Plastic film production parameter dynamic regulation and control method utilizing program process scheduling

Abstract

The invention relates to the technical field of information processing, and discloses a method for dynamically regulating and controlling production parameters of a plastic film by utilizing program process scheduling. The method comprises the following steps: converging a real-time parameter sequence to a digital twinborn body to establish virtual mapping, and constructing a multi-dimensional data model; analyzing a material state evolution trend in the digital twinborn body, synchronizing with a physical production line state, fusing historical data to predict a potential influence range, and generating a regulation and control demand report; generating an adjustment scheme draft, performing priority ranking on simulation tasks in the adjustment scheme draft by adopting a program process scheduling algorithm, dynamically allocating computing resources, executing the simulation tasks in parallel, and determining an optimization parameter set in combination with a gradient descent algorithm; and iteratively correcting the parameters in the digital twinborn body through feedback circulation to generate a final coordination control instruction, inputting the final coordination control instruction into the digital twinborn body, confirming synchronism and outputting a report. According to the invention, the precision of dynamic regulation and control of the production parameters of the plastic film is improved.

Description

Technical Field

[0001] This application relates to the field of information processing technology, and in particular to a method for dynamically controlling plastic film production parameters using program process scheduling. Background Technology

[0002] Plastic film production lines are core equipment in the materials processing field, and their operational stability directly determines product quality, energy consumption, and resource utilization efficiency. The continuous linkage of multiple stages such as extrusion, stretching, cooling, and winding causes the material state to continuously evolve with the production process. Small parameter fluctuations in previous stages can be amplified in subsequent stages, easily leading to product defects or even the scrapping of the entire roll.

[0003] Currently, thin film production control mainly relies on real-time sensor monitoring and manual adjustment based on experience. Although it can collect key parameters such as temperature, viscosity, tension, and wind speed, it has significant limitations: First, it is difficult to perceive the evolution trend of material state in advance, and the slow drift signals of parameters such as temperature and viscosity are easily masked by noise, making it impossible to predict potential risks. Second, the scope of impact is delayed after an anomaly occurs. By the time the parameters exceed the threshold alarm, the defect has already spread, increasing losses. Third, the efficiency of control scheme formulation is low. Manual trial and error is not only time-consuming, but also lacks an overall grasp of the entire production line status, which can easily lead to parameters deviating from the optimal operating range and affecting production stability.

[0004] Therefore, how to achieve early perception of material status trends, quickly define the scope of abnormal impact, and generate accurate and efficient control solutions has become a key issue in ensuring the efficient and stable operation of plastic film production lines. A dynamic control method that integrates program process scheduling and digital twin technology is urgently needed to break through the existing technical bottlenecks. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method for dynamically controlling plastic film production parameters using program process scheduling, which improves the accuracy of dynamic control of plastic film production parameters.

[0006] This application provides a method for dynamically controlling plastic film production parameters using program process scheduling, the method comprising:

[0007] Real-time parameter sequences from multiple stages of a plastic film production line are collected using a sensor network. These real-time parameter sequences are then aggregated into a digital twin to establish a virtual mapping and construct a multi-dimensional data model.

[0008] Based on the multidimensional data model and the pre-trained support vector machine model, the material state evolution trend is analyzed in the digital twin and synchronized with the physical production line state. If the material state evolution trend exceeds the preset evolution threshold, the potential impact range is predicted by integrating historical data through the digital twin, and a control demand report is generated.

[0009] Based on the aforementioned regulation demand report, virtual simulation of process parameters is performed in the digital twin to generate a draft regulation scheme. A program process scheduling algorithm is used to prioritize the simulation tasks in the draft regulation scheme, dynamically allocate computing resources to execute the simulation tasks in parallel, and combine the gradient descent algorithm to determine the set of optimization parameters.

[0010] The system detects whether the winding tension value in the optimized parameter set deviates from the stable range. If it does, the parameters are iteratively corrected in the digital twin through a feedback loop to generate a final coordinated control command. The final coordinated control command is input into the digital twin, and the material state evolution trend is simulated in combination with the verification sample data. The result is compared with the real-time state of the physical production line to confirm the synchronization and output a report, thus completing the dynamic control of plastic film production parameters.

[0011] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0012] First, a sensor network collects real-time parameters across the entire process. After median filtering for noise reduction and z-score standardization preprocessing, a digital twin virtual mapping is constructed using linear interpolation. This, combined with a multi-dimensional data model, achieves dynamic synchronization between physical and virtual states, significantly improving data transmission accuracy and mapping consistency, providing reliable data support for subsequent control. Next, a pre-trained support vector machine model is used to accurately classify material state evolution types. Kalman filtering effectively integrates historical and real-time data to quickly predict the potential impact range of anomalies. Bubble sorting clarifies control priorities, enabling early warning of potential risks and precise definition of control needs, reducing production losses caused by anomaly escalation. Following this, virtual simulation of process parameters is conducted based on the digital twin. A program process scheduling algorithm optimizes the sequencing of simulation tasks, dynamically allocates computing resources, and processes them in parallel. Combined with gradient descent, the collaborative parameters of blow molding temperature and traction speed are efficiently optimized, significantly improving the efficiency of control scheme generation and parameter optimization accuracy, and shortening control response time. Finally, a feedback loop mechanism is introduced to iteratively correct parameters for the winding tension deviation problem. The effect of parameter adjustment is accurately predicted by deep learning neural network. Finally, the synchronization of physical and virtual states is confirmed by Euclidean distance comparison, and real-time response coordination control commands are generated to effectively ensure the stable operation of the production line, reduce the product defect rate, and improve production energy efficiency and resource utilization efficiency. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of a method for dynamically controlling plastic film production parameters using program process scheduling, as described in an embodiment of this application. Figure 2 This is a schematic diagram illustrating the classification performance verification results of the support vector machine model in this application embodiment; Figure 3 This is a schematic diagram illustrating the synchronization verification results between the real-time status of the physical production line and the predicted status of the digital twin in an embodiment of this application. Detailed Implementation

[0015] This application provides a method for dynamically controlling plastic film production parameters using program process scheduling. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0016] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of a method for dynamically controlling plastic film production parameters using program process scheduling, as described in this application, includes:

[0017] Step S1: Collect real-time parameter sequences from multiple stages of the plastic film production line through a sensor network, aggregate the real-time parameter sequences into a digital twin to establish a virtual mapping, and construct a multi-dimensional data model corresponding to the physical production line.

[0018] The construction of a multidimensional data model includes:

[0019] Temperature, viscosity, tension, and air velocity data from each stage of the production line—extrusion, stretching, cooling, and winding—are collected to form a real-time parameter sequence covering the entire production process. Median filtering and z-score normalization are used to process the real-time parameter sequence. This processed sequence is then transmitted to a digital twin, where a virtual mapping relationship with the physical production line is established using linear interpolation. The processed real-time parameter sequence is aligned by timestamps to form a two-dimensional raw data matrix of "parameter dimension - time dimension." Based on this matrix, a sliding window method is used to extract historical parameter data before and after each time point, constructing a three-dimensional extended data matrix of "time window - parameter dimension." This extended data matrix is ​​then flattened to convert it into a high-dimensional feature dataset containing temporal correlation information. Principal component analysis is then applied. The analysis algorithm reduces the dimensionality of the high-dimensional feature dataset, calculates the covariance matrix, and solves for eigenvalues ​​and eigenvectors. Principal components with eigenvalues ​​greater than a preset threshold are selected as key features. Based on these key features, the time series dimension is integrated to generate a multidimensional data model. The multidimensional data model includes a core execution module. A deep learning model is built based on the key features as the core execution module. The deep learning model is trained using historical operating data from the full-process database. The gradient descent algorithm is used to iteratively optimize the model parameters until the prediction error is less than a preset error threshold. The standardized parameter data that is updated in real time is processed and input into the trained model, which dynamically outputs the production line operating status evaluation value. The model is fine-tuned online every preset period based on new historical data, realizing the dynamic reflection of the physical production line operating status by the multidimensional data model.

[0020] Specifically, specialized sensors are deployed in the four core production stages of the plastic film production line: extrusion, stretching, cooling, and winding. In the extrusion stage, temperature sensors monitor the temperature of the molten plastic material in real time, while viscosity sensors capture the material's flow characteristics, forming a basic data stream of temperature and viscosity. In the stretching stage, tension meters accurately measure the tensile force of the material during stretching, i.e., tension data, which refers to the internal interaction forces generated when the material is stretched, directly affecting the stretching and forming effect. In the cooling stage, wind speed sensors record the wind speed data of the cooling air; wind speed refers to the flow velocity of the cooling air, and its magnitude affects the cooling rate and uniformity of the material. In the winding stage, all data streams from the extrusion, stretching, and cooling stages are integrated and summarized according to a unified time sampling interval to generate a real-time parameter sequence covering all production stages with a unified data format. This real-time parameter sequence is a set of parameter data collected continuously in chronological order, dynamically reflecting the operating status of the production line at each moment, ensuring that the sequence comprehensively reflects the real-time operating status of each stage of the production line.

[0021] Median filtering is a commonly used nonlinear signal processing method. Its principle is to replace each point in the signal with the median, thereby effectively removing isolated noise points and avoiding interference from noise data in subsequent analysis. Therefore, the input data is the original real-time parameter sequence, including time-series data of temperature, viscosity, tension, and wind speed, and the output data is the denoised real-time parameter sequence. Specifically, each data point in the real-time parameter sequence and several adjacent data points form a data window, and the median of the data within this window is taken as the denoised value of the current data point. z-score standardization (standard deviation standardization) is a data standardization method. Its principle is to convert the original data into standardized data with a mean of 0 and a standard deviation of 1, eliminating the influence of differences in dimensions and numerical ranges of different parameters, facilitating subsequent comprehensive analysis of multiple parameters. Dimensions refer to the unit of measurement of the parameter, such as degrees Celsius for temperature and meters per second for wind speed. Differences in dimensions of different parameters can lead to large differences in numerical ranges. The calculation formula is: , where x is the original parameter data in the real-time parameter sequence, μ is the mean of the real-time parameter sequence, and σ is the standard deviation of the real-time parameter sequence. The standardized data corresponding to each original parameter data is calculated by this formula. That is, the input data is the denoised real-time parameter sequence, and the output data is the standardized real-time parameter sequence.

[0022] The processed real-time parameter sequence is transmitted to the digital twin server via a wireless communication module. A digital twin is a virtual model that corresponds completely to the physical production line and is constructed using digital technology. It can map the operating status, parameter changes, and material evolution process of the physical production line in real time and is the core carrier for achieving dynamic control.

[0023] Virtual mapping refers to mapping the physical structure, operating parameters, and material states of a physical production line to corresponding virtual elements in a digital twin through mathematical models and data association, forming a one-to-one correspondence to ensure that the virtual model can accurately replicate the operation of the physical production line. Linear interpolation, on the other hand, estimates the value at any point between two known data points by constructing a linear function. Therefore, the input data consists of discrete real-time parameter sequences and time intervals between adjacent data points, while the output data is continuous virtual mapping parameter values. In this embodiment, discrete real-time parameter sequences are synchronously mapped onto virtual nodes of the digital twin. Virtual nodes are virtual units in the digital twin that simulate the various equipment and processes of the physical production line. Each virtual node corresponds to a specific piece of equipment or a production process on the physical production line. Through linear interpolation, the discrete parameters can be continuously mapped onto the virtual nodes, ensuring that the virtual mapping relationship can accurately reflect the layout of the actual production line and the correspondence between the parameters of each process. Subsequently, by comparing the output data of the virtual model with the historical physical production line data, the mapping parameters are adjusted to verify and improve the accuracy of the mapping. The mapping parameters refer to various coefficients, function parameters, etc., used to establish the virtual mapping relationship, and their values ​​directly affect the accuracy of the virtual mapping.

[0024] The processed real-time parameter sequence is aligned by timestamps. A timestamp is the specific time marker corresponding to the acquisition of each parameter data. Timestamp alignment ensures consistency of parameter data from different stages and of different types across the time dimension, forming a two-dimensional raw data matrix of "parameter dimension - time dimension." The "parameter dimension" corresponds to the four core parameters: temperature, viscosity, tension, and wind speed. The "time dimension" corresponds to the time node of data acquisition. For example, if the temperature T, viscosity n, tension F, and wind speed V data at a certain moment (row i) constitute a row in the matrix, then that row's data is [T, η, F, v]. The two-dimensional raw data matrix is ​​a structured data organization form that clearly presents the values ​​of each parameter at different time points. The sliding window method involves setting a fixed-length time window (e.g., including the current moment, the data from the previous 5 moments, and the data from the next 2 moments), and sequentially moving the window based on the time node to extract all parameter data within the corresponding time period. Therefore, the input data for this method is a two-dimensional raw data matrix and a preset time window length, and the output data is a three-dimensional extended data matrix. A time window is the time range used to extract time-series correlation data. Its length can be set according to the characteristics of the production process and the needs of data analysis. The parameter data within each time window constitutes a slice in a three-dimensional extended data matrix. This method can expand isolated data from a single time node into a dataset containing time-series correlation information. Time-series correlation information refers to the changing patterns and mutual influences of parameter data at different time points, such as the effect of temperature changes at a certain moment on viscosity at subsequent moments. This more comprehensively reflects the changing patterns and mutual influences of parameters over time, thus constructing a three-dimensional extended data matrix of "time window - parameter dimension". The first dimension (time window dimension) represents the sequence number of the sliding window, with each sequence number corresponding to an independent time window; the second dimension (time dimension within the window) represents the sequence number of the time node within a single time window; the third dimension (parameter dimension) is consistent with the column dimensions of the two-dimensional original data matrix, fixed as temperature, viscosity, tension, and wind speed. The core parameters are as follows. Flattening refers to converting a multidimensional matrix into a one-dimensional vector in a certain order (e.g., first the data within the time window, then the parameter dimensions), eliminating dimensional differences and facilitating subsequent model processing. The input data for this processing method is a three-dimensional expanded data matrix, and the output data is a one-dimensional feature vector. Each one-dimensional vector corresponds to a high-dimensional feature data point at a given time. The combination of one-dimensional vectors from all time points forms a high-dimensional feature dataset containing temporal correlation information. That is, a high-dimensional feature dataset refers to a data set containing multiple feature dimensions, including both the original parameters such as temperature and viscosity, and the correlation features obtained through temporal expansion.

[0025] Principal Component Analysis (PCA) is a commonly used data analysis method. Its principle is to map high-dimensional data to a low-dimensional space through linear transformation. While preserving the main information of the data (i.e., the direction of maximum variance), it reduces the dimensionality of the data, thereby reducing the complexity of subsequent model calculations. The larger the variance, the more significant the changes in the data and the richer the information it contains. Therefore, the input data for this algorithm is a high-dimensional feature dataset containing time-series correlation information, and the output data is the dimensionality-reduced key feature dataset, the eigenvalues ​​of each principal component, and the eigenvectors. The specific process is as follows: First, calculate the covariance matrix of the high-dimensional feature dataset. The covariance matrix describes the degree of linear correlation between different features, and its elements C... ij The covariance between the i-th and j-th features is represented by the matrix. A positive covariance indicates a positive correlation between the two features, a negative covariance indicates a negative correlation, and zero indicates no linear correlation. Then, the eigenvalues ​​and eigenvectors of the covariance matrix are calculated. The eigenvalues ​​reflect the magnitude of the variance of the data along the corresponding eigenvector direction, while the eigenvectors represent the projection direction of the data in that direction. The larger the eigenvalue, the more data information is contained in the corresponding eigenvector direction. Next, principal components with eigenvalues ​​greater than a preset threshold are selected as key features. Principal components are new features obtained through linear transformation; they are linear combinations of the original features and can retain the information of the original data to the maximum extent. Key features are the principal components that have the most significant impact on the production line's operating status. Selecting key features can effectively reduce data redundancy. Based on the selected key features, a time series dimension is integrated, that is, the correlation information of key features changing over time is retained to generate a multidimensional data model. A multidimensional data model is a mathematical model that integrates multiple parameter dimensions and time dimension information, which can comprehensively and dynamically reflect the operating status and parameter change patterns of the production line, providing a reliable data foundation for subsequent production line operating status assessment.

[0026] The core execution module refers to the core functional unit in a multidimensional data model responsible for data processing, state assessment, and parameter prediction. Its performance directly determines the assessment accuracy and response speed of the multidimensional data model. Deep learning models are a type of machine learning model based on neural networks. Through a stacked structure of multiple neurons, they can automatically learn complex nonlinear relationships in data. The constructed deep learning model adopts a multilayer perceptron (MLP) structure, including an input layer, hidden layers, and an output layer. Therefore, the input data is a dataset of key features after dimensionality reduction, and the output data is the production line operating status assessment value. Neurons are the basic computational units of a neural network, capable of receiving input signals and performing nonlinear transformations before outputting results. The input layer receives key feature data after dimensionality reduction through principal component analysis. The hidden layer performs nonlinear transformations on the data through activation functions. Activation functions are functions used to introduce nonlinear factors, enabling the neural network model to fit complex nonlinear relationships. The output layer outputs the production line operating status assessment value. The full-process database refers to a database that stores all past operational data of the plastic film production line, including parameter data and corresponding actual operational status results under various scenarios such as normal operation and abnormal operation. Historical operational data is the foundation for model training, providing rich learning samples for the model. During training, historical key feature data is used as model input, and the corresponding actual operational status results are used as labels. Labels are data used to identify the category to which a sample belongs. In this embodiment, the label is the quantified value of the actual operational status of the production line. Gradient descent is a commonly used optimization algorithm. Its principle is to continuously update the model parameters along the negative gradient direction of the loss function to minimize the model's prediction error. Therefore, the input data of this algorithm are the initial values ​​of the model parameters, the loss function, the learning rate, and the preset error threshold, and the output data is the optimized model parameters. The loss function is a function used to measure the difference between the model's predicted value and the actual label value. Mean squared error (MSE) can be used. The gradient descent algorithm continuously adjusts the model's weights and bias parameters. Weights refer to the parameters of the connection strength between neurons, and bias parameters are parameters used to adjust the baseline of the neuron's output, so that the loss function value gradually decreases until the prediction error is less than the preset error threshold, thus completing model training.

[0027] The standardized parameter data, updated in real time, is processed through timestamp alignment, sliding window extraction, and principal component analysis for dimensionality reduction before being input into the trained deep learning model. The model dynamically outputs a production line operating status assessment value, ranging from 0 to 1, by performing layer-by-layer calculations and nonlinear transformations on the input key feature data. The closer the value is to 1, the more stable the production line's operating status; the closer it is to 0, the more abnormal the operating status. Simultaneously, the model undergoes online fine-tuning every preset period based on newly accumulated historical operating data. Online fine-tuning refers to iteratively updating the model parameters using newly generated historical data during normal model operation, without retraining the entire model, effectively adapting to dynamic changes in data distribution. Therefore, the input data consists of newly accumulated historical key feature data and corresponding actual operating status labels, and the output data is the fine-tuned model parameters. Online fine-tuning ensures that the model can adapt to dynamic situations such as parameter drift and environmental changes during production line operation. Parameter drift refers to the slow change in the statistical distribution of parameter data caused by factors such as changes in sensor performance and fluctuations in material characteristics over time. Environmental changes refer to changes in external environmental conditions such as temperature and humidity in the production workshop. These factors can affect the model's prediction accuracy. Online fine-tuning can maintain a high prediction accuracy of the model, thereby enabling the multidimensional data model to dynamically reflect the physical production line's operating status and providing accurate and real-time status information for subsequent dynamic control of production parameters.

[0028] Step S2: Based on the multidimensional data model and the pre-trained support vector machine model, analyze the material state evolution trend in the digital twin and synchronize it with the physical production line state. If the material state evolution trend exceeds the preset evolution threshold, predict the potential impact range by integrating historical data through the digital twin and generate a control demand report.

[0029] The analysis of material state evolution trends includes:

[0030] The processed real-time parameter sequence is input into a pre-trained support vector machine model. The data is mapped to a high-dimensional space using radial basis function kernels to construct a classification hyperplane, classifying the material state evolution types and identifying three state types: stable, gradual, and abrupt. Based on the classification results and the temporal correlation information contained in the multi-dimensional data model, the dynamic change trend of the current material state is determined. The dynamic change trend is compared with the state data collected in real time from the physical production line to calculate the deviation value. If the deviation value exceeds a preset deviation threshold, the temperature mapping coefficient and tension mapping ratio in the virtual mapping parameters are adjusted through a digital twin, and iterative optimization is performed until the deviation value is less than the preset deviation threshold. The material state evolution trend is updated based on the adjusted virtual mapping parameters, and a state synchronization report containing parameter adjustment records and trend analysis results is generated.

[0031] This includes generating a regulatory demand report, which includes:

[0032] Key parameters such as temperature, viscosity, and tension in the evolution trend of material states are extracted and compared with preset threshold values ​​for each key parameter to determine if any exceedances are detected. If an exceedance is detected, historical data with the same parameter type and similarity to the production conditions are extracted from the full-process database. A Kalman filter algorithm is used to fuse and analyze the historical data and current real-time parameters to smooth data noise and construct a chain reaction prediction model. Based on the chain reaction prediction model, the impact path of the exceedance parameters on downstream production links is simulated and predicted. Based on the simulation prediction results, the affected production links, related parameters, and parameter control requirements are determined, and a severity score is assigned to each parameter control requirement based on the potential loss level. A bubble sort algorithm is used to prioritize each parameter control requirement according to its severity score, generating a control requirement report that includes details of the affected links, control requirement content, and priority sequence.

[0033] Specifically, for analyzing the evolution trend of material states, the processed real-time parameter sequence is first input into a pre-trained support vector machine (SVM) model. The SVM model is a supervised learning model that classifies data by constructing an optimal classification hyperplane. It effectively handles high-dimensional data and has strong generalization ability, accurately identifying different evolution types in material state classification scenarios. The model's training process is based on historical parameter sequences and corresponding material state evolution type labels stored in the full-process database. The historical parameter sequences cover time-series data of temperature, viscosity, tension, and wind speed from each stage of extrusion, stretching, cooling, and winding in past production processes. The material state evolution type labels are categorized into three types based on actual production records: stable, gradual, and abrupt, confirmed by product quality inspection results and production logs. During training, the historical parameter sequences are divided into a training set and a validation set in a 7:3 ratio. The training set is used for model parameter learning, and the validation set is used to evaluate model performance, generating a pre-trained SVM model. Therefore, the input data is the processed real-time parameter sequence, including time-series data vectors of temperature, viscosity, tension, and wind speed. The output data is the classification results of the material state evolution type, corresponding to the three categories of stable, gradual, and abrupt changes, along with the corresponding classification confidence scores. The classification confidence score is a value between 0 and 1; a higher confidence score indicates a more reliable classification result. A radial basis function (RBF) kernel can be used to map the data to a high-dimensional space. The RDF kernel is a commonly used kernel function in support vector machine (SVM) models. Its function is to map linearly inseparable data in a low-dimensional space to a high-dimensional space, making it linearly separable, thus facilitating the construction of a classification hyperplane. After mapping with the RDF kernel, a classification hyperplane is constructed in the high-dimensional space. A classification hyperplane is a plane that can clearly separate data samples of different categories. Its construction is based on maximizing the margin between samples of different categories. The larger the margin, the stronger the model's classification accuracy and generalization ability, thereby classifying the material state evolution type and accurately identifying the three state types: stable, gradual, and abrupt changes. Among them, the stable state type refers to the material's state parameters fluctuating smoothly within a preset range without a significant trend; the gradual change state type refers to the material's state parameters changing slowly over time, with the change amplitude within a preset gradual change range; and the abrupt change state type refers to the material's state parameters changing drastically within a short period of time, exceeding a preset abrupt change threshold. For example... Figure 2 The image shows the classification performance verification results of the support vector machine model. Figure 2 The left figure shows the classification performance metrics of the SVM model, with precision, recall, and F1 score all exceeding 0.90 for the three state types: stable, gradual, and abrupt. The right figure shows the classification confidence distribution, indicating that the model has a high degree of confidence in recognizing different state types.

[0034] Based on the above classification results, and combined with the temporal correlation information contained in the multidimensional data model, the dynamic change trend of the current material state is determined. The dynamic change trend refers to the direction and rate of change of the material state over time, such as the temperature gradually increasing and the tension slowly decreasing. Specifically, this includes: after the support vector machine model outputs the classification results of the material state evolution type (stable, gradual, abrupt), the multidimensional data model will call its stored time-series correlation information for in-depth supplementary analysis. For the state classified as "gradual," the model will extract the parameter time-series sequence corresponding to the state, and through the deep learning model in the core execution module, analyze the rate of parameter change (the magnitude of temperature increase per hour, the gradient of viscosity decrease over time) and the continuity of change (whether it continuously conforms to the time-series logic of "parameter change in the previous stage → parameter response in the next stage"), thereby determining the specific direction of the gradual trend (continuous temperature increase, slow tension decay). For the state classified as "abrupt," the model will combine historical similar abrupt change scenario data in the time-series correlation information to analyze the parameter precursors before the abrupt change and the immediate impact after the abrupt change, clarifying the strength and diffusion path of the abrupt trend (whether the abrupt change in winding tension is caused by a sudden change in cooling wind speed, and whether it will subsequently affect the material flatness). For the state classified as "stable," the model will verify through the time-series correlation information whether the fluctuation range of parameters in multiple continuous time windows meets the stability threshold, confirming that there is no obvious trend of change. The dynamic trend is then compared with the real-time status data collected from the physical production line. The real-time status data of the physical production line refers to the current operating status data of each link of the production line collected in real time through the sensor network, including real-time parameters such as temperature, viscosity, tension, and wind speed. During the comparison process, the deviation value is calculated. The deviation value is a quantitative indicator used to measure the degree of difference between the dynamic trend and the real-time status data of the physical production line. The formula for calculating the deviation value Δ is: Δ=|T-T0|, where T is the predicted value of the parameter corresponding to the dynamic trend, and T0 is the actual value of the parameter collected in real time from the physical production line.

[0035] The preset deviation threshold is an allowable deviation range set based on historical production data and process requirements. When the deviation value exceeds this range, it indicates a mismatch between the virtual mapping in the digital twin and the physical production line status. It is necessary to adjust the temperature mapping coefficient and tension mapping ratio in the virtual mapping parameters through the digital twin. The virtual mapping parameters are key parameters used to establish the mapping relationship between the digital twin and the physical production line. The temperature mapping coefficient refers to the conversion ratio between the temperature parameters of the physical production line and the corresponding virtual temperature parameters in the digital twin. The tension mapping ratio refers to the ratio between the tension parameters of the physical production line and the corresponding virtual tension parameters in the digital twin. These two parameters are continuously adjusted through iterative optimization. Iterative optimization refers to the process of gradually reducing the deviation value by repeatedly adjusting the parameters. After each adjustment, the deviation value is recalculated until the deviation value is less than the preset deviation threshold, ensuring that the digital twin can accurately replicate the operating status of the physical production line.

[0036] The material state evolution trend is updated based on the adjusted virtual mapping parameters, making the updated evolution trend more consistent with the actual operation of the physical production line. Subsequently, a state synchronization report is generated, which includes parameter adjustment records and trend analysis results. The parameter adjustment records include information such as the virtual mapping parameter values ​​before and after the adjustment, the number of adjustments, and the deviation changes for each adjustment. The trend analysis results include the current material state evolution type, dynamic change trend, and degree of synchronization with the physical production line. This state synchronization report will serve as an important basis for subsequent control operations.

[0037] Specifically, the process of generating a control demand report first involves extracting key parameters such as temperature, viscosity, tension, and wind speed from the material's state evolution trend. These key parameters are core parameters that play a decisive role in the quality of plastic film production and the stable operation of the production line; fluctuations in these parameters directly affect product performance and production efficiency. Each key parameter is then compared with its corresponding preset parameter threshold. These preset thresholds are allowable ranges for each key parameter, set based on production process requirements, product quality standards, and historical experience data. For example, the preset temperature threshold is set at 180℃-200℃. The comparison determines whether any exceedances are possible; exceedances refer to situations where the actual value of a key parameter exceeds the preset parameter threshold.

[0038] The full-process database includes parameter data, production condition information, and corresponding processing results under various scenarios such as normal operation and abnormal operation. It extracts historical data that is consistent with the type of parameters exceeding the limit and whose production condition similarity is consistent. The production condition similarity means that the matching degree between the material type, production speed, equipment operating status, and other operating conditions in the historical production process and the current production condition reaches a preset similarity threshold, ensuring that the extracted historical data has reference value.

[0039] The Kalman filter algorithm is a recursive filtering algorithm based on the state equation of a linear system. Its principle is to continuously correct the system state estimate through two steps: prediction and updating. This effectively smooths data noise, eliminates data redundancy, and improves data reliability and accuracy. Data noise refers to irregular interference signals generated during data acquisition due to sensor errors, environmental interference, and other factors. The algorithm's input data consists of historical parameter sequences (such as time-series data of temperature and viscosity under similar past exceedance scenarios) and current real-time parameters (such as the current exceedance temperature value and real-time viscosity value). The output data is the fused optimal state estimate sequence, including smoothed time-series data of parameters such as temperature and viscosity. Specifically, the process includes: First, establishing system state equations and observation equations based on historical data. The system state equations describe the changes in parameters over time, while the observation equations establish the relationship between observed data and the system state. Next, a prediction step is performed: based on the previous state estimate and state equation, the current system state estimate and error covariance matrix are predicted. The error covariance matrix describes the uncertainty of the state estimate. Then, an update step is performed: combining the current real-time parameters (observation data) and the observation equation, the Kalman gain is calculated. The Kalman gain is used to balance the reliability of the predicted and observed values. The predicted system state estimate is then corrected based on the Kalman gain to obtain the optimal state estimate, and the error covariance matrix is ​​updated simultaneously. Through fusion analysis using the Kalman filter algorithm, after smoothing data noise, a chain reaction prediction model is constructed. This model is a mathematical model that analyzes the correlation between parameters based on the fused data and simulates the impact path of parameter exceedances on subsequent production stages. Therefore, the input data consists of the fused optimal state estimate sequence and current production parameters, while the output data includes the potential chain reaction's impact range, the degree of impact on each stage, and parameter change prediction curves. Based on this model, the impact path of excessive parameters on downstream production processes can be simulated and predicted. The impact path refers to the order in which excessive parameters propagate from the current process to each downstream process and the degree of influence on relevant parameters in each process. For example, if the temperature exceeds the limit in the extrusion process, it will lead to a decrease in material viscosity, which in turn will affect the tension control in the stretching process, and ultimately affect the product flatness in the winding process.

[0040] Affected production processes refer to downstream production processes impacted by the exceeding of critical parameters. Related parameters refer to other parameters that are associated with the exceeding parameter and fluctuate accordingly. Parameter control requirements refer to parameter adjustments needed to eliminate the impact of exceeding limits and restore stable production line operation, such as adjusting cooling fan speed or reducing traction speed. Each parameter control requirement is assigned a severity score based on its potential loss level. The potential loss level refers to the severity of potential losses such as product scrap, production downtime, and energy waste that might result from not addressing the control requirement in a timely manner. The severity score uses a scale of 1-10, with higher scores indicating higher loss levels.

[0041] Bubble sort is a simple sorting algorithm. Its principle is to repeatedly traverse the list of elements to be sorted, comparing the size of adjacent elements pairwise. If an element is out of order, they are swapped until the entire list is sorted. This algorithm is simple to operate, easy to implement, and suitable for sorting small-scale data. Therefore, the input data is a list of severity scores corresponding to the control requirements of each parameter, and the output data is a sequence of control requirements sorted from highest to lowest severity score. The bubble sort algorithm sorts the control requirements of each parameter from highest to lowest severity score, with higher severity scores indicating higher priority. The final result is a control requirement report containing details of the affected process, the content of the control requirements, and the priority sequence. Details of the affected process include the name of the affected process and the degree of impact. The content of the control requirements includes the name of the parameter to be adjusted, the target adjustment value, and the adjustment method. The priority sequence clarifies the processing order of each control requirement, providing clear guidance for subsequent virtual simulation of process parameters and the formulation of adjustment schemes.

[0042] Step S3: Based on the control demand report, perform virtual simulation of process parameters in the digital twin to generate a draft control scheme that includes adaptive adjustment of blow molding temperature. Use a program process scheduling algorithm to prioritize the simulation tasks in the draft control scheme, dynamically allocate computing resources to execute the simulation tasks in parallel, and combine the gradient descent algorithm to determine the set of optimized parameters for coordinated control of blow molding temperature and traction speed.

[0043] This includes generating a draft adjustment scheme that incorporates adaptive adjustment of blow molding temperature, including:

[0044] Based on the priority sequence in the control demand report, the wind speed-related data and corresponding control targets for the cooling process are acquired first. The wind speed data, current process parameters, and control targets are input into the finite element simulation model of the digital twin. The finite element simulation model discretizes the cooling process into a preset number of elements, sets wind speed as the boundary input, and solves the partial differential equations of heat transfer and fluid dynamics to simulate the impact of different process parameter combinations on the material state. The material state evolution path under each parameter combination is analyzed using a support vector machine model to identify the optimal parameter adjustment direction. Based on the analysis results, the adaptive adjustment amount of the blow molding temperature is calculated using a proportional-integral-derivative controller, and the traction speed is integrated with the coordinated adjustment parameters to generate a preliminary control scheme draft. Three scenarios—high load, low load, and normal load—are simulated in the digital twin to verify the stability and feasibility of all preliminary control scheme drafts. Preliminary control scheme drafts that exceed the stable production range are eliminated, and the final control scheme draft is output.

[0045] The optimal parameter set for the coordinated control of blow molding temperature and traction speed includes:

[0046] Computational resources are dynamically allocated to simulation and optimization tasks based on the ranking results. Multithreading technology is used to create multiple execution threads to process simulation tasks with different parameter combinations in parallel, simultaneously generating multiple parameter adjustment schemes that include adjustments to blow molding temperature and traction speed. These schemes are initially screened, eliminating those that do not meet production process constraints and retaining only the effective ones. A loss function is constructed, with material stability, production efficiency, and resource utilization as optimization objectives. The blow molding temperature and traction speed parameters from the multiple effective adjustment schemes are used as inputs, and the optimal parameter values ​​are iteratively calculated using a gradient descent algorithm. These optimal values ​​are then integrated to form an optimized parameter set. The optimized parameter set is preliminarily verified using a digital twin, and the verification results are output.

[0047] Specifically, the priority sequence in the control demand report clarifies the processing order of each control task. The higher the priority of the task, the more data support needs to be guaranteed. The wind speed data in the cooling process is a key parameter that affects the material cooling effect and is related to the adjustment of the blow molding temperature. The control target refers to the process standard that needs to be achieved in this process.

[0048] A finite element simulation model is a simulation model built based on the finite element analysis method. The finite element analysis method discretizes a complex physical system or structure into a finite number of interconnected elements, and obtains the simulation results of the entire system by solving the mechanical, thermal, and other related partial differential equations of each element. In this embodiment, it can accurately simulate the impact of process parameter changes on the material state. Therefore, the input data of the finite element simulation model includes cooling air velocity data, current process parameters such as extrusion temperature, tensile tension, blow molding temperature, and traction speed, as well as the target control parameters. The output data includes material state data under different combinations of process parameters, such as material temperature distribution, viscosity change, cooling rate, and tensile uniformity. Specifically, this includes: discretizing the cooling process into a predetermined number of units. Discretization refers to dividing a continuous physical region into multiple finite-sized sub-regions, each of which is a unit. By analyzing each unit, the entire cooling process can be accurately simulated. Wind speed is set as a boundary input. Boundary input refers to the boundary condition parameters set in the simulation model to simulate the external environmental factors affecting the cooling process in actual production. Subsequently, partial differential equations of heat transfer and fluid dynamics are solved. The heat transfer partial differential equation describes the laws of heat transfer during the cooling process, while the fluid dynamics partial differential equation describes the characteristics of the cooling airflow. By solving these two types of equations, data such as the temperature distribution and wind speed field distribution of the cooling process under different wind speed conditions can be obtained. This allows for the simulation of the influence of different combinations of process parameters on the material state, such as the cooling rate and viscosity changes of materials under different combinations of wind speed and temperature.

[0049] The Support Vector Machine (SVM) model has been trained using historical data. Its principle is to map linearly inseparable data in low-dimensional space to a high-dimensional space using a kernel function, constructing an optimal classification hyperplane to achieve data classification. Here, it is used to classify and identify the material state evolution types under different parameter combinations. Therefore, the input data of this SVM model consists of material state data (including time-series data such as temperature, viscosity, tension, and wind speed) corresponding to each parameter combination, and the output data is the classification results of the material state evolution types (stable, gradual, and abrupt). The material state evolution path refers to the process and pattern of material change from its initial state to its final state under the influence of different process parameters. By analyzing the evolution path corresponding to each parameter combination using the SVM model, the optimal parameter adjustment direction can be accurately identified. The optimal parameter adjustment direction refers to the parameter change trend that enables the material state to develop towards a stable state that meets production requirements, such as increasing wind speed or decreasing blow molding temperature.

[0050] The Proportional-Integral-Derivative (PID) controller is a commonly used automatic control algorithm. Its principle combines proportional, integral, and derivative control methods, adjusting based on the system's deviation signal to output a suitable control quantity. Proportional control is used for rapid response to deviations, integral control for eliminating steady-state error, and derivative control for predicting deviation trends. The input data for this PID controller is the deviation between the material's current state parameters (including viscosity, temperature, and tensile tension after cooling) and the target control value. The output data is the adaptive adjustment amount for the blow molding temperature (a specific temperature correction value). Specifically, it uses the deviation between the material's current state and the target control value as input to the PID controller. Through preset proportional, integral, and derivative coefficients, it calculates the adaptive adjustment amount for the blow molding temperature. This adaptive adjustment amount is a temperature correction value dynamically adjusted according to real-time changes in the material's state, ensuring that the blow molding temperature always matches production requirements. Simultaneously, it integrates a traction speed adjustment parameter, which is a traction speed adjustment value matched to the blow molding temperature adjustment. This coordinated control ensures the stability of the material during stretching and cooling processes, preventing abnormal material states caused by adjusting a single parameter, and thus generating a preliminary adjustment plan.

[0051] High-load scenarios refer to production lines operating at full capacity with maximum material handling capacity. Low-load scenarios refer to production lines operating at low capacity with minimum material handling capacity. Normal-load scenarios refer to production lines operating under normal conditions with rated material handling capacity. These three scenarios essentially cover the main operating conditions of the production line. The stability and feasibility of a preliminary adjustment plan draft are verified by simulating its implementation in these three scenarios within a digital twin. Stability means that the material state can be maintained within a preset stable range after implementation, without significant fluctuations. Feasibility means that the plan can be smoothly executed in actual production without conflicting with existing production equipment and processes. Preliminary adjustment plan drafts exceeding the production stable range (the allowable range of parameters ensuring product quality and normal production line operation) are eliminated, and the final output is a compliant adjustment plan draft.

[0052] Specifically, computing resources include hardware resources such as CPU cores, memory, and storage. Dynamic allocation refers to adjusting the resource allocation ratio in real time based on task priority and current resource usage to ensure that high-priority tasks receive sufficient resource support and improve task execution efficiency. Simulation tasks refer to tasks that perform virtual simulations of different parameter combinations, while optimization tasks refer to tasks that calculate the optimal parameters based on simulation results. The input data for the program process scheduling algorithm consists of information about each simulation task in the draft adjustment scheme, including task identifiers, associated process parameters, required computing resources, and task dependencies. The output data is the simulation task priority ranking result, i.e., a list of tasks arranged from highest to lowest priority.

[0053] Multithreading technology refers to creating multiple independent execution units (i.e., threads) within a single process. Multiple threads can execute different tasks simultaneously and in parallel, significantly improving task processing efficiency compared to single-threaded execution. Multithreading technology is used to process simulation tasks with different parameter combinations in parallel, synchronously generating multiple parameter adjustment schemes that include adjustments to blow molding temperature and traction speed. Each parameter adjustment scheme refers to a specific combination of adjustment values ​​for blow molding temperature and traction speed designed for different process requirements. The input data for multithreading technology consists of a priority-sorted list of simulation tasks and the corresponding parameter combination ranges for each task. The output data consists of multiple parameter adjustment schemes, each containing specific blow molding temperature adjustment values, traction speed adjustment values, and corresponding material state prediction results.

[0054] Production process constraints refer to limitations set based on production equipment performance, product quality standards, and safe operating procedures. Examples include blow molding temperature not exceeding the equipment's maximum withstand temperature and traction speed not falling below the minimum speed requirement for material stretching. Effective parameter adjustment schemes that comply with these constraints should be retained. Effective parameter adjustment schemes are those that can be implemented under existing production conditions and are expected to achieve the control targets.

[0055] The loss function is constructed by weighting and summing material state stability indicators, production energy efficiency indicators, and resource utilization indicators. The weight coefficients are set according to the importance of each optimization objective. The input data of this loss function are the blow molding temperature and traction speed parameters in the parameter adjustment scheme, along with the corresponding material state stability data, production energy efficiency data, and resource utilization rate data. The output data is a single loss value, used to quantify the degree of fit between the scheme and the optimization objective. Gradient descent is a commonly used optimization algorithm. Its core principle is to continuously update the parameter values ​​along the negative gradient direction of the loss function, gradually reducing the loss function value until it reaches the minimum or convergence state. The corresponding parameter value at this point is the optimal parameter value. The input data of this gradient descent algorithm are the initial parameter values ​​of blow molding temperature and traction speed in an effective parameter adjustment scheme, the loss function, the preset learning rate, and the error threshold. The output data are the optimal blow molding temperature and optimal traction speed values ​​that minimize the loss function. Specifically, the process includes: first, initializing the initial parameter values ​​for blow molding temperature and traction speed, calculating the initial loss function value and gradient direction, and then updating the parameter values ​​according to a preset learning rate. The learning rate refers to the step size of each parameter update, which is used to control the magnitude of parameter adjustment and avoid non-convergence due to excessively large step sizes or excessively low iteration efficiency due to excessively small step sizes. The above process is repeated for iterative calculation until the loss function value is less than the preset error threshold. The blow molding temperature and traction speed parameters obtained at this time are the optimal parameter values. These optimal parameter values ​​are then integrated to form an optimized parameter set.

[0056] The verification process involves inputting the optimized parameter set into the simulation model of the digital twin to simulate its operation in a real production scenario, acquiring data such as material state evolution and production energy efficiency. The input data for the digital twin verification process includes the optimized parameter set and historical verification sample data from the full-process database. The output data is the verification result, including material state stability assessment values, production energy efficiency assessment values, and whether it meets the stability range. The simulation results are compared with the control targets to determine whether the optimized parameter set can meet production requirements. The final verification result is output, indicating whether the verification passed or failed. If the verification fails, the parameter adjustment plan or optimization algorithm parameters need to be readjusted.

[0057] Step S4: Check whether the winding tension value in the optimized parameter set deviates from the stable range. If it does, iterate and correct the parameters in the digital twin through feedback loop, generate the final coordinated control command, input the final coordinated control command into the digital twin, combine the verification sample data to simulate the material state evolution trend, compare it with the real-time state of the physical production line, confirm the synchronization and output a report, and complete the dynamic control of plastic film production parameters.

[0058] The generation of final coordination and control instructions includes:

[0059] If the winding tension value deviates from the stable range, the upstream data is processed through a digital twin to extract historical temperature data and real-time tension data from the extrusion and stretching processes. The cause of the deviation is identified and the degree of deviation is quantified using a support vector machine model. Based on the cause and degree of deviation, the parameter correction range is determined, the feedback loop is initialized, the optimized parameter set is input into the digital twin, the adjusted winding tension change is predicted, and the parameters are iteratively corrected until the winding tension value enters the stable range. The optimal parameters after iteration are summarized to generate the final coordinated control command that includes real-time response tasks and delayed execution tasks, and the parameter adjustment path during the command generation process is recorded.

[0060] The predicted changes in winding tension after adjustment include:

[0061] Based on the key features output by the multidimensional data model and combined with relevant historical data on winding tension in the full-process database, a deep learning neural network model containing an input layer, hidden layer, and output layer is constructed. The dimension of the input layer matches the number of key features, the hidden layer fits the nonlinear mapping relationship between key features and winding tension using a hyperbolic tangent activation function, and the output layer outputs the predicted value of winding tension. The neural network model is trained using historical parameter adjustment data as input and the corresponding actual change value of winding tension as the label. An adaptive moment estimation optimization algorithm is used to iteratively optimize the model weight parameters, and an early stopping mechanism is introduced to avoid overfitting until the prediction error of the neural network model is less than a preset error threshold. The current set of optimized parameters and the real-time production status data output by the multidimensional data model are input into the trained neural network model, and real-time temperature and viscosity change data of the upstream extrusion and stretching stages and wind speed adjustment data of the cooling stage are input simultaneously to construct a multidimensional input feature vector. Through forward propagation calculation of the neural network model, the dynamic change curve of winding tension under different parameter adjustment ranges is predicted, and the tension change rate, peak value, and stable value are quantified. The judgment result of whether the winding tension in the predicted change curve has entered the preset stable range is extracted. If it has not entered, the parameter adjustment direction suggestion is output.

[0062] This includes confirming synchronization and outputting a report, including:

[0063] Historical verification sample data matching the current production conditions is extracted from the full-process database. This historical verification sample data includes complete sequences of temperature, viscosity, tension, and wind speed. The historical verification sample data and the final coordinated control commands are fused to form an input dataset, which is then transmitted to the simulation model via a digital twin interface. The simulation model simulates the state evolution of materials during extrusion, stretching, cooling, and winding, extracting time-series data of key parameters for each stage to generate state prediction data for the physical production line. Real-time state data of the physical production line is collected, and the Euclidean distance between the state prediction data and the real-time state data is calculated as a deviation metric. If the deviation metric is less than a preset synchronization threshold, state synchronization is confirmed. If the deviation metric is greater than the preset synchronization threshold, the simulation model parameters are adjusted through feedback loops, and the simulation is repeated until the synchronization requirements are met. A state synchronization confirmation report containing synchronization indicators, a summary of prediction data, and parameter adjustment suggestions is output.

[0064] Specifically, the stable range is the allowable range of winding tension values ​​set based on the requirements of plastic film production processes, equipment performance limits, and historical stable operating data. The winding tension value is a key parameter affecting the film's winding flatness, tightness, and subsequent processing performance. Deviating from the stable range can lead to quality problems such as wrinkles, tensile deformation, and even breakage of the film.

[0065] If the winding tension value deviates from the stable range, the upstream data, including the extrusion and stretching stages, is processed using a digital twin. Parameter changes in these stages have a cascading effect on the winding tension through material state transmission. Specifically, historical temperature data from the extrusion stage and real-time tension data from the stretching stage are extracted. Historical temperature data refers to the time-series temperature data of past operations in the extrusion stage stored in the full-process database, while real-time tension data refers to the tension values ​​collected in real-time by tension sensors in the stretching stage. Subsequently, the cause of the deviation is identified using the aforementioned Support Vector Machine (SVM) model. SVM is a supervised learning model whose core principle is to map linearly inseparable data in low-dimensional space to a high-dimensional space using kernel functions, constructing an optimal classification hyperplane to achieve data classification. It has strong generalization ability and classification accuracy. Here, the input data for this model includes historical temperature data from the extrusion stage, real-time tension data from the stretching stage, and the deviation of the current winding tension value from the stable range. The output data is the classification result of the deviation cause, such as viscosity change deviation due to temperature drift or stretching tension fluctuation deviation, as well as a quantified value of the deviation severity. The quantified value of the deviation severity is a value between 0 and 1, with values ​​closer to 1 indicating a more severe deviation. By classifying and processing the input data using this model, the core factors causing the winding tension deviation can be accurately located, and the severity of the deviation can be quantified.

[0066] The parameter correction magnitude refers to the specific numerical changes in parameters such as blow molding temperature, traction speed, and cooling air velocity required to bring the winding tension back to a stable range. The feedback loop is a closed-loop control mechanism that iteratively optimizes parameters by continuously comparing the predicted results after parameter adjustment with the winding tension values ​​within the stable range corresponding to the target values. The adjusted winding tension change is predicted using a simulation model in a digital twin. If the predicted result still deviates from the stable range, the parameters are adjusted based on the direction and magnitude of the deviation, repeating this process until the winding tension value enters the stable range, completing the parameter iterative correction. The optimal parameters refer to the combination of parameters that, after multiple iterative corrections, can stabilize the winding tension within a preset range and meet other production indicators, such as production energy efficiency and material stability. Real-time response tasks refer to parameter adjustment operations that require immediate execution, such as emergency adjustment of cooling air velocity to quickly stabilize the winding tension; delayed execution tasks refer to optimization operations that can be gradually executed in subsequent production cycles, such as fine-tuning the extrusion temperature to optimize long-term production energy efficiency. The parameter adjustment path refers to the specific values, adjustment time, and corresponding winding tension changes for each parameter adjustment, providing a traceability basis for subsequent production optimization and troubleshooting.

[0067] Specifically, key features refer to the feature parameters that reflect the core information of the production line's operating status, obtained through principal component analysis and dimensionality reduction, such as temperature-dominant features and viscosity-related features. Relevant historical data includes records of winding tension changes under different parameter combinations in past production, and corresponding production condition information. The input layer dimension of the deep learning neural network model matches the number of key features, meaning the number of neurons in the input layer equals the number of key features, used to receive the input feature data. The hidden layer is set to three layers, with 128, 64, and 32 neurons in each layer, respectively. A hyperbolic tangent activation function is used to fit the nonlinear mapping relationship between key features and winding tension. The output range of the hyperbolic tangent activation function is [-1, 1], which can effectively handle the nonlinear relationship of the data and enhance the model's fitting ability. The output layer has one neuron, used to output the predicted winding tension value.

[0068] Historical parameter adjustment data includes adjusted values ​​for parameters such as blow molding temperature, traction speed, and cooling air velocity from past production, along with corresponding key feature data. The actual change in winding tension is used as the label, where the label refers to the actual measured change in winding tension after parameter adjustment in the historical data, used to measure the accuracy of the model's prediction results. An adaptive moment estimation optimization algorithm (Adam algorithm) is used to iteratively optimize the model's weight parameters. Adaptive moment estimation is a commonly used neural network optimization algorithm that combines the advantages of momentum gradient descent and RMSProp algorithms. By calculating the first and second moment estimates of the gradient, it dynamically adjusts the learning rate, enabling rapid convergence and reducing the likelihood of getting trapped in local optima. Therefore, the input data for this algorithm consists of the model's loss function gradient, the current weight parameters, and the preset initial learning rate value; the output data is the updated model weight parameters. Simultaneously, an early stopping mechanism is introduced to avoid overfitting. Early stopping refers to using validation set data to verify model performance at regular intervals during model training. If the prediction error on the validation set increases consecutively, training is stopped, and the current optimal model parameters are retained. Overfitting refers to the phenomenon where a model performs excellently on the training set but its prediction accuracy drops significantly on new data. The early stopping mechanism effectively improves the model's generalization ability. Training continues until the prediction error of the neural network model is less than a preset error threshold, at which point model training is complete.

[0069] A multi-dimensional input feature vector is constructed to ensure that the input data comprehensively reflects the real-time operating status of the production line. Forward propagation refers to the process where data is transmitted from the input layer through a nonlinear transformation in the hidden layers to the output layer to obtain the prediction result. Specifically, after the input feature vector is received at the input layer, it passes through each hidden layer sequentially, is processed by the hyperbolic tangent activation function, and then transmitted to the next layer. Finally, the predicted winding tension value is output at the output layer. Based on the input data under different parameter adjustment ranges, a dynamic change curve of the winding tension is generated. This dynamic change curve can intuitively show the change pattern of the winding tension over time, from which key indicators such as the rate of change of tension (the amount of tension change per unit time), the peak value (the maximum value during the tension change process), and the stable value can be quantitatively extracted. Parameter adjustment direction suggestions refer to the parameter adjustment guidelines given based on the prediction results to bring the winding tension back to a stable range.

[0070] Production condition similarity refers to the degree to which the material type, production speed, equipment operating status, and other operating conditions in historical production processes match the current production conditions, reaching a preset similarity threshold to ensure the reference value of historical verification sample data. Historical verification sample data contains complete sequences of temperature, viscosity, tension, and wind speed, comprehensively reflecting the production line's operating status data under corresponding conditions. The fusion process for forming the input dataset involves aligning the time-series information in the historical verification sample data with the parameter adjustment information in the final coordinated control command, ensuring data consistency and correlation. This data is then transmitted to the simulation model through a digital twin interface, a standardized interface for data transmission and interaction, guaranteeing the real-time performance and accuracy of data transmission.

[0071] The simulation model is a mathematical model built based on the aforementioned finite element analysis method, capable of accurately simulating the physical processes of each stage of the production line. By solving relevant partial differential equations such as heat transfer and fluid dynamics, it simulates the state changes of materials under different parameter conditions, such as temperature changes, viscosity evolution, and tension transmission. It extracts time series data of key parameters of each stage and generates state prediction data of the physical production line. State prediction data refers to the predicted values ​​of key parameters of each stage of the production line output by the simulation model for a future period of time, including time series data such as temperature, viscosity, tension, and wind speed.

[0072] Real-time status data refers to the key parameter values ​​of each stage of the production line's current operation, collected in real time through a sensor network. The Euclidean distance d is a commonly used indicator to measure the degree of difference between two multidimensional data points; its calculation formula is: ,in, For the k-th parameter value in the state prediction data, Here, n represents the k-th parameter value in the real-time status data, where n is the number of parameter dimensions. A smaller Euclidean distance indicates a smaller difference between the two sets of data and better synchronization. If the deviation metric is less than the preset synchronization threshold, the virtual state in the digital twin is confirmed to be synchronized with the physical production line state. If the deviation metric is greater than the preset synchronization threshold, the simulation model parameters are adjusted through a feedback loop. These parameters include mesh generation accuracy and boundary condition coefficients. The simulation is then recalculated until the deviation metric meets the synchronization requirements. In the status synchronization confirmation report, synchronization indicators include the deviation metric and synchronization compliance rate (the percentage of parameters that meet the synchronization requirements). The predicted data summary includes statistical information such as the predicted maximum, minimum, and average values ​​of key parameters for each stage. Parameter adjustment suggestions refer to recommendations for further optimizing production parameters based on the synchronization verification results, such as fine-tuning the traction speed to improve the synchronization accuracy of the winding stage.

[0073] For example, Figure 3 The diagram shows the verification results of the synchronization between the real-time status of the physical production line and the predicted status of the digital twin. Figure 3 The paper presents time-series comparison curves of real-time and predicted data for three key parameters: temperature, tension, and viscosity. The results show a high degree of agreement between the predicted and real-time data, with an average relative error of less than 2%. The Euclidean distance of the three parameters gradually decreases from 12.5 in the initial stage to 4.2 in the stable stage as the production process progresses, indicating a significant improvement in the synchronization between the digital twin and the physical production line over time. The histogram of synchronization error distribution shows a normal distribution, confirming the high-precision synchronization prediction capability of this method. The improvement trend of synchronization at different time stages shows an increasing improvement from stage 1 to stage 5, exhibiting a gradual convergence trend. These synchronization verifications ensure that the digital twin accurately reflects the actual state of the physical production line, providing a reliable basis for dynamic control. Furthermore, the paper demonstrates that the control method of this invention can effectively reduce physical-virtual state deviation and improve the accuracy of production control.

[0074] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0075] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0076] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for dynamically controlling plastic film production parameters using program process scheduling, characterized in that, The method includes: Real-time parameter sequences from multiple stages of a plastic film production line are collected using a sensor network. These real-time parameter sequences are then aggregated into a digital twin to establish a virtual mapping and construct a multi-dimensional data model. Based on the multidimensional data model and the pre-trained support vector machine model, the material state evolution trend is analyzed in the digital twin and synchronized with the physical production line state. If the material state evolution trend exceeds the preset evolution threshold, the potential impact range is predicted by integrating historical data through the digital twin, and a control demand report is generated. Based on the aforementioned regulation demand report, virtual simulation of process parameters is performed in the digital twin to generate a draft regulation scheme. A program process scheduling algorithm is used to prioritize the simulation tasks in the draft regulation scheme, dynamically allocate computing resources to execute the simulation tasks in parallel, and combine the gradient descent algorithm to determine the set of optimization parameters. The system detects whether the winding tension value in the optimized parameter set deviates from the stable range. If it does, the parameters are iteratively corrected in the digital twin through a feedback loop to generate a final coordinated control command. The final coordinated control command is input into the digital twin, and the material state evolution trend is simulated in combination with the verification sample data. The result is compared with the real-time state of the physical production line to confirm the synchronization and output a report, thus completing the dynamic control of plastic film production parameters.

2. The method according to claim 1, characterized in that, Constructing a multidimensional data model includes: The system collects temperature, viscosity, tension, and air velocity data from each stage of the production line, including extrusion, stretching, cooling, and winding, to form a real-time parameter sequence covering the entire production process. The real-time parameter sequence is processed by median filtering for noise reduction and z-score normalization. The processed real-time parameter sequence is then transmitted to a digital twin, and a virtual mapping relationship with the physical production line is established through linear interpolation. The processed real-time parameter sequence is aligned by timestamps to form a two-dimensional original data matrix. Based on the two-dimensional original data matrix, the historical parameter data before and after each time node is extracted using the sliding window method to construct a three-dimensional extended data matrix. The three-dimensional extended data matrix is ​​flattened and converted into a high-dimensional feature dataset containing time-series correlation information. Principal component analysis (PCA) is applied to reduce the dimensionality of the high-dimensional feature dataset. The covariance matrix is ​​calculated and the eigenvalues ​​and eigenvectors are solved. Principal components with eigenvalues ​​greater than a preset feature threshold are selected as key features. The time series dimension is integrated based on the key features to generate the multidimensional data model.

3. The method according to claim 2, characterized in that, Analyze the evolution trend of material states, including: The processed real-time parameter sequence is input into a pre-trained support vector machine model. The data is mapped to a high-dimensional space through radial basis function kernels to construct a classification hyperplane, which classifies the material state evolution type and identifies three state types: stable, gradual and abrupt. Based on the classification results and the temporal correlation information contained in the multidimensional data model, the dynamic change trend of the current material state is determined, and the dynamic change trend is compared with the state data collected in real time from the physical production line to calculate the deviation value. If the deviation value exceeds the preset deviation threshold, the temperature mapping coefficient and tension mapping ratio in the virtual mapping parameters are adjusted through the digital twin, and iterative optimization is performed until the deviation value is less than the preset deviation threshold. The material state evolution trend is updated based on the adjusted virtual mapping parameters, and a state synchronization report containing parameter adjustment records and trend analysis results is generated.

4. The method according to claim 1, characterized in that, Generate a regulatory requirement report, including: The key parameters of temperature, viscosity, and tension in the evolution trend of material state are extracted and compared with the preset parameter thresholds corresponding to each key parameter to determine whether there is any exceeding of the limit. If a key parameter is detected to be out of limit, historical data that is consistent with the type of the out-of-limit parameter and has a similarity to the production conditions will be extracted from the full-process database. Kalman filtering algorithm is used to fuse and analyze historical data and current real-time parameters to smooth data noise, construct a chain reaction prediction model, and simulate and predict the impact path of excessive parameters on downstream production links based on the chain reaction prediction model. Based on the simulation and prediction results, the affected production links, related parameters and parameter control requirements are determined, and a severity score is assigned to each parameter control requirement based on the degree of potential loss. The bubble sort algorithm is used to prioritize the control requirements of each parameter according to their severity scores, and a control requirement report is generated, which includes details of the affected links, the content of the control requirements, and the priority sequence.

5. The method according to claim 1, characterized in that, Generate a draft adjustment plan, including: Based on the priority sequence in the aforementioned regulation demand report, wind speed-related data and corresponding regulation targets for the cooling process are obtained first. The wind speed data, current process parameters and control targets are input into the finite element simulation model of the digital twin. The finite element simulation model discretizes the cooling process into a preset number of elements, sets the wind speed as the boundary input, solves the partial differential equations of heat transfer and fluid dynamics, and simulates the influence of different combinations of process parameters on the material state. By analyzing the material state evolution path under various parameter combinations using a support vector machine model, the optimal parameter adjustment direction can be identified. Based on the optimal parameter adjustment direction, the adaptive adjustment amount of blow molding temperature is calculated by a proportional-integral-derivative controller, and the traction speed is integrated to generate a preliminary adjustment scheme draft. Simulate three scenarios—high load, low load, and normal load—in a digital twin to verify the stability and feasibility of all the preliminary adjustment scheme drafts. Eliminate the preliminary adjustment scheme drafts that exceed the stable production range and output the final adjustment scheme draft.

6. The method according to claim 1, characterized in that, Determine the optimal set of parameters for the coordinated control of blow molding temperature and traction speed, including: Computational resources are dynamically allocated to the simulation and optimization tasks based on the sorting results; Multithreading technology is used to create multiple execution threads to process simulation tasks with different parameter combinations in parallel, and to generate multiple parameter adjustment schemes that include parameters for adjusting blow molding temperature and traction speed simultaneously. A preliminary screening of the parameter adjustment schemes was conducted, eliminating those that did not meet the constraints of the production process and retaining the valid ones. A loss function is constructed with material state stability, production energy efficiency and resource utilization as optimization objectives. The blow molding temperature and traction speed parameters from multiple effective parameter adjustment schemes are used as inputs. The optimal parameter values ​​are iteratively calculated using the gradient descent algorithm. The optimal parameter values ​​are then integrated to form the set of optimized parameters. The optimized parameter set is initially verified using the digital twin, and the verification results are output.

7. The method according to claim 1, characterized in that, Generate the final coordinated control instructions, including: If the winding tension value is detected to deviate from the stable range, the upstream data is processed through the digital twin to extract historical temperature data and real-time tension data of the extrusion and stretching processes, and the cause of the deviation is identified and the degree of deviation is quantified by classifying the deviation using a support vector machine model. Based on the cause and degree of deviation, determine the parameter correction range, initialize the feedback loop, input the optimized parameter set into the digital twin, predict the change in winding tension after adjustment, and iteratively correct the parameters until the winding tension value enters the stable range. The optimal parameters after iteration are summarized to generate the final coordinated control command that includes real-time response tasks and delayed execution tasks, and the parameter adjustment path during the command generation process is recorded.

8. The method according to claim 7, characterized in that, Predicted changes in adjusted winding tension include: Based on the key features output by the multidimensional data model, and combined with the relevant historical data of winding tension in the full-process database, a neural network model containing an input layer, a hidden layer and an output layer is constructed. The dimension of the input layer matches the number of key features. The hidden layer fits the nonlinear mapping relationship between key features and winding tension through the hyperbolic tangent activation function. The output layer outputs the predicted value of winding tension. The neural network model is trained by taking historical parameter adjustment data as input and the corresponding actual change value of winding tension as label. The adaptive moment estimation optimization algorithm is used to iteratively optimize the model weight parameters until the prediction error of the neural network model is less than the preset error threshold. Input the current set of optimized parameters and the real-time output of the production process status data of the multi-dimensional data model into the trained neural network model, and simultaneously input the real-time temperature and viscosity change data of the upstream extrusion and stretching processes and the wind speed adjustment data of the cooling process to construct a multi-dimensional input feature vector. By using forward propagation calculations of a neural network model, the dynamic change curve of winding tension under different parameter adjustment ranges is predicted, and the tension change rate, peak value, and stable value are quantified. Extract the judgment result of whether the winding tension in the predicted change curve has entered the preset stable range. If it has not entered, output the parameter adjustment direction suggestion.

9. The method according to claim 1, characterized in that, Confirm synchronization and generate a report, including: Historical verification sample data that matches the current production conditions are extracted from the full-process database. The historical verification sample data includes a complete sequence of temperature, viscosity, tension, and wind speed. The historical verification sample data and the final coordinated control commands are fused to form an input dataset, which is then transmitted to the simulation model through the digital twin interface. The simulation model simulates the state evolution of materials during extrusion, stretching, cooling, and winding, extracts time series data of key parameters in each stage, and generates state prediction data for the physical production line. Collect real-time status data of the physical production line, calculate the Euclidean distance between the predicted status data and the real-time status data as a deviation metric, and confirm status synchronization if the deviation metric value is less than a preset synchronization threshold. If the deviation metric exceeds the preset synchronization threshold, the simulation model parameters are adjusted through feedback loops, and the simulation is repeated until the synchronization requirements are met. A status synchronization confirmation report containing synchronization indicators, a summary of predicted data, and parameter adjustment suggestions is output.

Images