and practicality of the entire system.

[0078] Embodiment 4 This embodiment further elaborates on the specific working mechanism of the error compensator in the execution feedback module, particularly how it generates feedforward compensation based on the correlation analysis of execution bias and product quality detection results. At the same time, it will delve deeper into the details of the construction of the process knowledge graph and the reasoning of the graph neural network in the knowledge evolution module.

[0079] The error compensator learns the complex mapping relationship between execution bias and product quality based on historical data, and generates accurate feedforward compensation accordingly. Inside the error compensator, a deep neural network model is deployed, specifically a fully connected network with 4 hidden layers, each with 256 neurons and ReLU activation function. The input features of the network include: the bias vector between the instruction value and the actual execution value at the current time (such as injection pressure bias, mold temperature bias, cooling time bias), the product quality detection results of the previous production cycle (such as the absolute deviation of product size, the numerical value of warping deformation, the measured value of weld line strength), and the current working condition parameters (such as raw material type, environmental temperature). The output of the network is the feedforward compensation of the key process parameters (such as injection pressure, holding time, mold temperature) for the next round of decision instructions.

[0080] The training data of the deep neural network model comes from the millions of production records accumulated in the historical production database. Each record contains: instruction parameters, actual execution parameters, execution bias, product quality detection results, and working condition parameters. The model is periodically updated through offline batch training to minimize the mean square error between the predicted compensation and the actual required compensation. The actual required compensation is calculated in reverse based on historical data by adjusting the feedforward compensation to achieve optimal product quality in subsequent production. Distributed computing frameworks such as TensorFlow or PyTorch are used for training to handle large-scale data sets. After training, the error compensator can quickly calculate accurate feedforward compensation based on real-time execution bias and the latest quality detection results within milliseconds, and add it to the next round of decision instructions generated by the adaptive decision-making module. For example, if it is detected that the size of the last two batches of products is slightly smaller, and the execution monitor finds that the injection pressure is systematically lower than the instruction value, the error compensator will generate a positive injection pressure compensation to ensure that the next injection cycle can achieve a more accurate pressure target, thereby correcting the product size.

[0081] Details of the construction of the process knowledge graph in the knowledge evolution module and the reasoning of the graph neural network.

[0082] The construction of the process knowledge graph is a multi-stage automated process. First, the system extracts entity data from structured databases (e.g., enterprise resource planning systems, manufacturing execution systems, sensor databases), such as raw material types, mold numbers, equipment models, process parameters, product quality labels, etc., which are defined as nodes of the graph. Second, through natural language processing techniques, entities and relationships are extracted from unstructured text data (e.g., engineers' production logs, fault reports, maintenance records). For example, from the sentence "excessive injection pressure causes product flash", two entities "injection pressure" and "flash" can be extracted, as well as the causal relationship "causes". These relationships are defined as edges of the graph. At the same time, the system presets a set of ontologies to define the hierarchical relationships and attributes between different entity types, ensuring the semantic consistency of the graph. The graph is stored in a graph database, such as Neo4j, which supports efficient graph traversal and query operations.

[0083] The graph neural network plays a core role in the knowledge evolution module, which is used for relationship reasoning of the process knowledge graph. This embodiment adopts a graph convolutional network model. The input of the graph convolutional network is the adjacency matrix and the node feature matrix of the knowledge graph. The node feature matrix contains the attribute information of each entity, such as the physical property parameters of the raw material type node and the numerical range of the process parameter node. The graph convolutional network learns the high-dimensional embedding representation of the nodes through multiple stacked graph convolutional layers. In each graph convolutional layer, the embedding representation of the node is updated by aggregating the feature information of itself and its neighbor nodes. This enables the graph neural network to capture local and global information in the graph structure.

[0084] The relationship reasoning of the graph neural network is embodied in the following aspects: Causal relationship prediction: by training the graph neural network, it can predict possible product quality problems that may occur according to a given set of process parameters and working conditions, or inversely infer the process parameters that cause a specific quality problem. For example, when detecting "product surface glossiness decreases", the graph neural network can infer the possible causal relationships such as "mold temperature is too low" and "injection speed is too fast".

[0085] Similarity relationship discovery: the node embedding representation learned by the graph neural network can cluster entities such as process parameters, raw materials, and molds with similar features or behaviors in the embedding space. By calculating the distance or similarity between node embedding vectors, the system can identify historical successful cases that are highly similar to the current working condition, thereby performing case reasoning.

[0086] Knowledge completion: when there are missing entity attributes or relationships in the knowledge graph, the graph neural network can predict and complete them according to the existing graph structure and node features, such as predicting the optimal holding pressure time of a new type of raw material in a specific mold.

[0087] The online learning mechanism injects the latest production data into the graph, and periodically updates the weights of the graph neural network. When new data is added to the graph, the graph neural network is incrementally trained, enabling it to learn and adapt to new knowledge and patterns. This mechanism enables the knowledge evolution module to continuously accumulate experience and continuously optimize its reasoning ability.

[0088] Through the in-depth elaboration of the error compensator and the knowledge evolution module, the precision and intelligent level of the system are further improved. The error compensator realizes accurate feedforward compensation of execution deviation through deep learning, significantly improving the precision and response speed of control. The knowledge evolution module constructs a comprehensive process knowledge graph and uses graph neural networks for advanced reasoning, enabling the system to learn from experience, self-optimize, and efficiently use historical knowledge to guide future production decisions. These details together constitute the strong self-learning and adaptive ability of the invention, enabling it to perform outstanding performance in the complex and variable injection molding environment.

Claims

1. An intelligent control system for injection molding process of an industrial control system, characterized in that, The application relates to a real-time adaptive control system for injection molding machines. The application comprises: a multi-source perception module for collecting real-time injection molding machine operating parameters, melt rheological properties, mold temperature field distribution, environmental temperature and humidity, and raw material batch identification information; a dynamic modeling module for building a hybrid digital twin model combining physical mechanisms and data-driven models based on the data collected by the multi-source perception module, and updating the model parameters in real time; an adaptive decision-making module for generating optimal process parameter adjustment instructions based on the current working condition state prediction results output by the dynamic modeling module, combining preset process target constraints and multi-objective optimization criteria; an execution feedback module for issuing the adjustment instructions generated by the adaptive decision-making module to the injection molding machine controller, and synchronously collecting actual process response data after execution for closed-loop verification and error compensation; 2. The intelligent regulation system for injection molding process of an industrial control system according to claim 1, wherein, a knowledge evolution module for structurally storing and causally reasoning successful cases and failure experiences in historical regulation and control processes, continuously optimizing the model structure of the dynamic modeling module and the decision-making strategy of the adaptive decision-making module.

3. The intelligent regulation system for injection molding process of an industrial control system according to claim 1, wherein, The multi-source perception module comprises a melt pressure-temperature composite sensor array installed at the front end of an injection molding machine cylinder, a distributed optical fiber temperature sensing unit embedded in a mold cooling water channel, an environmental monitoring node deployed in an injection molding workshop, and a batch information reading interface connected with a raw material storage system; the melt pressure-temperature composite sensor array synchronously obtains the pressure gradient and temperature distribution of the melt in the injection stage at a sampling frequency of not less than 100 times per second; the distributed optical fiber temperature sensing unit is arranged along the cooling circuits of the mold main flow channel and the cavity, has a spatial resolution of 5 millimeters and a temperature measurement accuracy of plus or minus 0.5 DEG C; the environmental monitoring node obtains the workshop environmental temperature, relative humidity and dust concentration in real time; and the batch information reading interface automatically obtains the supplier code, production date, drying treatment record and material physical property parameter nominal value of the current feeding raw material through an industrial Internet of Things protocol. The dynamic modeling module adopts a hierarchical hybrid modeling architecture, the bottom layer is a physical mechanism submodel constructed based on mass conservation, momentum conservation and energy conservation equations, the middle layer is a time series deviation compensation submodel based on a long short-term memory network, and the top layer is an uncertainty quantification submodel based on Bayesian inference; the physical mechanism submodel automatically switches the control equation according to the injection process stage, adopts a non-Newtonian fluid Navier-Stokes equation in the injection stage, introduces a volume shrinkage compensation term in the holding stage, and couples a heat conduction and phase change latent heat model in the cooling stage; the time series deviation compensation submodel receives the output of the physical mechanism submodel and the actual observation value of the multi-source perception module, trains a residual mapping function for correcting the systematic deviation of the mechanism model caused by simplification assumptions; and the uncertainty quantification submodel dynamically evaluates the confidence interval of the model prediction results based on the current raw material batch information and environmental disturbance intensity, and transmits the confidence index to the adaptive decision-making module.

4. The intelligent regulation system for injection molding process of an industrial control system according to claim 3, wherein, The time deviation compensation sub-model adopts a stacked long short-term memory network structure containing 2 hidden layers, each of which is configured with 128 neurons, and a batch normalization layer is added after each long short-term memory network layer; the uncertainty quantification sub-model adopts a Gaussian process regression model, the input features of which include current working condition parameters, raw material batch information, environmental disturbance intensity, and the predicted outputs of the physical mechanism sub-model and the time deviation compensation sub-model, and the output is the mean and variance of the predicted value.

5. The intelligent regulation system for injection molding process of an industrial control system according to claim 1, wherein, The adaptive decision module includes a target analysis unit, a constraint checking unit, a multi-objective optimization solver, and a safety boundary checking unit; the target analysis unit converts the product quality indicators set by the user into quantitative process objective functions, including size tolerance band center offset, warping deformation, weld line strength, and surface glossiness; the constraint checking unit verifies in real time whether the candidate process parameters meet the equipment capacity boundary, the mold safety temperature upper limit, and the material thermal degradation threshold; the multi-objective optimization solver uses an improved non-dominated sorting genetic algorithm to search for a Pareto optimal solution set under the constraint condition, and dynamically adjusts the weights of each objective according to the current model confidence; the safety boundary checking unit verifies the robustness of the optimization result to ensure that the product quality remains within the qualified range under a ±5% parameter disturbance.

6. The intelligent regulation system for injection molding process of an industrial control system according to claim 5, wherein, The improved non-dominated sorting genetic algorithm introduces adaptive crossover and mutation operators, and dynamically adjusts the crossover probability and mutation probability according to the population diversity index and convergence index; the safety boundary checking unit uses a Monte Carlo simulation method based on Latin hypercube sampling to introduce a ±5% random disturbance on the key process parameters, and simulates the product quality indicators through the digital twin model to verify whether they remain within the qualified range.

7. The intelligent regulation system for injection molding process of an industrial control system according to claim 1, wherein, The execution feedback module includes an instruction converter, an execution monitor, and an error compensator; the instruction converter converts the process parameter adjustment instructions output by the adaptive decision module into a control signal format recognized by the injection molding machine PLC; the execution monitor compares the instruction value with the actual execution value fed back by the injection molding machine in real time, and triggers an execution exception alarm when the deviation exceeds the preset threshold; the error compensator generates a feedforward compensation amount based on the correlation analysis of the execution deviation and the product quality detection results, and adds it to the next round of decision instructions.

8. The intelligent regulation system for injection molding process of an industrial control system according to claim 7, wherein, The error compensator internally deploys a fully connected deep neural network containing 4 hidden layers, each of which is configured with 256 neurons and uses a ReLU activation function; the network input includes the deviation vector of the instruction value and the actual execution value, the product quality detection results of the previous production cycle, and the current working condition parameters, and the output is the feedforward compensation amount of the key process parameters of the next round of decision instructions.

9. The intelligent regulation system for injection molding process of an industrial control system according to claim 1, wherein, The knowledge evolution module is constructed with a process knowledge graph, the nodes include raw material types, mold numbers, equipment models, process parameter combinations and corresponding quality evaluation labels, and the edges represent causal relationships or similarity relationships; the knowledge evolution module automatically injects the input, decision, execution and result data of this time's regulation process into the knowledge graph after each production cycle is completed, and uses a graph neural network to perform relationship reasoning; when the system identifies that the similarity between a new working condition and a historical high-confidence case exceeds 90%, the historical optimal strategy is directly called as the initial decision; the knowledge evolution module triggers an offline retraining process regularly to update the network weights of the dynamic modeling module and the time series bias compensation sub-model.

10. The intelligent regulation system for injection molding process of an industrial control system according to claim 1, wherein, The system runs in an edge-cloud collaborative computing architecture; The edge computing node is deployed locally in the injection molding workshop and is responsible for real-time processing of multi-source sensing data, dynamic modeling and millisecond-level response of adaptive decision-making; the cloud platform is responsible for global maintenance of the knowledge graph, multi-factory data aggregation analysis and model version management; the edge node and the cloud are synchronized through an industrial security gateway, and the synchronization period is every 100 production cycles or every 24 hours, whichever comes first.

Images