Adaptive control method and system based on etee film stretching process

Abstract

The application discloses a self-adaptive control method and system based on an ETEE film stretching process, and particularly relates to the field of film stretching process control, and comprises the steps of constructing a multi-dimensional target process fingerprint and a self-adaptive tolerance band fused with material and process parameters, generating a real-time process fingerprint vector through a distributed sensing network and a deep learning model, performing a dimension-by-dimension comparison and coupling analysis, generating a deviation trend chart and evaluating a process state severity level, starting a multi-objective optimization decision based on the severity level, dynamically decoupling a control instruction and generating an adjustment strategy to guide the process state back into the tolerance band, and finally dynamically optimizing the tolerance band width through post-analysis and machine learning to realize continuous evolution of the control system. The application solves the problem that a traditional control method is difficult to adapt to material fluctuation and process coupling, and realizes intelligent and self-adaptive precise control of the whole process from sensing, decision-making to learning in the film stretching process.

Description

Technical Field

[0001] This invention relates to the field of thin film stretching process control technology, and more specifically, to an adaptive control method and system based on ETEE thin film stretching process. Background Technology

[0002] Due to their excellent mechanical properties, weather resistance, and light transmittance, thin films are widely used in building curtain walls, aerospace, and other fields. The performance of thin films is highly dependent on the process control during the stretching process, such as the precise matching of temperature, stretching ratio, and speed. Traditional stretching process control relies heavily on preset process curves and operator experience, which makes it difficult to adapt to batch fluctuations in raw materials and changes in equipment status, resulting in unstable film product quality and a need to improve yield.

[0003] Existing technologies are based on PID (proportional-integral-derivative) control and fixed parameter settings in automated control systems. Based on historical production experience or pilot test results, a set of fixed target values ​​for temperature, tension, and speed are set for each stage of the stretching process. Temperature sensors, tension sensors, and other sensors deployed on the production line are used to compare the detected real-time signals with the target values. The deviation is calculated by the PID controller and then outputs control signals to actuators such as heaters and variable frequency motors, thereby attempting to stabilize the process parameters near the preset fixed targets.

[0004] However, in practical use, it still has some shortcomings. For example, the control target is a single, isolated process parameter, while the final quality of the thin film is the comprehensive result of the complex coupling effect of multiple process parameters in the spatiotemporal domain. Secondly, by using fixed target setpoints and PID parameters, it lacks the ability to perceive and adaptively adjust to changes in the characteristics of incoming batches. The scheme is static and can only be compensated after deviations occur. It cannot perform predictive control based on the evolution trend of the process state and does not have the ability to learn from historical data and self-optimize, resulting in limited control accuracy and intelligence level. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an adaptive control method and system based on the ETEE thin film stretching process, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adaptive control method based on ETEE thin film stretching process, comprising S1: constructing a dynamically updated multidimensional target process fingerprint based on material batch characteristics and initial process parameters, and setting a dynamically adjustable adaptive tolerance band for each dimension by initializing the reference fingerprint; S2: During the stretching process, the physical state data of the film is collected synchronously through a distributed sensor array, and the data is calculated and mapped in real time into a real-time process fingerprint vector corresponding to the process fingerprint dimension. S3: Compare the real-time process fingerprint vector with the target process fingerprint dimension by dimension to generate a multi-dimensional deviation situation map, calculate the instantaneous deviation value of each dimension, and evaluate the stability of the current process state and the severity of the deviation situation by analyzing the evolution trend of the deviation in time and space and the coupling relationship between each dimension. S4: Based on the deviation situation map, initiate a multi-objective optimization decision-making process, select the intensity level of control intervention according to the severity level of the situation, dynamically decouple control commands according to the spatiotemporal distribution characteristics and coupling relationship of the deviation, generate adjustment strategies, and guide the real-time process fingerprint back to the adaptive tolerance band of the target fingerprint. S5: After a single or multiple consecutive process cycles, perform a post-hoc analysis on the target process fingerprint, adjustment strategy, and its effects. By analyzing the correlation between the control strategy and fingerprint convergence in historical data, dynamically optimize the adaptive tolerance band width of the target process fingerprint.

[0007] An adaptive control system based on ETEE thin film stretching process includes a process fingerprint construction module: receiving material batch characteristic data and initializing a baseline process fingerprint model based on predefined process parameters, and assigning an adaptive tolerance band to each dimension; Multimodal sensing and fusion module: It communicates with the distributed sensor array, synchronously collects the physical state data of the thin film during the stretching process, and calculates and maps it in real time into a real-time process fingerprint vector corresponding to the dimension of the process fingerprint construction module; Fingerprint Deviation Situation Assessment Module: The received real-time process fingerprint vector is compared with the target process fingerprint dimension by dimension to generate a multi-dimensional deviation situation map. By calculating the instantaneous deviation value of each dimension and analyzing the evolution trend of the deviation in time and space and the coupling relationship between each dimension, the stability of the current process state and the severity level of the deviation situation are assessed. Adaptive strategy decision-making module: Based on the multi-dimensional deviation situation map and severity level, it initiates a multi-objective optimization decision-making process, selects the intensity level of control intervention according to the severity level, and dynamically decouples control commands and generates adjustment strategies based on the spatiotemporal distribution characteristics and coupling relationship of the deviation. Closed-loop evolution learning module: Connected to the process fingerprint construction module and the adaptive strategy decision module, after the process cycle ends, it dynamically optimizes and outputs the adjustment command of the adaptive tolerance band width of the target process fingerprint by analyzing the correlation between the control strategy and the convergence of the fingerprint in historical data, and continuously optimizes the control system.

[0008] The technical effects and advantages of this invention are as follows: This invention constructs a "process fingerprint" that integrates material properties and process parameters as a control benchmark. It uses multimodal sensing and fusion technology to generate corresponding real-time fingerprint vectors in real time, elevating the control target from the traditional, isolated single parameter to the "state" level that comprehensively characterizes the intrinsic quality of the thin film. By analyzing the spatiotemporal evolution and coupling relationship of deviations in various dimensions, it performs collaborative decision-making and decoupled control, fundamentally solving the control problem caused by strong coupling of multiple process parameters. This invention introduces a dynamically updated adaptive tolerance band and a closed-loop evolutionary learning mechanism to dynamically initialize the target process fingerprint based on the characteristics of material batches. After each production cycle, through post-hoc analysis, it autonomously learns and optimizes the tolerance band width and control strategy, enabling the control system to proactively adapt to changes in the characteristics of raw materials in different batches and exhibit stronger robustness to unexpected disturbances in the process. This effectively reduces the scrap rate and decreases the reliance on the experience of senior operators. This invention implements control actions through real-time monitoring, situation assessment, and multi-objective optimization decision-making. After the cycle ends, it analyzes the control effect and convergence, and optimizes the control benchmark and strategy model in reverse. This enables the entire control system to continuously iterate and improve with the accumulation of production data, ultimately achieving a continuous improvement in the accuracy and efficiency of process control. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall structure of the method of the present invention; Figure 2 This is a schematic diagram of the dynamic multidimensional target process fingerprint construction process of the present invention; Figure 3 This is a schematic diagram of the real-time process fingerprint vector mapping process of the present invention; Figure 4 This is a schematic diagram of the dynamic decoupling control command generation process of the present invention; Figure 5 This is a schematic diagram of the posterior analysis and tolerance band optimization process of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] As attached Figure 1 and appendix Figure 2The adaptive control method based on ETEE thin film stretching process shown includes S1: constructing a dynamically updated multidimensional target process fingerprint based on material batch characteristics and initial process parameters, and setting a dynamically adjustable adaptive tolerance band for each dimension by initializing the reference fingerprint.

[0012] It should be specifically noted that the construction method of the multidimensional target process fingerprint includes small-batch pre-stretching tests and orthogonal tests. The dimensions are 12-dimensional, including 5 material parameter dimensions and 7 process parameter dimensions. The material parameter dimensions include number-average molecular weight, molecular weight distribution index, melt index, crystallinity and end-group fluorine content. The process parameter dimensions include preheating temperature, stretching temperature, MD stretching ratio, TD stretching ratio, stretching rate, cooling temperature and traction tension.

[0013] It should be further explained that the establishment of the benchmark fingerprint adopts "small batch pre-stretch test + orthogonal test". For each batch of newly arrived ETEE resin, 50g of sample is taken for small batch pre-stretch test. Different stretching temperatures, MD / TD stretching ratios, stretching rates, and preheating times are set. The key quality indicators of tensile strength (MD / TD), elongation at break (MD / TD), haze, and thickness uniformity (within ±5% error) of the film after stretching are collected simultaneously.

[0014] The orthogonal experimental design set four levels for five key process parameters: tensile temperature: 160℃, 170℃, 180℃, 190℃; MD tensile ratio: 2.0, 2.5, 3.0, 3.5; TD tensile ratio: 2.0, 2.5, 3.0, 3.5; tensile rate: 5mm / s, 10mm / s, 15mm / s, 20mm / s; preheating time: 30s, 60s, 90s, 120s.

[0015] Through grey relational analysis, material-process parameter combinations with a correlation degree ≥ 0.85 with quality indicators were selected as the initial baseline fingerprint vector for this batch. The dimensions were set to 12, including 5 material parameters and 7 process parameters. The 5 material parameters include number average molecular weight (Mn), molecular weight distribution index (PDI), melt index (MI), crystallinity, and end-group fluorine content. The 7 process parameters include preheating temperature, stretching temperature, MD stretching ratio, TD stretching ratio, stretching rate, cooling temperature, and traction tension.

[0016] A sliding window algorithm is introduced, with the window size set to 5 process cycles. The posterior optimization data of each cycle is incorporated in real time. When the mean quality index of the new cycle deviates from the historical mean within the window by more than 10%, the fingerprint is updated. Partial least squares regression (PLSR) is used to establish a material-process-quality mapping model. With the optimal quality index as the goal, the process parameter dimension of the target process fingerprint is updated by inversion. The material parameter dimension is updated once for each batch to ensure dynamic matching between fingerprint and material properties.

[0017] Unlike traditional fixed tolerances, a dynamic tolerance band is designed for each fingerprint dimension. Statistical process control (SPC) combined with process robustness analysis is used to calculate the process capability index Cp / Cpk for each dimension parameter. For material parameters, due to uncontrollable batch fluctuations, the tolerance band width is set to ±3σ based on the standard deviation of historical batch data (σ is the standard deviation of that parameter in historical batches). For process parameters, due to their high controllability, the tolerance band width is set to ±1.5σ based on the standard deviation under stable process conditions, and is dynamically adjusted in conjunction with material parameters. The tolerance band boundaries are marked with three colors: red (out of tolerance), yellow (warning), and green (normal), providing a basis for subsequent deviation assessment.

[0018] As attached Figure 3 As shown, S2: During the stretching process, the physical state data of the thin film is synchronously collected through a distributed sensor array, and the data is calculated and mapped in real time into a real-time process fingerprint vector corresponding to the process fingerprint dimension.

[0019] It should be specifically noted that the physical state data synchronous acquisition method includes deploying a "3-layer 5-point" distributed sensor network covering the preheating zone, stretching zone and cooling zone, and using a hierarchical preprocessing strategy to process the multi-physics field data. The processing methods include Kalman filtering, wavelet threshold denoising and timestamp alignment algorithm.

[0020] It should be further explained that a "3-layer, 5-point" distributed sensor network was designed to cover the preheating zone, stretching zone, and cooling zone of the stretching machine. Five key sampling points were deployed in each zone to simultaneously collect multi-physics field data, as detailed below: In the preheating zone, infrared thermal imagers and fiber optic grating sensors collect data on the surface temperature field distribution and internal stress of the thin film. Uneven preheating can lead to localized orientation differences during stretching, and internal stress directly affects the risk of subsequent fracture. In the stretching zone, laser thickness gauges, high-speed cameras, and torque sensors collect data on the film thickness distribution, wrinkle degree during stretching, stretching roller torque, and thickness uniformity, which are core quality indicators. Wrinkles are a precursor to stretching instability, and torque reflects the material's real-time stretching resistance. In the cooling zone, thermocouple arrays and capacitive humidity sensors collect data on the cooling temperature distribution and film surface humidity. The ETEE crystallization rate is sensitive to cooling temperature, and humidity affects the film's surface tension and subsequent processability.

[0021] To address the heterogeneity and noise interference of multi-sensor data, a hierarchical preprocessing strategy is adopted. For continuous data of temperature and thickness, Kalman filtering is used with a Q value of 0.01 and an R value of 0.1 to remove random noise. For non-stationary data of wrinkles and torque, wavelet threshold denoising is used with db4 wavelet and decomposition into 3 layers. For sensor synchronization error, a timestamp alignment algorithm is used with the trigger signal of the high-speed camera in the stretching zone as the reference, and the deviation is ≤1ms.

[0022] A deep learning model (CNN-LSTM) is used to extract data features. Multi-source data that has been processed by Kalman filtering, wavelet denoising, and time-series alignment is input into the model for classification. Spatial data is input into the CNN module in the form of a 256×256 pixel matrix, and temporal data is input into the LSTM module in the form of a sequence of 100 sampling points.

[0023] The CNN module performs three rounds of convolution on the spatial data using 3×3 convolutional kernels (32 kernels, stride 1). After each round, the data is reduced in dimensionality by ReLU activation and 2×2 max pooling, outputting a 6-dimensional spatial feature vector. The LSTM module performs 50 steps of temporal memory on the temporal data using 64 hidden units, compresses it into a 6-dimensional temporal feature vector through a fully connected layer, and concatenates them to obtain a 12-dimensional comprehensive feature vector. A fully connected layer is used to map the 12-dimensional feature vector to the 12 dimensions of the target process fingerprint one by one. The mapping error is minimized by the Adam optimizer (learning rate 0.001). After 500 rounds of iterative training, the mapping error is stabilized within ±2%, generating a real-time process fingerprint vector.

[0024] S3: Compare the real-time process fingerprint vector with the target process fingerprint dimension by dimension to generate a multi-dimensional deviation situation map, calculate the instantaneous deviation value of each dimension, and evaluate the stability of the current process state and the severity of the deviation situation by analyzing the evolution trend of the deviation in time and space and the coupling relationship between each dimension.

[0025] It should be specifically noted that the severity level, in particular, includes establishing a three-dimensional evaluation index system based on "deviation amplitude - trend slope - coupling strength". The severity level of the process status is divided into 1 to 5 levels according to whether the deviation of each dimension enters the warning area of ​​the adaptive tolerance zone, the spatiotemporal evolution trend slope of the deviation, and the number of strong coupling groups between dimensions.

[0026] It should be further explained that, using the 12-dimensional real-time process fingerprint vector and the target process fingerprint vector as input, a visualized deviation situation map is generated through "dimensional comparison + coupling analysis," and the instantaneous deviation value is calculated dimension by dimension. The formula is as follows: A heatmap is used to present the distribution of deviations in each dimension, with the horizontal axis representing time and the vertical axis representing the fingerprint dimension. The color depth indicates the magnitude of the deviation. An association matrix is ​​introduced to show the coupling relationship between dimensions, and dimensions with a correlation coefficient ≥ 0.7 are marked as "strongly coupled groups".

[0027] The introduction of an association matrix to illustrate the coupling relationship between dimensions involves using a 12-dimensional real-time process fingerprint vector as the analysis object. Real-time data from the past 10 process cycles is extracted, with a sampling frequency of 50-100Hz. 1000 sets of valid data are collected per cycle to ensure data coverage of different states of process stability, early warning, and adjustment. The Pearson correlation coefficient formula is used to calculate the linear correlation between any two dimensions. The specific formula is as follows: in , Real-time data in two dimensions. , The mean of the data in the corresponding dimension is used to calculate a 12×12 correlation coefficient matrix (the diagonal is the autocorrelation coefficient of each dimension).

[0028] The correlation coefficient matrix is ​​visualized as a heatmap. The darker the color, the larger the absolute value of the correlation coefficient. A correlation coefficient ≥ 0.7 is defined as "strong coupling", 0.5-0.7 as "medium coupling" and < 0.5 as "weak coupling". Strong coupling dimension groups are marked with red borders. The coupling mechanism is explained next to the matrix to provide an intuitive basis for subsequent dynamic decoupling control.

[0029] The spatiotemporal trend analysis employs the Holt-Winters exponential smoothing method. A smoothing coefficient α of 0.3 controls the current data weight, β of 0.2 controls the trend weight, and γ of 0.1 controls the seasonal weight. Since the stretching process has no significant seasonal fluctuations, γ is set to a relatively low value. The horizontal smoothing value reflects the current deviation baseline, the trend smoothing value reflects the rate of deviation change, and finally, the seasonal smoothing value is calculated to mitigate the impact of the lack of seasonal fluctuations in the process. This yields a smoothed sequence for the deviation in each dimension. Using nearly 100 sets of historical deviation data as initial input, the deviation evolution trend is predicted for the next 5 sampling periods (50ms). If the predicted deviation increases at a uniform rate, it is classified as a "gradual deviation"; if the instantaneous increase is ≥5%, it is classified as a "sudden deviation," providing a trend basis for assessing the stability of the process.

[0030] A three-dimensional evaluation index system of "deviation amplitude - trend slope - coupling strength" was established, and fuzzy comprehensive evaluation method was used to classify it into 5 severity levels: Level 1 (Normal): All dimensional deviations are within the green tolerance band, the trend slope is ≤0.1% / cycle, and there is no strong coupling deviation; Level 2 (Hint): There is exactly one dimension of bias entering the yellow tolerance band, the bias trend slope is in the range of 0.1% / cycle-0.3% / cycle, and there is no strong coupling bias; Level 3 (Warning): 2-3 dimensional deviations enter the yellow tolerance zone, with a trend slope of 0.3%-0.5% / cycle, indicating the existence of one set of strongly coupled deviations; Level 4 (Severe Warning): Four or more dimensions have deviations that enter the yellow tolerance zone (but have not exceeded the red tolerance zone), or one dimension has deviations that exceed the red tolerance zone (the remaining dimensions are all within the yellow tolerance zone), the deviation trend slope is in the range of 0.5% / cycle-0.8% / cycle, and there are 2-3 sets of strongly coupled deviations.

[0031] Level 5 (Emergency Warning): Two or more dimensional deviations exceed the red tolerance zone, the trend slope is ≥0.8% / cycle, and there are three or more sets of strongly coupled deviations. Immediate shutdown and adjustment are required.

[0032] As attached Figure 4 As shown, S4: Based on the deviation situation map, initiate the multi-objective optimization decision-making process, select the intensity level of control intervention according to the severity level of the situation, dynamically decouple the control commands according to the spatiotemporal distribution characteristics and coupling relationship of the deviation, generate adjustment strategies, and guide the real-time process fingerprint back to the adaptive tolerance band of the target fingerprint.

[0033] It should be specifically noted that, regarding the intensity level, the control intervention method is as follows: when the severity level is level 1 or 2, a mild intervention fine-tuning mode is adopted, with an adjustment range not exceeding ±3%; when the severity level is level 3 or 4, a moderate intervention combined adjustment mode is adopted, with synchronous adjustment for the strongly coupled dimension group, with an adjustment range of ±3% to ±8%; when the severity level is level 5, a severe intervention emergency intervention mode is adopted, prioritizing the reduction of the stretching rate and adjusting the key process parameters by a range not exceeding ±15%.

[0034] The adjustment strategy generation method includes establishing relevant transfer function models for process parameter dimensions with strong coupling relationships, designing dynamic decoupling compensation matrices based on the transfer function models to eliminate or reduce cross interference between parameters, constructing a multi-objective optimization function, solving it using a non-dominated sorting genetic algorithm, and selecting the final control adjustment strategy from the Pareto optimal solution set.

[0035] It should be further explained that a "three-level response" mechanism is established by matching the corresponding intensity of control intervention based on the severity level: Mild intervention (corresponding to levels 1-2): The "fine-tuning mode" is adopted, adjusting only a single weakly sensitive process parameter with an adjustment range of ≤±3% to avoid process fluctuations; Moderate intervention (corresponding to level 3-4): adopt the "combined adjustment mode", and make synchronous adjustments for the strongly coupled group, with an adjustment range of ±3% to ±8%; Severe intervention (corresponding to level 5): adopt the "emergency intervention mode", first reduce the stretching rate by 30%, and then adjust the key parameters according to priority (priority: stretching temperature > stretching ratio > preheating time), with an adjustment range of ≤ ±15%.

[0036] To address the strongly coupled dimensions, an improved dynamic matrix control (DMC) was employed for decoupling. A transfer function model for each process parameter and deviation (a first-order inertial model of stretching temperature → thickness deviation) was established. Synchronous data of stretching temperature and thickness deviation (-5% to +5%) over nearly 20 process cycles were selected, with a sampling frequency of 100Hz. 1000 sets of valid data were collected per cycle. Kalman filtering (Q=0.01, R=0.1) was used to remove outliers caused by equipment vibration and sensor noise, ensuring a data signal-to-noise ratio ≥30dB.

[0037] The expression for the first-order inertial model is: Where K is the gain (reflecting the magnitude of the effect of the stretching temperature change on the thickness deviation) and T is the time constant (reflecting the response lag of the thickness deviation to the temperature change). Through a step test (the stretching temperature is stepped from 180℃ to 185℃, and the thickness deviation is continuously monitored), the transition process data of the thickness deviation from the initial value to the steady state value are recorded.

[0038] The least squares method was used to fit the transient process data, and the results were calculated. (For every 1°C increase in temperature, the thickness deviation decreases by 0.02%) (Thickness deviation response lag of 2 seconds) Substitute the identified parameters into the model, and verify the goodness of fit of the model by comparing the model output with the actual thickness deviation data. This ensures that the model can reflect the dynamic relationship between stretching temperature and thickness deviation.

[0039] A decoupling compensation matrix is ​​introduced to eliminate cross-interference between parameters (the compensation coefficient is dynamically adjusted according to the coupling strength; when the correlation coefficient is ≥0.8, the compensation coefficient is set to 0.6). Specifically, this is for... The coupling terms in the model are designed using the diagonal dominance method to create a decoupling compensation matrix. For example, regarding the coupling of stretching temperature (parameter 1) and MD stretch ratio (parameter 2) on thickness deviation (deviation 1), in Set compensation item at position (1,2) , Let be the transfer function of MD stretch ratio-thickness deviation, with -0.3 as the compensation coefficient (dynamically adjusted based on the correlation coefficient of 0.8), so that the transfer function matrix of the decoupled system is... A near-diagonal matrix.

[0040] With the objectives of "minimizing deviation, minimizing adjustment range, and optimizing quality", a multi-objective optimization function is constructed: in , , The quality score is calculated by weighting tensile strength and uniformity indices, and is solved using a non-dominated sorting genetic algorithm. The number of iterations is set to 100 generations, the population size is set to 50, and the Pareto optimal solution is output as the adjustment strategy.

[0041] Tensile strength (MD / TD) is standardized using a "higher is better" principle. If the real-time value exceeds the target value, it is scored out of 100. The formula is as follows: ; Thickness uniformity is standardized using the "smaller the better" principle. If the real-time error is ≤1%, it is scored out of 100. The formula is as follows: .

[0042] The final scoring formula is: The scoring range is 0-100 points, with 80 points or above considered "excellent", 60-80 points considered "qualified", and below 60 points considered "unqualified". It serves as one of the core output indicators for multi-objective optimization.

[0043] The non-dominated sorting genetic algorithm uses the current process parameters as a benchmark and randomly generates 50 initial control strategies within an adaptive tolerance range of ±10%. For each generation of the population, the non-dominated levels are divided according to the dominance rule of "the smaller the deviation value, the smaller the adjustment range, and the higher the quality score" (the first level is the optimal solution with no other individuals dominating it). Calculate the crowding degree of each individual in each layer (reflecting the dispersion of solutions), prioritize retaining individuals with high crowding degree to avoid solutions being concentrated in local areas, and select 30% of high-quality individuals from the population as parents; For continuous parameters, simulated binary crossover (SBX) is used with a crossover probability of 0.8; for discrete parameters, single-point crossover is used to ensure the rationality of offspring parameters. The mutation probability is set to 0.05. The parameter values ​​are slightly perturbed to maintain population diversity. After 100 iterations, the first layer of non-dominated solutions forms a Pareto optimal solution set. According to the real-time process requirements (e.g., if thickness deviation is prioritized, the solution with the smallest uniformity error is selected; if adjustment cost is prioritized, the solution with the smallest adjustment range is selected), one optimal solution is selected as the final control adjustment strategy to ensure that the decision output is completed within 50ms and the real-time requirements of the process are met.

[0044] The adjustment strategy is transformed into control commands and transmitted to the actuator via industrial Ethernet. The execution delay is ≤50ms. Real-time fingerprints are collected synchronously after execution to form a closed loop of "decision-execution-feedback".

[0045] As attached Figure 5 As shown, S5: After a single or multiple consecutive process cycles, a post-hoc analysis is performed on the target process fingerprint, adjustment strategy and effect. By analyzing the correlation between the control strategy and fingerprint convergence in historical data, the adaptive tolerance band width of the target process fingerprint is dynamically optimized.

[0046] It should be specifically noted that the adaptive tolerance bandwidth is dynamically optimized by collecting full-process data, including the target process fingerprint, control command sequence, and final film quality index, after a single or multiple consecutive process cycles. A random forest algorithm is used to establish a correlation model between the control strategy, fingerprint convergence, and tolerance bandwidth. Based on the output of the correlation model, the tolerance bandwidth of fingerprint dimensions with low convergence efficiency is reduced, while the tolerance bandwidth of fingerprint dimensions with stable quality but frequent adjustments is increased.

[0047] It should be further explained that after each process cycle (approximately 30 minutes) or after 5 consecutive batches, three types of historical data are collected to construct an analysis library. The input data are the target process fingerprint, material batch characteristics, and initial process parameters; the process data are the real-time process fingerprint sequence, deviation trend diagram, and control command sequence; the output data are the film quality indicators (tensile strength, uniformity), fingerprint convergence time (time for real-time fingerprint to return to the tolerance band), and number of adjustments.

[0048] Two core indicators are used to evaluate the effectiveness of the control strategy, including convergence efficiency: fingerprint convergence time ≤10s is excellent, 10-20s is good, and >20s is inefficient; stability: the quality index fluctuation ≤±3% within 10 consecutive periods is stable. Inefficient or unstable strategies are marked as "strategies to be optimized".

[0049] A random forest algorithm is used to establish a correlation model of "control strategy-convergence-tolerance band". The tolerance band width is used as the input feature, and the convergence time and quality fluctuation are used as the output labels. For dimensions with low convergence efficiency, the tolerance band width is reduced to improve control accuracy; for dimensions with stable quality but frequent adjustments, the tolerance band width is expanded to reduce control cost.

[0050] When a fingerprint dimension has conflicting requirements of both "low convergence efficiency" and "stable quality but frequent adjustments", the tolerance band width is narrowed first according to the requirement of "low convergence efficiency". The standard for "good convergence efficiency" is "the time for real-time process fingerprint to return to the tolerance band is 10-20 seconds", and the standard for "frequent adjustments" is "the number of adjustments for this dimension is ≥5 times in a single process cycle". The optimized tolerance band is updated synchronously through the fingerprint update mechanism to achieve closed-loop optimization of the entire process.

[0051] An adaptive control system based on ETEE thin film stretching process includes a process fingerprint construction module: receiving material batch characteristic data and initializing a baseline process fingerprint model based on predefined process parameters, and assigning an adaptive tolerance band to each dimension.

[0052] Multimodal sensing and fusion module: It communicates with the distributed sensor array, synchronously collects the physical state data of the thin film during the stretching process, and calculates and maps it in real time into a real-time process fingerprint vector corresponding to the dimension of the process fingerprint construction module.

[0053] Fingerprint Deviation Situation Assessment Module: The received real-time process fingerprint vector is compared with the target process fingerprint dimension by dimension to generate a multi-dimensional deviation situation map. By calculating the instantaneous deviation value of each dimension and analyzing the evolution trend of the deviation in time and space and the coupling relationship between each dimension, the stability of the current process status and the severity level of the deviation situation are assessed.

[0054] Adaptive strategy decision-making module: Based on the multi-dimensional deviation situation map and severity level, it initiates a multi-objective optimization decision-making process, selects the intensity level of control intervention according to the severity level, and dynamically decouples control commands and generates adjustment strategies based on the spatiotemporal distribution characteristics and coupling relationship of the deviation.

[0055] Closed-loop evolution learning module: Connected to the process fingerprint construction module and the adaptive strategy decision module, after the process cycle ends, it dynamically optimizes and outputs the adjustment command of the adaptive tolerance band width of the target process fingerprint by analyzing the correlation between the control strategy and the convergence of the fingerprint in historical data, and continuously optimizes the control system.

[0056] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive control method based on ETEE thin film stretching process, characterized in that, include: S1: Based on the material batch characteristics and initial process parameters, construct a dynamically updated multi-dimensional target process fingerprint. By initializing the baseline fingerprint, set a dynamically adjustable adaptive tolerance band for each dimension. S2: During the stretching process, the physical state data of the film is collected synchronously through a distributed sensor array, and the data is calculated and mapped in real time into a real-time process fingerprint vector corresponding to the process fingerprint dimension. S3: Compare the real-time process fingerprint vector with the target process fingerprint dimension by dimension to generate a multi-dimensional deviation situation map, calculate the instantaneous deviation value of each dimension, and evaluate the stability of the current process state and the severity of the deviation situation by analyzing the evolution trend of the deviation in time and space and the coupling relationship between each dimension. S4: Based on the deviation situation map, initiate a multi-objective optimization decision-making process, select the intensity level of control intervention according to the severity level of the situation, dynamically decouple control commands according to the spatiotemporal distribution characteristics and coupling relationship of the deviation, generate adjustment strategies, and guide the real-time process fingerprint back to the adaptive tolerance band of the target fingerprint. S5: After a single or multiple consecutive process cycles, perform a post-hoc analysis on the target process fingerprint, adjustment strategy, and its effects. By analyzing the correlation between the control strategy and fingerprint convergence in historical data, dynamically optimize the adaptive tolerance band width of the target process fingerprint.

2. The adaptive control method based on ETEE thin film stretching process according to claim 1, characterized in that: The multidimensional target process fingerprint is constructed by a method including small-batch pre-stretching tests and orthogonal tests. The dimensions are 12, including 5 material parameter dimensions and 7 process parameter dimensions. The material parameter dimensions include number-average molecular weight, molecular weight distribution index, melt index, crystallinity and end-group fluorine content. The process parameter dimensions include preheating temperature, stretching temperature, MD stretching ratio, TD stretching ratio, stretching rate, cooling temperature and traction tension.

3. The adaptive control method based on ETEE thin film stretching process according to claim 1, characterized in that: The physical state data is synchronously acquired by deploying a "3-layer, 5-point" distributed sensor network covering the preheating zone, stretching zone, and cooling zone. A hierarchical preprocessing strategy is used to process the multi-physics data. The processing methods include Kalman filtering, wavelet threshold denoising, and timestamp alignment algorithm.

4. The adaptive control method based on ETEE thin film stretching process according to claim 1, characterized in that: The severity level is specifically defined by establishing a three-dimensional evaluation index system based on "deviation amplitude - trend slope - coupling strength". The severity level of the process status is divided into 1 to 5 levels according to whether the deviation of each dimension enters the warning area of ​​the adaptive tolerance zone, the spatiotemporal evolution trend slope of the deviation, and the number of strong coupling groups between dimensions.

5. The adaptive control method based on ETEE thin film stretching process according to claim 1, characterized in that: The intensity level control intervention method is as follows: when the severity level is level 1 or 2, a mild intervention fine-tuning mode is adopted, with an adjustment range not exceeding ±3%; when the severity level is level 3 or 4, a moderate intervention combined adjustment mode is adopted, and synchronous adjustments are made to the strongly coupled dimension group, with an adjustment range of ±3% to ±8%; when the severity level is level 5, a severe intervention emergency intervention mode is adopted, prioritizing the reduction of the stretching rate and adjusting the key process parameters by no more than ±15%.

6. The adaptive control method based on ETEE thin film stretching process according to claim 1, characterized in that: The adjustment strategy generation method includes establishing relevant transfer function models for process parameter dimensions with strong coupling relationships, designing dynamic decoupling compensation matrices based on the transfer function models to eliminate or reduce cross interference between parameters, constructing a multi-objective optimization function, solving it using a non-dominated sorting genetic algorithm, and selecting the final control adjustment strategy from the Pareto optimal solution set.

7. The adaptive control method based on ETEE thin film stretching process according to claim 1, characterized in that: The adaptive tolerance band width is specifically optimized by a dynamic method that includes collecting full-process data, including the target process fingerprint, control command sequence, and final film quality index, after a single or multiple consecutive process cycles. A random forest algorithm is used to establish a correlation model between the control strategy, fingerprint convergence, and tolerance band width. Based on the output of the correlation model, the tolerance band width is reduced for fingerprint dimensions with low convergence efficiency and expanded for fingerprint dimensions with stable quality but frequent adjustments.

8. An adaptive control system based on ETEE thin film stretching process, used to implement the adaptive control method based on ETEE thin film stretching process as described in any one of claims 1-7, characterized in that, include: Process fingerprint construction module: Receives material batch characteristic data and initializes a baseline process fingerprint model based on predefined process parameters, assigning an adaptive tolerance band to each dimension; Multimodal sensing and fusion module: It communicates with the distributed sensor array, synchronously collects the physical state data of the thin film during the stretching process, and calculates and maps it in real time into a real-time process fingerprint vector corresponding to the dimension of the process fingerprint construction module; Fingerprint Deviation Situation Assessment Module: The received real-time process fingerprint vector is compared with the target process fingerprint dimension by dimension to generate a multi-dimensional deviation situation map. By calculating the instantaneous deviation value of each dimension and analyzing the evolution trend of the deviation in time and space and the coupling relationship between each dimension, the stability of the current process state and the severity level of the deviation situation are assessed. Adaptive strategy decision-making module: Based on the multi-dimensional deviation situation map and severity level, it initiates a multi-objective optimization decision-making process, selects the intensity level of control intervention according to the severity level, and dynamically decouples control commands and generates adjustment strategies based on the spatiotemporal distribution characteristics and coupling relationship of the deviation. Closed-loop evolution learning module: Connected to the process fingerprint construction module and the adaptive strategy decision module, after the process cycle ends, it dynamically optimizes and outputs the adjustment command of the adaptive tolerance band width of the target process fingerprint by analyzing the correlation between the control strategy and the convergence of the fingerprint in historical data, and continuously optimizes the control system.

Images