Emulsion paint packaging full-process digital twinning method and system

Abstract

The invention relates to the technical field of intelligent manufacturing and industrial digital twinning, in particular to a latex paint packaging full-process digital twinning method and system. The method comprises the following steps: collecting multi-source data in real time through a sensor network, and preprocessing to form a standard data stream; a coupled digital twinborn model fusing material characteristics and process parameters is constructed, and real-time prediction is realized by combining physical laws and data driving residual correction and online parameter self-adaption; constructing a multi-scale abnormal index by comparing model output with actual data, identifying an abnormal mode by adopting PCA, calculating local and overall health indexes, and predicting a trend; and generating a preliminary control strategy according to the abnormal mode and the health trend, forming closed-loop adaptive control in combination with health state weighted optimization and look-ahead adjustment, and issuing execution and feeding back an update strategy. According to the invention, the quality control qualification rate and the energy-saving efficiency of the whole process of emulsion paint packaging are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial digital twin technology, specifically to a digital twin method and system for the entire process of latex paint packaging. Background Technology

[0002] As a typical non-Newtonian fluid, latex paint presents common challenges in its packaging process, such as complex filling accuracy control and difficulty in preventing leakage risks. This type of material has significant thixotropic properties and bubble sensitivity, and its viscosity changes nonlinearly with shear rate, temperature, and settling time, making it difficult for traditional fixed threshold control strategies to adapt to batch differences in raw materials and fluctuations in production line conditions. Current packaging lines mainly rely on discrete sensor monitoring and offline parameter adjustment, which has limited ability to collaboratively analyze multi-source heterogeneous data, and the control response is subject to time constraints.

[0003] Chinese invention patent CN119692572B discloses a full-process packaging control system and method based on a digital twin model. Through the fusion of digital twin models and multi-source data, a closed-loop control system is formed, encompassing external demand perception and internal packaging process optimization. By integrating data from order systems, market feedback, production plans, and equipment status, the system solves the data silo problem in traditional packaging production, achieving real-time fusion of external demand and internal data. Employing a multi-source data weighting and cleaning fusion mechanism, and through digital twin simulation and evaluation, it ensures that the packaging solution more closely resembles the actual production environment. A multi-objective optimization algorithm accurately captures the conflicts and coupling effects between multiple objectives, improving decision-making accuracy. Through rapid deployment and a closed-loop feedback mechanism, the system monitors and adjusts the production line in real time, improving production efficiency and quality, reducing production costs, increasing efficiency and product quality, and minimizing environmental impact.

[0004] Current applications of digital twin technology in manufacturing primarily focus on mechanical structure simulation or general process monitoring, failing to adequately address the deep coupling modeling of fluid material properties and packaging process parameters. Specifically, there is a lack of systematic description of the nonlinear correction mechanism of bubble content on container stiffness and fluid resistance, and the cumulative effect of viscosity dynamic evolution on filling stability. Furthermore, conventional anomaly detection methods are limited by single-point threshold alarm mechanisms, making it difficult to identify multi-state coupled fault modes and lacking predictive control capabilities based on health status evolution. Therefore, there is an urgent need to develop an intelligent control method that integrates material-mechanical coupling mechanisms with data-driven optimization to achieve dynamic perception, accurate prediction, and autonomous optimization throughout the entire latex paint packaging process. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a digital twin method and system for the entire process of latex paint packaging.

[0006] The technical solution of this invention: a digital twin method for the entire process of latex paint packaging, comprising the following specific implementation steps: S1. In the entire process of latex paint packaging, multi-source heterogeneous sensing data is collected in real time through a sensor network, and the collected multi-source heterogeneous sensing data is preprocessed in a unified manner to form a standard data stream. S2. Based on standard data flow, construct a coupled digital twin model that integrates the material properties and process parameters of latex paint. This model jointly models the process domain and mechanical domain, and combines a data-driven residual correction mechanism and an online parameter adaptive estimation strategy to make real-time predictions of the system state. S3. Compare the predicted sequence of the coupled digital twin model with the actual observation data, construct a multi-scale anomaly index that takes into account both transient fluctuations and long-term drift through residual analysis, extract cross-state anomaly patterns using principal component analysis, dynamically calculate the local health index of key states and the overall health index of the packaging line, and then use an autoregressive model to predict the health trend. S4. Based on the identified abnormal patterns, current local and overall health indices and their predicted trends, specific abnormal patterns are mapped to preliminary process adjustment strategies for filling speed and sealing pressure. The control amplitude is weighted and optimized in combination with the health status, and adjustments based on health trend prediction are introduced. The final control command is generated and sent to the execution unit. At the same time, the strategy parameters are updated in real time based on the control effect feedback.

[0007] Preferably, the construction steps of the coupled digital twin model specifically include: Define a state vector that includes at least the apparent viscosity, structure factor, bubble content, volume of latex paint in the container, dynamic pressure at the nozzle end, and temperature data. Define a control vector that includes at least filling speed, sealing pressure, and labeling position; A physical skeleton model is constructed, which specifically describes the volumetric flow rate and flow-pressure relationship determined by the nozzle pressure difference and equivalent flow resistance, as well as the container pressure-volume relationship determined by the container equivalent stiffness parameter and the volume change of the latex paint. A material structure dynamics model is established to simulate the nonlinear variation of the apparent viscosity of latex paint with local shear rate and temperature, and to describe the dynamic evolution of structural factors under shear failure and restorative effects.

[0008] Preferably, the data-driven residual correction is specifically implemented as follows: Calculate the residuals between the state predictions obtained from the physical skeleton model and the material structure dynamics model and the actual observation data; A lightweight data-driven model is used to learn and map the residuals to obtain the residual correction value; The residual correction value is added to the physical model prediction value to generate the final system state prediction value.

[0009] Preferably, the online parameter adaptive estimation strategy adopts a joint state and parameter estimation method. By introducing process noise and observation noise, and using the Kalman gain matrix to update the state vector and key model parameters in real time, it can adapt to the differences in raw material batches and the fluctuations in production line conditions.

[0010] Preferably, the construction process of multi-scale anomaly indicators specifically includes: For each key state variable, calculate its residual moving average at three different time scales: short-term, medium-term, and long-term. By assigning different weights to the moving averages of the residuals at three time scales and performing a weighted summation, a comprehensive anomaly index reflecting the instantaneous anomalies and trend drift of the state variable is obtained.

[0011] Preferably, the extraction of cross-state anomalous patterns using principal component analysis specifically includes: Within a set time window, collect the residuals of each key state quantity to construct a short-term residual matrix; Calculate the covariance matrix of the short-term residual matrix, and perform principal component analysis on the covariance matrix to extract eigenvectors; The residuals of each state variable are projected onto the extracted main feature vectors, and the mode anomaly index representing different types of coupled faults is obtained by weighted summation, thus completing the interpretable diagnosis of multi-state coupled anomalies.

[0012] Preferably, the process of constructing the overall health index specifically includes: For each key state variable, its anomaly index, state sensitivity coefficient calibrated by historical data, and contribution weight of relevant anomaly patterns are combined to calculate a local health index that reflects the health status of the state variable. The overall health index, which characterizes the overall operating status of the packaging line, is calculated by weighting and summing the local health indices of each key state quantity according to their contribution to the overall packaging quality.

[0013] Preferably, using an autoregressive model to predict health trends specifically involves: Based on the historical data sequences of the local and overall health indices, an autoregressive prediction model is established to predict changes in health status at specific future points in time, thereby providing early warning of potential risks.

[0014] Preferably, the control amplitude is weighted and optimized based on health status, specifically as follows: The adjustment amount of each control parameter in the initial control strategy vector is associated with the local health index of the corresponding key state quantity. The weighted calculation is performed through the weight vector so that the control parameters corresponding to the key state with the worse health status receive greater adjustment intervention.

[0015] The technical solution of this invention: A digital twin system for the entire process of latex paint packaging, used to execute the above-mentioned digital twin method for the entire process of latex paint packaging, comprising: The multi-source sensing data acquisition and preprocessing module is configured to acquire multi-source heterogeneous data in real time through a sensor network deployed at key workstations of raw material tanks, filling devices, sealing machines, and labeling equipment, and to perform time synchronization, noise filtering, and data compression preprocessing to form a standard data stream. The coupled digital twin model construction and prediction control module is connected to the multi-source sensing data acquisition and preprocessing module. It is configured to build and maintain a dynamic digital twin model that integrates material properties and process parameters, and executes residual correction and online parameter adaptation based on a combination of physical model and data-driven approach to achieve real-time high-precision prediction of system status. The multi-scale anomaly detection and health prediction module is connected and coupled with the digital twin model construction and prediction control module. It is configured to calculate multi-scale anomaly indicators, identify anomaly patterns through principal component analysis, and dynamically calculate and predict local and overall health indices to achieve fault diagnosis and risk warning. The closed-loop adaptive control and strategy optimization module is connected to the multi-scale anomaly detection and health prediction module. It is configured to generate and issue optimized control commands to the execution unit based on anomaly patterns, health indices and trends. Based on feedback, it performs strategy self-learning and continuous optimization, forming an integrated closed-loop control system of perception-modeling-diagnosis-regulation.

[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs a digital twin method and system for the entire latex paint packaging process. By constructing a digital twin system for the entire latex paint packaging process, the stability of packaging quality and the reliability of equipment are significantly improved. Its core lies in integrating multi-source real-time sensing data with a material-mechanical coupling model. Utilizing a residual correction mechanism that combines physical and data-driven approaches, it accurately describes the impact of dynamic characteristics of latex paint, such as thixotropic viscosity and bubble content, on the filling process, solving problems such as uneven filling and leakage risks caused by nonlinear fluid behavior. Furthermore, through multi-scale anomaly detection technology, combined with short-term transient fluctuations and long-term drift analysis, and based on principal component analysis to extract cross-state anomaly patterns, it achieves the detection of nozzle blockage, bubble accumulation, and viscosity issues. The system provides interpretable diagnosis of complex faults such as degree drift; it quantifies anomalies into local and overall health indices, predicts health trends through autoregressive models, triggers early warnings, and generates proactive control strategies; based on anomaly pattern mapping, health status weighting, and trend prediction, it dynamically optimizes process parameters such as filling speed and sealing pressure, and continuously learns through feedback mechanisms to form an integrated closed loop of "perception-modeling-diagnosis-control"; the system adaptively responds to batch differences and operating condition fluctuations, comprehensively optimizes packaging yield and energy efficiency, and reduces equipment maintenance costs, providing intelligent solutions for packaging of highly complex fluids such as latex paint. Attached Figure Description

[0017] Figure 1 This is a flowchart of a digital twin method for the entire process of latex paint packaging proposed in this invention; Figure 2 This is a system architecture diagram of a digital twin system for the entire process of latex paint packaging proposed in this invention. Detailed Implementation

[0018] Example 1, as Figure 1 As shown, the present invention proposes a digital twin method for the entire process of latex paint packaging, which includes the following specific implementation steps: S1. During the latex paint packaging process, a sensor network deployed on raw material tanks, filling devices, sealing machines, and labeling equipment is used to collect multi-source sensing data in real time, including but not limited to the viscosity, fluidity, temperature, and bubble content of the latex paint, as well as the force, deformation, and movement trajectory of the packaging container. The collected heterogeneous data undergoes unified preprocessing, including time synchronization, noise filtering, and data compression, to form a basic input stream y(t) with consistent timing and physical semantics.

[0019] S2. A coupled digital twin model of latex paint material properties and process parameters is constructed using multi-source sensing data. The process domain and mechanical domain are jointly modeled, and data-driven residual correction and online parameter adaptation are combined to perform real-time prediction and control closed-loop optimization, ensuring stable filling quality under different batches and operating conditions. The specific implementation process is as follows: S21. By defining a unified state vector and control vector, material, mechanical, and process parameter information are fused to form a model input / output interface, specifically: Define the state vector x(t): ; Define the control vector u(t): ; in, S(t) represents the apparent viscosity of the latex paint, reflecting the flow resistance of the material at the current shear rate; S(t) represents the structure factor (0~1), characterizing the integrity of the microstructure within the latex paint, a dynamic index of thixotropic destruction and recovery; B(t) represents the bubble content (volume fraction), characterizing the proportion of bubbles in the liquid, affecting flow and pressure response; V b (t) represents the volume of latex paint (m³) inside the container, used to calculate pressure changes and remaining quantity; P noz (t) represents the nozzle end dynamic pressure (Pa), reflecting the filling flow pressure; T(t) represents the temperature (°C), the ambient and latex paint temperature, which affects viscosity and chemical properties; u speed (t) represents the filling speed, controlling the filling pump speed; u press (t) represents the sealing pressure, controlling the sealing strength of the lid; u label (t) indicates the labeling position and controls the position of the automatic labeling device; S22. Construct a physical framework model to describe the flow-pressure difference and container pressure-volume relationships, ensuring that the material-mechanical coupling is solvable. Specifically: Define the flow rate-pressure difference relationship: ; ; Define the pressure-volume relationship of the container: ; ; Where Q(t) represents the volumetric flow rate; Indicates the nozzle pressure difference, the pressure difference that drives the flow of latex paint; R eff (t) represents the equivalent flow resistance, calculated by combining the flow channel geometry and fluid characteristics; Indicates geometric coefficients; This represents the equivalent viscosity, calculated by combining the thixotropic structure factor and temperature. P represents a non-Newtonian exponent; b (t) represents the pressure inside the container; P env(t) represents environmental pressure; K represents the corrected equivalent stiffness parameter of the container. b This indicates the inherent elastic stiffness of the container itself; This represents the bubble sensitivity coefficient, used to correct for changes in container stiffness as the proportion of bubbles changes. V0 represents the normalized function that varies with the bubble content B; V0 represents the reference volume, the initial or empty volume of the bucket; Vb(t) represents the liquid volume inside the latex paint bucket or container at time t, which is a key state quantity describing the filling progress and internal pressure response. S23. Establish a material structure dynamics model to simulate the change of thixotropic viscosity with shear and temperature: Describe the dynamic changes in viscosity of latex paint under shear or static conditions: ; Perform apparent viscosity coupling: ; in, θ represents the local shear rate; α represents the structural failure coefficient, describing the rate of structural failure under shear; m represents the shear sensitivity index, reflecting the nonlinearity of failure with shear rate; β represents the structural recovery coefficient, describing the microstructure recovery rate under static or low shear conditions; θ and T0 represent temperature-sensitive parameters used to adjust the structural recovery rate as a function of temperature. Indicates the reference viscosity, temperature-dependent initial viscosity; c S S represents the coefficient of influence of the structure factor on viscosity, describing the amplification effect of S on apparent viscosity; n represents the rheological index, describing the shear dependence characteristics. S24. Generate a sensitivity matrix through offline high-fidelity simulation, and correct the parameters of the online reduced-order model, specifically as follows: Offline generation of sensitivity matrix S using high-fidelity CFD and multiphysics simulation hf The online order reduction model uses its correction parameters: ; Among them, S hf Represents the high-fidelity sensitivity matrix, offline CFD and multiphysics simulation results, and input state / control mapping to parameter correction; This represents the parameter vector of the online model, used to correct the reduced-order model; This represents the offline basic parameters used for initial model calculations; W represents the weight matrix used to map sensitivity to parameter updates. S25. Utilize data-driven models to correct physical model residuals and improve prediction accuracy, specifically: The physical skeleton model established in steps S22-S24 is used to predict the state of the latex paint, and the predicted physical skeleton value is output. ; The actual observed data y(t) is compared with the predicted value, and the residual is calculated. : ; Lightweight data-driven models (such as neural networks, regression models, or recursive filters) are used to learn and map the residuals, and the residual correction values ​​are added to the physical prediction to generate the final prediction. ; S26. The status and key parameters are updated in real time using an online parameter adaptive method, specifically as follows: Joint estimation of state and parameters: ; ; in, This represents the predicted state of the i-th sample; Represents the prediction function of the physical model; K represents process noise, used to represent model uncertainty. t H represents the Kalman gain matrix, used for state and parameter updates; H represents the observation matrix, which maps the state to the observation space. This represents the updated state and parameter estimates; It should be noted that the physical model prediction function It is used to predict the system state at the next moment based on the current state vector and control vector. Essentially, it maps the coupling relationship between material properties, process parameters, and mechanical response to the future state through physical equations; specifically, It combines the fluid dynamics equations of latex paint, the pressure-volume relationship of containers, and the dynamics model of thixotropic structures. It can calculate the changes in apparent viscosity with shear rate and temperature, the evolution of structural factors with shear failure and resting recovery, and the correction of flow resistance and pressure by bubble content. At the same time, it couples control inputs such as filling speed and nozzle pressure to achieve unified prediction of process domain and mechanical domain, update each component in the state vector, output the predicted state at the next moment, and provide input for residual correction and online adaptation. S27. Conduct sensitivity analysis to guide sensor weighting and sampling strategies, namely: assess the impact of key parameters on product quality indicators through linearized Jacobian, and predict the sequence. The input is fed into a simplified MPC to generate predictive control mapping, achieving soft real-time constraints (edge ​​computing <50ms) to ensure closed-loop linkage between the digital twin and the actual filling process; and periodic adaptive calibration is performed, uploading historical batch data and high-fidelity simulation results every N batches or event triggers, and refitting. The residual model is then used to update the parameters to the edge nodes.

[0020] S3. By analyzing the predicted sequence output in step S2 By performing residual analysis on actual multi-source sensing data, multi-scale anomaly indicators are constructed. Pattern recognition is then performed by combining inter-state correlations to achieve real-time anomaly detection, dynamic evolution of health indices, and future trend prediction in the latex paint packaging process. Ultimately, the detection results are mapped to control recommendations to form a closed-loop regulation, improving packaging quality and equipment reliability. The specific implementation process is as follows: S31. By calculating the residuals between the predicted sequence and the observed data, filtering and denoising, and quantifying the bias, a basis for anomaly analysis is provided, specifically: Quantify the deviation between actual observations and the predicted sequence in step S2: ; Dynamic window length adaptive filtering is used to ensure smooth signal under different flow rates and filling speeds; Where r(t) represents the residual vector, i.e. the deviation between the actual and the predicted values; This represents the predicted sequence output in step S2; S32. Construct short-term, medium-term, and long-term weighted anomaly indicators, taking into account both transient fluctuations and long-term drift, to achieve multi-scale anomaly detection. Specifically: ; Among them, A i (t) represents the anomaly index of state i; , and These represent the weights for short-term, medium-term, and long-term time scales, respectively. and This indicates the length of the medium- to long-term sliding window, used to capture slowly changing trends; Let represent the residual of the i-th state at time k; S33. Construct a short-term residual matrix and calculate the inter-state covariance. Extract anomalous patterns through principal component analysis, quantify multi-state coupled anomalies into pattern anomaly indices, and complete cross-state anomaly identification and interpretable diagnosis. Specifically: Define window length Collect short-term residuals: ; Where R(t) represents the short-term residual matrix, the rows represent time steps, and the columns represent state variables (such as flow rate, pressure, viscosity, and bubble content). Calculate the residual covariance matrix : ; Using Principal Component Analysis (PCA) to analyze the covariance matrix Perform feature decomposition to extract abnormal patterns; Define the abnormal pattern index: ; Among them, M j(t) represents the anomaly index of pattern j; u ij (t) represents the weight of state i in pattern j in the PCA feature vector; Indicates the number of patterns extracted; S34. The accumulated anomalies are transformed into local and overall health indices to dynamically quantify the equipment and process status. A local health index is constructed, which reflects the evolution of health status over time through weighted cumulative anomalies. The overall health index is then calculated by combining the status weights, resulting in a local-to-overall two-layer health assessment. Specifically: For each critical state i (such as flow rate, pressure, viscosity, bubble content, etc.), a health index H is defined. i (t): ; in, This represents the state sensitivity coefficient, calibrated using historical batch data, reflecting the sensitivity of the state to the overall packaging quality. The UM represents the weight of pattern j's contribution to state i, reflecting the impact of anomalous patterns on local states; i Represents the set of abnormal patterns associated with state i; This indicates the cumulative impact of abnormalities, reflecting trend risks; Define the overall health index of the packaging line : ; in, The state weights represent the proportion of each state's contribution to the overall packaging quality, and are determined through regression analysis of historical data. Indicates the number of critical states; S35. Utilizing autoregressive models combined with physical residual correction to predict future health status and abnormal trends, and through visualization of local and overall health trends, early warnings and control strategies are triggered in advance, achieving interpretable and quantifiable predictive anomaly prevention and control, and digital twin closed-loop regulation, specifically: Local health index H i (t) Establish autoregressive prediction: ; ; in, Indicates prediction of future time Local health index; wh j The weights for historical changes are determined through a sliding window fitting; p represents the autoregression order, which determines the prediction time span. This indicates the historical change in health index; Perform overall health index predict: .

[0021] S4. Based on the abnormal patterns, local and overall health indices, and abnormal trends output in step S3, closed-loop adaptive control and strategy optimization are performed on the latex paint packaging process. Through abnormal pattern-driven, health index-weighted, and trend-predictive methods, an executable control strategy is generated and adjusted in real time, achieving predictive, preventative, and optimizing digital twin closed-loop control. The specific implementation process is as follows: S41. Using the extracted abnormal patterns and local health indices, a preliminary process control strategy is generated, mapping different abnormal patterns to corresponding process parameter adjustments. That is, defining a corresponding process control mapping for each abnormal pattern and generating a preliminary control vector u0. For example, nozzle clogging mode → reduce filling speed, adjust nozzle pressure, clean nozzle; bubble accumulation mode → adjust mixer speed, extend settling time, adjust temperature control parameters; viscosity drift mode → adjust raw material ratio or temperature control parameters. S42. Based on the initial control strategy and combined with local and overall health indices, adaptively weighted optimization of the control amplitude is performed to prioritize intervention in critical states, namely: Define weighted optimization: ; ; in, H(t) represents the optimized control strategy vector; H(t) represents the local health index vector. Indicates element-wise multiplication; w H This represents a health weight vector, set based on the importance of the current state or historical sensitivity. S43. Utilize predicted local and overall health trends to proactively adjust optimized control strategies, achieving predictive anomaly prevention and early intervention for potential risks; that is: adjust control strategies based on predicted trends. ; in, Indicates the final control strategy; H ref This represents the reference value (ideal state) for the health index; k f This represents the look-ahead control gain vector; This indicates a prediction of future health indicators; S44, Final Control Strategy The system distributes data to the execution unit and collects feedback in real time. By evaluating the effectiveness of the strategy, it dynamically updates the control mapping and health weights, enabling the strategy to learn itself and continuously optimize, thus forming a complete closed-loop digital twin control system.

[0022] Example 2, as Figure 2As shown, the present invention proposes a digital twin system for the entire process of latex paint packaging, which is used to execute a digital twin method for the entire process of latex paint packaging proposed in Embodiment 1. It includes: a multi-source sensing data acquisition and preprocessing module, a coupled digital twin model construction and predictive control module, a multi-scale anomaly detection and health prediction module, and a closed-loop adaptive control and strategy optimization module.

[0023] The multi-source sensing data acquisition and preprocessing module is responsible for collecting multi-source heterogeneous data in real time, such as the viscosity, temperature, and bubble content of latex paint, as well as the stress, deformation, and equipment movement trajectory of packaging containers, through a sensor network deployed at key workstations such as raw material tanks, filling devices, sealing machines, and labeling equipment. It performs time synchronization, noise filtering, and data compression on the collected raw data to form a standard data stream with consistent timing and clear physical semantics, providing high-quality input for subsequent digital twin modeling. The coupled digital twin model construction and predictive control module is connected to the multi-source sensing data acquisition and preprocessing module. Based on the preprocessed data, a dynamic digital twin model integrating latex paint material properties and process parameters is constructed. This model integrates physical laws such as fluid mechanics, container mechanics and material thixotropic behavior, and introduces a data-driven residual correction mechanism and an online parameter adaptive estimation strategy. The multi-scale anomaly detection and health prediction module is connected and coupled with the digital twin model construction and prediction control module. By comparing the residuals of the digital twin model output and the actual observation data, it constructs short-term, medium-term and long-term multi-scale anomaly indicators. It combines principal component analysis to identify cross-state anomaly patterns, dynamically calculates the local health index of each key state and the overall health index of the packaging line, and uses an autoregressive model to predict health trends, thereby realizing real-time anomaly diagnosis, equipment status assessment and early risk warning. The closed-loop adaptive control and strategy optimization module connects to the multi-scale anomaly detection and health prediction module. Based on the anomaly pattern recognition results, the current status of the health index and its future trends, it maps specific anomaly patterns into preliminary process adjustment strategies, combines the health status to weighted optimize the control amplitude, and introduces an adjustment mechanism to generate final control commands, which are then issued to execution units such as filling, sealing, and labeling. At the same time, it updates the strategy parameters in real time based on the control effect feedback, forming a self-learning, adaptive, and continuously optimizing digital twin closed-loop control system.

[0024] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A latex paint packaging full-process digital twin method, characterized in that, The method comprises the following specific implementation steps: S1. In the whole process of latex paint packaging, real-time acquisition of multi-source heterogeneous sensing data is performed through a sensor network, and the acquired multi-source heterogeneous sensing data is uniformly preprocessed to form a standard data stream; S2. Based on the standard data stream, a coupled digital twin model integrating the material characteristics and process parameters of latex paint is constructed, the model jointly models the process domain and the mechanical domain, and combines a data-driven residual correction mechanism and an online parameter adaptive estimation strategy to perform real-time prediction of the system state; S3. The prediction sequence of the coupled digital twin model is compared with the actual observation data, a multi-scale anomaly index considering transient fluctuations and long-term drifts is constructed through residual analysis, an abnormal mode across states is extracted using principal component analysis technology, and the local health index of the key state and the overall health index of the packaging line are dynamically calculated, and then the health trend is predicted using an autoregressive model; S4. According to the identified abnormal mode, the current local and overall health index and its prediction trend, a specific abnormal mode is mapped to a preliminary process adjustment strategy for the filling speed and sealing pressure, the control range is optimized by weighting in combination with the health state, and an adjustment based on the health trend prediction is introduced, to generate a final control instruction and issue it to an execution unit, and meanwhile, the strategy parameters are updated in real time according to the control effect feedback.

2. The latex paint packaging full-process digital twin method according to claim 1, characterized in that, The construction steps of the coupled digital twin model specifically include: Defining a state vector, which at least includes apparent viscosity, structure factor, bubble content, latex paint volume in the container, nozzle end dynamic pressure, and temperature data; Defining a control vector, which at least includes filling speed, sealing pressure, and labeling position; Constructing a physical skeleton model, which specifically describes the volume flow rate and flow rate-pressure difference relationship determined by the nozzle differential pressure and equivalent flow resistance, and the container pressure-volume relationship determined by the equivalent stiffness parameter of the container and the change of the latex paint volume; Establishing a material structure dynamics model, which simulates the nonlinear change of the apparent viscosity of the latex paint with the local shear rate and temperature, and describes the dynamic evolution process of the structure factor under the action of shear destruction and static recovery.

3. The latex paint packaging full-process digital twin method according to claim 2, characterized in that, The data-driven residual correction is specifically implemented as follows: Calculating the residual between the state prediction value obtained by the physical skeleton model and the material structure dynamics model and the actual observation data; Using a lightweight data-driven model to learn and map the residual, to obtain a residual correction value; Adding the residual correction value to the physical model prediction value to generate the final system state prediction value.

4. The latex paint packaging full-process digital twin method according to claim 3, characterized in that, The online parameter adaptive estimation strategy adopts a state and parameter joint estimation method, introduces process noise and observation noise, and uses Kalman gain matrix to update the state vector and key model parameters in real time to adapt to the batch differences of raw materials and the fluctuations of production line conditions.

5. The latex paint packaging full-process digital twin method according to claim 4, characterized in that, The construction process of the multi-scale anomaly index specifically includes: For each key state variable, the residual moving average values at three different time scales, i.e., short-term, medium-term and long-term, are calculated respectively; The residual moving average values at the three time scales are given different weights and summed by weighting, to obtain a comprehensive anomaly index reflecting the instantaneous anomaly and trend drift of the state variable.

6. The latex paint packaging full-process digital twin method according to claim 5, characterized in that, The abnormal pattern across states is extracted by using principal component analysis technology, specifically including: In a set time window, the residual of each key state quantity is collected to construct a short-term residual matrix; The covariance matrix of the short-term residual matrix is calculated, and principal component analysis is performed on the covariance matrix to extract the characteristic vector; The residual of each state quantity is projected onto the extracted main characteristic vector, and the mode abnormality index representing different coupling fault types is calculated by weighted summation, completing the interpretable diagnosis of multi-state coupling abnormalities.

7. The latex paint packaging full-process digital twin method according to claim 6, characterized in that, The construction process of the overall health index specifically includes: For each key state quantity, the abnormality index, the state sensitivity coefficient calibrated by historical data, and the contribution weight of the related abnormal pattern are combined to calculate the local health index reflecting the health degree of the state quantity; The local health indexes of each key state quantity are weighted and summed according to the contribution proportion of their influence on the overall packaging quality to calculate the overall health index representing the overall running state of the packaging line.

8. The latex paint packaging full-process digital twin method according to claim 7, characterized in that, The health trend is predicted by using an autoregressive model, specifically including: Based on the historical data sequence of the local health index and the overall health index, an autoregressive prediction model is established to predict the health state change at a specific future time point, completing early warning of potential risks.

9. The latex paint packaging full-process digital twin method according to claim 8, characterized in that, The control amplitude is weighted and optimized in combination with the health state, specifically including: Each control parameter adjustment of the preliminary control strategy vector is associated with the local health index of the corresponding key state quantity, and the weighted operation is performed through the weight vector, so that the control parameter corresponding to the key state with worse health state is adjusted and intervened to a greater extent.

10. A latex paint packaging whole-process digital twin system for performing the latex paint packaging whole-process digital twin method of any one of claims 1-9, characterized in that, Including: A multi-source perception data acquisition and preprocessing module configured to collect multi-source heterogeneous data in real time through a sensor network deployed at key workstations of raw material tanks, filling devices, sealing machines, and labeling equipment, and to complete time synchronization, noise filtering, and data compression preprocessing to form a standard data stream; A coupled digital twin model construction and predictive control module connected to the multi-source perception data acquisition and preprocessing module and configured to construct and maintain a dynamic digital twin model that integrates material properties and process parameters, perform residual correction and online parameter adaptation based on a combination of physical models and data-driven methods, and realize real-time high-precision prediction of system state; A multi-scale anomaly detection and health prediction module connected to the coupled digital twin model construction and predictive control module and configured to calculate multi-scale anomaly indexes, identify abnormal patterns through principal component analysis, dynamically calculate and predict local and overall health indexes, and realize fault diagnosis and risk warning; A closed-loop adaptive control and strategy optimization module connected to the multi-scale anomaly detection and health prediction module and configured to generate and issue optimized control instructions to the execution unit based on abnormal patterns, health indexes, and trends, and perform strategy self-learning and continuous optimization based on feedback to form an integrated closed-loop control system of perception, modeling, diagnosis, and regulation.

Images