A pet film preparation control method based on data analysis

Abstract

The application discloses a PET film preparation control method based on data analysis. The method obtains real-time process parameter data, obtains a structured data set, adopts a pre-established machine learning algorithm to train a parameter performance relationship model, further obtains a real-time film surface image of the PET film, generates defect distribution data and a defect severity quantization result, analyzes the interaction of the defect distribution data, the defect severity quantization result and the process parameter data, and obtains an optimized parameter combination. Finally, the optimized parameter combination is adopted to update a PET film preparation equipment control system and adjust the process parameter data. The application realizes real-time collection of process parameters, optimizes parameter configuration by using a machine learning algorithm, and combines real-time detection of defects and the interaction of process parameter data to obtain a PET film preparation scheme meeting high performance requirements.

Description

Technical Field

[0001] This invention relates to the field of thin film preparation technology, and in particular to a data analysis-based method for controlling the preparation of PET films. Background Technology

[0002] Polyethylene terephthalate (PET) film, as a high-performance polymer material, is widely used in food packaging, electronic devices, and optical displays due to its excellent mechanical strength, thermal stability, and transparency. However, with the increasing demands on film performance from downstream applications (such as thickness uniformity, surface defect control, and mechanical properties), the limitations of traditional PET film manufacturing processes are becoming increasingly apparent.

[0003] Existing preparation methods mostly rely on experience to adjust or fix parameter combinations, which makes it difficult to adapt to dynamically changing process requirements. Furthermore, the detection and handling of defects mainly depend on manual experience, which is inefficient and prone to introducing errors. Summary of the Invention

[0004] The technical solution adopted by this invention to solve its technical problem is: a data analysis-based method for controlling the preparation of PET film, comprising the following steps:

[0005] S1: Acquire real-time process parameter data from sensors in the PET film preparation equipment and store it in the database to obtain a structured dataset;

[0006] S2: Using a pre-established machine learning algorithm, a parameter performance relationship model is trained based on the structured dataset. The parameter performance relationship model is a model showing the relationship between the process parameter data and the performance of the PET film.

[0007] S3: Obtain a real-time image of the PET film surface, process the image using a convolutional neural network algorithm, identify the type and location of surface defects, and generate defect distribution data and defect severity quantification results;

[0008] S4: By analyzing the interaction between the defect distribution data, the defect severity quantification results, and the process parameter data through the parameter performance relationship model, an optimized parameter combination is obtained;

[0009] S5: Using the optimized parameter combination, update the control system of the PET film preparation equipment and adjust the process parameter data.

[0010] Furthermore, in the data analysis-based PET film preparation control method described in this invention, the process parameter data includes: temperature data, stretching rate data, and cooling rate data.

[0011] Furthermore, in the data analysis-based PET film preparation control method of the present invention, step S1 includes the following steps:

[0012] S101: Obtain real-time process parameter data from sensors in the PET film preparation equipment;

[0013] S102: Determine whether the process parameter data meets the preset threshold range. If it does, mark it as valid data and obtain a valid dataset.

[0014] S103: Based on the valid dataset, the temperature data, the stretching rate and the cooling rate are classified and stored according to timestamps using database storage technology, and a unique identifier is generated for each parameter;

[0015] S104: Confirm that the stored procedure is complete and obtain the structured dataset.

[0016] Furthermore, in the data analysis-based PET film preparation control method of this invention, the training of the parameter performance relationship model based on the structured dataset includes:

[0017] The process parameter data is obtained from the structured dataset. The process parameter data is normalized using data preprocessing techniques to generate a first dataset. The first dataset is determined to meet the training requirements to obtain a standardized dataset. Based on the standardized dataset, a parameter performance relationship model is trained using a logistic regression algorithm.

[0018] Furthermore, in the data analysis-based PET film preparation control method of the present invention, step S3 includes the following steps:

[0019] S301: Acquire a real-time image of the PET film surface, process the image using a convolutional neural network algorithm, identify the type and location of surface defects, and generate defect distribution data;

[0020] S302: Calculate the severity score of the defect using the defect distribution data to obtain a quantitative result of the defect severity.

[0021] Furthermore, in the data analysis-based PET film preparation control method of the present invention, step S301 includes the following steps:

[0022] S3011: Acquire a real-time image of the PET film surface, and use mean filtering technology to denoise the image to generate a first image set;

[0023] S3012: If the resolution of the first image set meets the preset threshold, then a standardized image set is obtained;

[0024] S3013: Based on the standardized image set, feature extraction is performed using a convolutional neural network algorithm to identify the defect types on the membrane surface and generate a first defect type set;

[0025] S3014: If the first defect type set satisfies the preset defect classification threshold, then the defect feature set is obtained;

[0026] S3015: The defect feature set is processed using image segmentation technology to determine the specific location of the defects on the membrane surface and generate a first defect location set;

[0027] S3016: If the positioning accuracy of the first defect location set meets the preset threshold, then the defect location set is obtained;

[0028] S3017: Integrate the defect feature set and the defect location set through data formatting technology to generate defect distribution data containing defect types and locations.

[0029] Furthermore, in the data analysis-based PET film preparation control method of the present invention, step S302 includes the following steps:

[0030] S3021: Obtain the defect type and location information from the defect distribution data, perform structured processing on the defect distribution data using data parsing technology, and generate a first structured dataset from the processed defect distribution data whose integrity meets a preset threshold;

[0031] S3022: Based on the first structured dataset, feature extraction technology is used to separate the defect type and location information. If the separation degree of the separated feature set meets the preset classification threshold, a first feature set is generated.

[0032] S3023: Using a pre-established scoring standard table, a weighted calculation method is used to calculate the severity score based on the defect type and location information in the first feature set. If the score consistency meets a preset threshold, a first score set is generated.

[0033] S3024: The first scoring set is integrated using data formatting technology to generate a quantitative result dataset containing the defect type, location information, and severity score;

[0034] S3025: If the format of the quantification result dataset meets the preset business objectives, then the final defect severity quantification result is generated.

[0035] Furthermore, in the data analysis-based PET film preparation control method of the present invention, step S4 includes the following steps:

[0036] S401: Through the parameter performance relationship model, analyze the interaction between the defect distribution data, the defect severity quantification results, and the process parameter data to generate performance prediction adjustment values ​​and process parameter optimization boundary conditions;

[0037] S402: Based on the performance prediction adjustment value and the process parameter optimization boundary conditions, the optimized value of the process parameter data is calculated using the gradient descent method of the machine learning algorithm;

[0038] S403: Obtain the optimized parameter combination through the optimized value.

[0039] Furthermore, in the data analysis-based PET film preparation control method of the present invention, step S403 specifically includes the following steps:

[0040] S4031: Using a pre-established defect solution database, based on the defect distribution data, the quantification results of the severity of prime number defects, and the performance prediction adjustment value, match them with the defect type index and historical data in the database to determine the defect processing parameters;

[0041] S4032: The optimized parameter combination is obtained by combining the optimized value with the defect handling parameters.

[0042] Furthermore, in the data analysis-based PET film preparation control method of the present invention, step S403 specifically includes the following steps:

[0043] S40301: By using the parameter performance relationship model and inputting the process parameter data, the membrane performance indicators, including mechanical properties, transparency, and chemical stability, are predicted to obtain the performance prediction results.

[0044] S40302: If the performance prediction result is lower than the preset membrane quality threshold, the deviation between the process parameters and historical data is analyzed through the parameter performance relationship model to generate a quantitative value of performance deviation.

[0045] S40303: Based on the quantified performance deviation value, the optimized value is recalculated using the gradient descent method of a machine learning algorithm to obtain the optimized parameter combination.

[0046] The present invention offers the following advantages: It discloses a data analysis-based control method for PET film fabrication. This method acquires real-time process parameter data to obtain a structured dataset, and uses a pre-established machine learning algorithm to train a parameter performance relationship model. Furthermore, it acquires real-time images of the PET film surface, generating defect distribution data and quantified defect severity results. Then, it analyzes the interaction between the defect distribution data, the quantified defect severity results, and the process parameter data to obtain an optimized parameter combination. Finally, it uses this optimized parameter combination to update the PET film fabrication equipment control system and adjust the process parameter data. This invention, by acquiring process parameters in real-time, optimizing parameter configuration using machine learning algorithms, and combining real-time defect detection with the interaction of process parameter data, achieves a PET film fabrication scheme that meets high-performance requirements. Attached Figure Description

[0047] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0048] Figure 1 This is a schematic diagram of the steps of a data analysis-based PET film preparation control method according to the present invention;

[0049] Figure 2 This is a detailed schematic diagram of step S1 of a data analysis-based PET film preparation control method of the present invention;

[0050] Figure 3 This is a detailed schematic diagram of step S3 of the PET film preparation control method based on data analysis of the present invention;

[0051] Figure 4 This is a detailed schematic diagram of step S301 of the PET film preparation control method based on data analysis of the present invention;

[0052] Figure 5 This is a detailed schematic diagram of step S302 of the PET film preparation control method based on data analysis of the present invention;

[0053] Figure 6 This is a detailed schematic diagram of step S4 of the PET film preparation control method based on data analysis of the present invention;

[0054] Figure 7 This is a detailed schematic diagram of step S403 of the first embodiment of the data analysis-based PET film preparation control method of the present invention;

[0055] Figure 8 This is a detailed schematic diagram of step S403 of the second embodiment of the PET film preparation control method based on data analysis of the present invention. Detailed Implementation

[0056] To provide a clearer understanding of the technical features, objectives, and effects of this invention, specific embodiments are now described in detail with reference to the accompanying drawings. In the following description, it should be understood that the orientations or positional relationships indicated by terms such as "front," "rear," "upper," "lower," "left," "right," "longitudinal," "horizontal," "vertical," "horizontal," "top," "bottom," "inner," "outer," "head," and "tail" are based on the orientations or positional relationships shown in the accompanying drawings, and are constructed and operated in a specific orientation. They are only for the convenience of describing this technical solution and do not indicate that the device or element referred to must have a specific orientation; therefore, they should not be construed as limitations on this invention.

[0057] It should also be noted that, unless otherwise explicitly specified and limited, terms such as "installation," "connection," "linking," "fixing," and "setting" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. When an component is referred to as being "on" or "below" another component, the component can be located "directly" or "indirectly" on the other component, or there may be one or more intermediary components. The terms "first," "second," "third," etc., are only for the convenience of describing this technical solution and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, features defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0058] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0059] Specifically, this invention provides a data analysis-based method for controlling the preparation of PET films, such as... Figure 1 As shown, the process includes the following steps (S1, S2, S3, S4, and S5):

[0060] Step S1: Obtain real-time process parameter data from sensors in the PET film preparation equipment and store it in the database to obtain a structured dataset.

[0061] Specifically, real-time process parameter data acquired from sensors in the PET film fabrication equipment includes temperature, stretching rate, and cooling rate. It's important to note that, compared to other parameters (such as extrusion pressure and roller speed), temperature, stretching rate, and cooling rate are directly related to the material's intrinsic properties (crystallinity, orientation, phase) and macroscopic properties (strength, transparency, thermal stability). Furthermore, their data can be acquired online at high frequencies (1-100Hz) using infrared thermal imagers, laser velocimeters, and thermocouples, providing actionable input dimensions for data-driven real-time optimization. Other parameters (such as humidity and vacuum level) are typically used as auxiliary variables in the secondary control loop, either due to large measurement lag or low influence weight.

[0062] Furthermore, such as Figure 2 As shown, step S1 includes the following steps:

[0063] S101: Obtain real-time process parameter data from sensors in the PET film preparation equipment;

[0064] S102: Determine whether the process parameter data meets the preset threshold range. If it does, mark it as valid data and obtain a valid dataset.

[0065] S103: Based on the valid dataset, use database storage technology to classify and store temperature data, stretching rate and cooling rate according to timestamps, and generate a unique identifier for each parameter;

[0066] S104: Confirm that the stored procedure is complete and obtain the structured dataset.

[0067] For example, the sensor collects temperature data, stretching rate, and cooling rate in real time, and the sampling frequency needs to be set to ensure data accuracy.

[0068] For example, in thin film processing scenarios, the sampling frequency is set to once per second. The temperature sensor's acquisition range is 20-1000℃, the stretching rate sensor monitors 0-10 mm / s, and the cooling rate sensor records 0-50℃ / min. Preset threshold ranges are 500-800℃ for temperature, 2-8 mm / s for stretching rate, and 10-40℃ / min for cooling rate. Data meeting the thresholds is marked as valid, and outliers such as a temperature of 100℃ or a stretching rate of 15 mm / s are removed, resulting in a valid dataset. This process ensures data quality and provides a reliable foundation for subsequent analysis.

[0069] In one possible implementation, the valid dataset is stored categorized using database storage technology. A relational database such as MySQL is used, with each record stored by timestamp, and temperature, stretching rate, and cooling rate each generating a unique identifier, such as T_ID, S_ID, and C_ID.

[0070] For example, a timestamp of 2025-04-21 10:00:00 records a temperature of 600℃, a tensile rate of 5mm / s, and a cooling rate of 30℃ / min, which, after storage, forms a structured dataset. This method facilitates data retrieval and traceability, improving data management efficiency.

[0071] Specifically, the parameter correlation analysis is based on a structured dataset and uses a logistic regression algorithm to explore the dynamic relationship between temperature and stretching rate and cooling rate.

[0072] For example, analysis showed that increased temperature may lead to an increase in stretching rate, with a correlation coefficient of 0.85, exceeding the preset threshold of 0.7, generating a parameter correlation matrix. This matrix reveals the interactions between parameters, providing a basis for process optimization, such as adjusting the temperature to control the stretching rate and optimize film performance.

[0073] Preferably, the data consistency check is based on the correlation matrix to identify abnormal data points.

[0074] For example, a record showing a temperature of 700℃ and a tensile rate of 1 mm / s deviates from the correlation pattern and is marked as an anomaly. Using time series analysis techniques, such as the moving average method, the anomaly is corrected to a tensile rate of 4 mm / s. The corrected dataset undergoes integrity verification to ensure there are no missing or redundant records. This process improves data reliability and provides accurate support for subsequent process analysis.

[0075] For example, after correcting abnormal data, the processing technology can optimize parameter settings based on the final structured dataset, such as maintaining a temperature of 650℃, a stretching rate of 6 mm / s, and a cooling rate of 35℃ / min to improve material toughness. Time series analysis can also predict parameter trends, allowing for advance process adjustments and reducing production defect rates.

[0076] Understandably, the implementation of the aforementioned technologies forms a closed loop, from data collection to storage, analysis, and correction, ensuring high-quality data and process stability. Each stage supports the others, collectively improving processing efficiency and product quality, reducing resource waste, and enhancing production controllability.

[0077] Step S2: Using a pre-established machine learning algorithm, a parameter performance relationship model is trained based on a structured dataset. The parameter performance relationship model is a model showing the relationship between process parameter data and PET film performance.

[0078] Specifically, the parametric performance relationship model is a model relating temperature, stretching rate, cooling rate to mechanical properties, transparency, and chemical stability, used to determine the nonlinear mapping relationship between temperature, stretching rate, cooling rate, and mechanical properties, transparency, and chemical stability.

[0079] Furthermore, the training of the parameter performance relationship model based on the structured dataset in step S2 includes:

[0080] Process parameter data is obtained from a structured dataset. Data preprocessing techniques are used to normalize the process parameter data to generate the first dataset. The first dataset is determined to meet the training requirements to obtain a standardized dataset. Based on the standardized dataset, a logistic regression algorithm is used to train a parameter performance relationship model.

[0081] For example, when obtaining temperature, stretching rate, and cooling rate data from a structured dataset, data integrity must be ensured.

[0082] For example, in the thin film preparation scenario, the dataset contains records with the timestamp 2025-04-21 10:00:00, a temperature of 550℃, a stretching rate of 4 mm / s, and a cooling rate of 25℃ / min. Data preprocessing employs normalization techniques to map the temperature range of 20-1000℃, the stretching rate of 0-10 mm / s, and the cooling rate of 0-50℃ / min to the 0-1 interval.

[0083] For example, 550℃ is normalized to 0.54, and the tensile rate of 4 mm / s is normalized to 0.4, generating the first dataset. This process ensures consistent data dimensions, facilitating model training.

[0084] In one possible implementation, determining whether the first dataset meets the training requirements requires checking the data distribution and missing values.

[0085] For example, if a temperature record is missing in a scanned dataset, it is filled with 600℃ using mean interpolation to ensure dataset integrity and create a standardized dataset. This dataset provides a unified format for subsequent modeling.

[0086] Step S3: Acquire real-time images of the PET membrane surface, process the images using a convolutional neural network algorithm, identify the types and locations of surface defects, and generate defect distribution data and quantitative results of defect severity.

[0087] Furthermore, such as Figure 3 As shown, step S3 includes the following steps (S301 and S302):

[0088] S301: Acquire a real-time image of the PET film surface, process the image using a convolutional neural network algorithm, identify the type and location of surface defects, and generate defect distribution data;

[0089] S302: Calculate the severity score of the defect using the defect distribution data to obtain a quantitative result of the defect severity.

[0090] Furthermore, such as Figure 4As shown, step S301 includes the following steps (S3011, S3012, S3013, S3014, S3015, S3016 and S3017):

[0091] S3011: Acquire real-time images of the PET film surface, and use mean filtering technology to denoise the images to generate the first image set;

[0092] S3012: If the resolution of the first image set meets the preset threshold, then a standardized image set is obtained;

[0093] S3013: Based on the standardized image set, feature extraction is performed using a convolutional neural network algorithm to identify the defect types on the membrane surface and generate a first defect type set;

[0094] S3014: If the first defect type set satisfies the preset defect classification threshold, then the defect feature set is obtained;

[0095] S3015: The defect feature set is processed using image segmentation technology to determine the specific location of the defects on the membrane surface and generate a first defect location set;

[0096] S3016: If the positioning accuracy of the first defect location set meets the preset threshold, then the defect location set is obtained;

[0097] S3017: Integrate the defect feature set and the defect location set through data formatting technology to generate defect distribution data containing defect types and locations.

[0098] For example, in a thin film fabrication scenario, the production line acquires real-time images of the film surface for quality monitoring. Mean filtering technology removes noise interference by smoothing the image pixel values. Its principle is to take the average value of each pixel's neighborhood to generate a smooth image.

[0099] For example, if the acquired image resolution is 1920x1080, and the grayscale value of pixel (100, 100) abruptly changes to 200, while the grayscale values ​​of other pixels in the neighborhood are approximately 150, after mean filtering, the grayscale value of this pixel is adjusted to 152, generating the first image set. This method is suitable for high-noise environments and ensures the reliability of image data.

[0100] In one possible implementation, the resolution of the first image set is checked to see if it meets a preset threshold, such as requiring a resolution of no less than 1280x720. If it does, a standardized image set is generated through normalization processing.

[0101] For example, the image grayscale values ​​are mapped from 0-255 to 0-1, and the grayscale value 152 is normalized to 0.596, unifying the data format and facilitating subsequent analysis.

[0102] Specifically, convolutional neural networks are used to extract features from a standardized image set to identify the types of defects on the membrane surface. The principle is to extract texture and shape features from the image through multiple layers of convolution and pooling operations.

[0103] For example, the network identifies two types of defects in an image: scratches and bubbles, generating a first defect type set. If the classification accuracy is higher than 90%, a preset threshold is met, forming a defect feature set. This method can efficiently distinguish between multiple defect types.

[0104] It should be noted that image segmentation technology is used to process defect feature sets and determine the specific location of defects. Its principle is to divide the image into multiple regions and locate the pixel region where the defect is located.

[0105] For example, after segmentation, if the scratch is located at image coordinates (200, 300) to (250, 350) and the bubble is located at (400, 500), a first defect location set is generated. If the positioning accuracy error is less than 5 pixels, the threshold is met, and a defect location set is formed. This technique improves the accuracy of defect positioning.

[0106] In one embodiment, data formatting technology integrates a defect feature set and a defect location set to generate defect distribution data.

[0107] For example, the type and location information of scratches and bubbles can be formatted into a JSON structure, including the defect name and coordinate range. If the data format meets business requirements, such as supporting parsing by downstream analysis software, a final defect distribution dataset can be generated. This method ensures a clear data structure and facilitates process adjustments.

[0108] Understandably, the above process, through denoising, feature extraction, defect localization, and data integration, forms a complete solution for membrane surface defect analysis. Each step supports the others, ensuring a logically rigorous process from image acquisition to defect distribution.

[0109] For example, mean filtering ensures image quality, convolutional neural networks accurately classify defects, image segmentation precisely locates defects, and data formatting facilitates process optimization. This multi-stage collaborative approach improves the efficiency of quality monitoring in thin film preparation.

[0110] For example, when acquiring multi-source sensor data from an IoT framework, the first step is to process the collected abnormal temperature fluctuations using data cleaning techniques. For instance, a sliding window algorithm is used to smooth the temperature data with a window size of 5, removing abnormal fluctuations and retaining data within the normal range. Next, the peak vibration frequencies are analyzed. A Fast Fourier Transform (FFT) algorithm is used to convert the time-domain signal to the frequency-domain signal, extracting the main frequency components. For example, a peak frequency of 50Hz is identified as a normal vibration frequency of the device. Then, current abrupt changes are detected. A differential algorithm is used to calculate the rate of change of current, with a threshold of 0.5A / s. When the rate of change exceeds the threshold, the abrupt change time point is recorded. For example, a current abrupt change at 10:05:23 is marked as an abnormal event. Finally, the cleaned temperature, vibration, and current data are integrated into a real-time operating dataset for the device. Timestamp alignment ensures data consistency; for example, the temperature, vibration frequency, and current value at 10:05:23 are stored together to form a complete operating status record. The entire process is implemented through automated scripts, ensuring efficient and accurate data processing.

[0111] Furthermore, such as Figure 5 As shown, step S302 includes the following steps (S3021, S3022, S3023, S3024 and S3025):

[0112] S3021: Obtain the defect type and location information from the defect distribution data, perform structured processing on the defect distribution data using data parsing technology, and generate a first structured dataset from the processed defect distribution data whose integrity meets a preset threshold;

[0113] S3022: Based on the first structured dataset, feature extraction technology is used to separate the defect type and location information. If the separation degree of the separated feature set meets the preset classification threshold, a first feature set is generated.

[0114] S3023: Using a pre-established scoring standard table, a weighted calculation method is used to calculate the severity score based on the defect type and location information in the first feature set. If the score consistency meets a preset threshold, a first score set is generated.

[0115] S3024: The first scoring set is integrated using data formatting technology to generate a quantitative result dataset containing the defect type, location information, and severity score;

[0116] S3025: If the format of the quantification result dataset meets the preset business objectives, then the final defect severity quantification result is generated.

[0117] For example, in the context of thin film fabrication, defect distribution data includes defect type and location information, and data parsing techniques are used for structured processing.

[0118] For example, defect distribution data is stored in JSON format, including scratches located at coordinates (200, 300) to (250, 350) and bubbles located at (400, 500). Data parsing technology extracts fields using scripts to ensure information integrity. If the missing data rate after parsing is less than 5%, a first structured dataset is generated. This dataset is stored in tabular form, with columns including defect name and coordinate range, and rows corresponding to each defect instance, facilitating subsequent processing.

[0119] In one possible implementation, feature extraction techniques separate defect types and spatial features from a first structured dataset.

[0120] Specifically, defect types such as scratches or bubbles are extracted through classification labels, while spatial features such as the area and shape of defects are generated through geometric analysis.

[0121] For example, a scratch has an area of ​​500 pixels and a long, thin strip shape; a bubble has an area of ​​300 pixels and a circular shape. If the classification accuracy of the separated feature set is higher than 85%, a first feature set is generated. This set is represented in vector form, with each defect corresponding to a type label and spatial feature description, facilitating scoring.

[0122] It should be noted that the pre-established scoring criteria table is used to assess the severity of defects.

[0123] For example, the standard table stipulates that for every 100 pixels increase in the scratch area, 1 point is added to the score; if a bubble is close to the center of the membrane, 2 points are added to the score.

[0124] In one embodiment, the weighted calculation method assigns weights of 0.6 and 0.4 to scratches and bubbles, respectively. Assuming a scratch area of ​​500 pixels receives a score of 5, and a bubble near the center receives a score of 3, the weighted total score is 4.2. If the scoring consistency error is less than 10%, a first score set is generated. This set records the score for each defect for quantitative analysis.

[0125] Preferably, the data formatting technique integrates the first scoring set to generate a quantitative result dataset.

[0126] For example, the defect type, spatial features such as area and location, and severity score (4.2) can be formatted into a CSV file containing columns for: defect name, area, coordinates, and score. If the format meets business requirements, such as compatibility with quality analysis software, a final defect severity quantification dataset is generated. This dataset has a clear structure and supports process optimization decisions.

[0127] Understandably, the above process, from data analysis to result quantification, is interconnected. Data analysis ensures information completeness, feature extraction separates key attributes, scoring criteria quantify severity, and formatting facilitates application. Each step is logically rigorous, collectively supporting the high efficiency of thin-film quality monitoring.

[0128] Step S4: Analyze the interaction between the defect distribution data, the defect severity quantification results, and the process parameter data using the parameter performance relationship model to obtain an optimized parameter combination.

[0129] Furthermore, such as Figure 6 As shown, step S4 includes the following steps (S401, S402, and S403):

[0130] S401: Through the parameter performance relationship model, analyze the interaction between the defect distribution data, the defect severity quantification results, and the process parameter data to generate performance prediction adjustment values ​​and process parameter optimization boundary conditions;

[0131] S402: Based on the performance prediction adjustment value and the process parameter optimization boundary conditions, the optimized value of the process parameter data is calculated using the gradient descent method of the machine learning algorithm;

[0132] S403: Obtain the optimized parameter combination through the optimized value.

[0133] In the first embodiment of the present invention, as Figure 7 As shown, S403 specifically includes the following steps (S4031 and S4032):

[0134] S4031: Using a pre-established defect solution database, based on the defect distribution data, the quantification results of the severity of prime number defects, and the performance prediction adjustment value, match them with the defect type index and historical data in the database to determine the defect processing parameters;

[0135] S4032: The optimized parameter combination is obtained by combining the optimized value with the defect handling parameters.

[0136] For example, in thin film fabrication scenarios, defect distribution data, severity scores, and real-time process parameters such as temperature need to be structured. Data parsing techniques can extract JSON-formatted data via scripts, including defect locations such as coordinates (150, 200), severity scores of 3.5, and temperatures of 230°C. The script validates each field, and if the missing rate is less than 3%, a first structured dataset is generated. This dataset is stored in tabular form, with columns for defect coordinates, scores, and temperatures, and rows corresponding to individual defect instances, facilitating correlation analysis.

[0137] It should be noted that data integrity checks ensure the reliability of subsequent analysis, for example, by statistically analyzing the proportion of missing values ​​and removing outlier data.

[0138] In one possible implementation, feature extraction techniques separate the feature correlations between defect location, score, and temperature from a first structured dataset.

[0139] Specifically, location information is used to generate region labels through cluster analysis, such as near the edge or center; ratings are extracted by dividing intervals to obtain high and low levels; and temperature is normalized to generate continuous features. If correlation analysis shows that the correlation among the three is higher than 80%, a first feature association set is generated. This set is stored in matrix form, with each row representing a defect instance and each column containing the location label, rating level, and temperature value, supporting regression analysis.

[0140] Understandably, a high correlation between features indicates that temperature may affect defect distribution, which is beneficial for accurate prediction.

[0141] For example, regression algorithms calculate prediction bias and adjustment magnitude for the first feature association set.

[0142] In one embodiment, linear regression analysis is used to analyze the relationship between defect location, score, and temperature, generating predicted deviations such as a score deviation of 0.2 and a temperature deviation of 5°C. Weighted calculations are performed, assigning weights of 0.5, 0.3, and 0.2 to location, score, and temperature, respectively, to determine the adjustment range, such as a temperature reduction of 3°C. If the deviation is less than 10%, a first predicted adjustment set is generated. This set records the deviation and adjustment suggestions for each defect, guiding process optimization.

[0143] Preferably, the regression algorithm is trained using historical data to ensure prediction accuracy.

[0144] In one embodiment, constrained optimization techniques integrate the deviations, adjustment magnitudes, and pressure parameters from a first predictive adjustment set to generate a process parameter optimization boundary dataset.

[0145] Specifically, pressure parameters such as 0.8 MPa are analyzed together with temperature and position deviations. Linear programming is used to set boundary ranges, such as temperature 210-240°C and pressure 0.7-0.9 MPa. If the optimization results meet process constraints, such as equipment limits, a final dataset is generated. This dataset is stored in CSV format and includes columns such as temperature range, pressure range, and position labels, supporting real-time control.

[0146] Understandably, constraint optimization balances multiple parameter conflicts, ensures the feasibility of adjustments, and helps stabilize film quality.

[0147] For example, when boundary datasets guide process adjustments, assuming defects are near the membrane edge, have higher scores, and are at higher temperatures, optimization suggestions might be made to reduce the temperature to 220°C and the pressure to 0.75 MPa. After adjustment, the defect score is expected to decrease by 0.3, and the defect locations will be more evenly distributed. This process improves process consistency through multi-parameter collaborative optimization.

[0148] It should be noted that the structured output of the boundary dataset facilitates integration into the quality monitoring system, providing real-time feedback on process status and supporting efficient decision-making.

[0149] For example, in a thin film fabrication scenario, defect distribution data and severity scores need to be structured using data parsing techniques. Data parsing techniques can extract key information from the raw logs, such as defect coordinates, scores, and temperatures. Assuming the raw data is in JSON format, containing defect coordinates such as x=100, y=150, a severity score of 4.0, and a temperature of 250°C, the parsing process extracts and verifies the completeness of each field using a script. If the proportion of missing values ​​is less than 5%, a second structured dataset is generated. This dataset is stored in tabular form, with columns including defect coordinates, scores, and temperatures, and each row corresponding to a defect instance, facilitating subsequent analysis.

[0150] It should be noted that the integrity check removes outlier data by statistically analyzing the proportion of missing values, thus ensuring the reliability of the dataset.

[0151] Specifically, feature extraction techniques separate the interactive features of defect distribution, severity, and temperature from the second structured dataset. Defect distribution is used to generate region labels, such as "central area" or "edge area," through spatial analysis; severity scores are generated into levels, such as high, medium, and low, through thresholding; and temperature is used to generate continuous features through standardization.

[0152] In one embodiment, correlation analysis is used to assess the correlation between the three interactive features. If the correlation is higher than 85%, a second feature association set is generated. This set is stored in matrix form, with each row representing a defect instance and each column containing a region label, rating level, and temperature value. This structured feature set facilitates classification and optimization analysis.

[0153] Understandably, the high correlation indicates a strong interaction between temperature and defect distribution and severity, providing a data foundation for subsequent optimization.

[0154] In one possible implementation, the gradient descent algorithm calculates the optimization direction and step size for the interaction features, stretching rate, and cooling rate in the second feature association set. Assuming the stretching rate is 2 mm / s and the cooling rate is 10°C / min, the algorithm calculates the optimization direction through weighted iteration.

[0155] For example, the weights are assigned as follows: interaction feature 0.4, stretching rate 0.3, cooling rate 0.3. Iterative calculations determine that the optimization direction is to reduce the cooling rate, with a step size of 0.5°C / min. The generated first set of optimization parameters records the optimization direction and step size, such as the cooling rate needing to be reduced to 9.5°C / min.

[0156] Preferably, the algorithm is pre-trained using historical data to ensure the accuracy of the optimization direction. This method iteratively approaches the optimal solution, improving the precision of parameter adjustment.

[0157] For example, constrained optimization integrates the optimization direction, step size, and constraints from the first set of optimization parameters to generate a second set of optimization parameters. Assume the process constraints include a temperature range of 220-260°C, a stretching rate of 1.5-2.5 mm / s, and a cooling rate of 8-12°C / min. Constrained optimization uses linear programming to analyze the optimization direction and step size, generating parameter combinations that satisfy the constraints, such as a temperature of 245°C, a stretching rate of 2.1 mm / s, and a cooling rate of 9.5°C / min. This combination is stored in CSV format, containing the temperature, stretching rate, and cooling rate ranges, facilitating process control.

[0158] It should be noted that constraint optimization balances multiple parameter conflicts, ensuring that adjustments are within equipment limits. This structured output facilitates integration into the process control system, supporting real-time decision-making.

[0159] In one embodiment, assuming defects are concentrated at the film edges, resulting in a high severity score and a high cooling rate, the optimization suggestion is to reduce the cooling rate to 9.5°C / min and adjust the temperature to 245°C. After adjustment, a more uniform defect distribution is expected, with the score decreasing by 0.4. This multi-parameter synergistic optimization improves film quality consistency through precise control.

[0160] Understandably, the generation of optimal parameter combinations relies on a comprehensive analysis of feature correlation and constraints, which is logically rigorous and has practical guiding significance.

[0161] In the second embodiment of the present invention, as Figure 8 As shown, S403 specifically includes the following steps (S40301, S40302, and S40303):

[0162] S40301: By using the parameter performance relationship model and inputting the process parameter data, the membrane performance indicators, including mechanical properties, transparency, and chemical stability, are predicted to obtain the performance prediction results.

[0163] S40302: If the performance prediction result is lower than the preset membrane quality threshold, the deviation between the process parameters and historical data is analyzed through the parameter performance relationship model to generate a quantitative value of performance deviation.

[0164] S40303: Based on the quantified performance deviation value, the optimized value is recalculated using the gradient descent method of a machine learning algorithm to obtain the optimized parameter combination.

[0165] In the second embodiment of the present invention, the optimized value was recalculated based on the performance deviation quantification value, which further improved the reliability of the optimized parameter combination compared with the optimized value in the first embodiment.

[0166] In step S40301, specifically, temperature data, stretching rate, and cooling rate are acquired from real-time collected process parameters. Data cleaning techniques are used to denoise the data, generating a first parameter dataset. The first parameter dataset is then determined to meet integrity requirements, resulting in a standardized parameter set. Based on the standardized parameter set, forward propagation techniques are used to extract features from the parameter inputs, generating a first feature set. If the dimension of the first feature set exceeds a preset threshold, principal component analysis is used to reduce the dimensionality of the first feature set, resulting in a second feature set. Using the second feature set, a logistic regression algorithm is used to train the mapping relationship between temperature data, stretching rate, and cooling rate and mechanical properties, transparency, and chemical stability, generating a first prediction dataset. The first prediction dataset is then determined to meet consistency requirements, resulting in a performance prediction set. Predicted values ​​for mechanical properties, transparency, and chemical stability are obtained from the performance prediction set. Data visualization techniques are used to format the predicted values, generating a prediction result dataset. The prediction result dataset is then determined to meet business objectives, resulting in the final prediction result.

[0167] For example, in thin film preparation scenarios, real-time acquisition of temperature, stretching rate, and cooling rate data requires ensuring sensor accuracy.

[0168] For example, a temperature sensor records data once per second, with the acquisition time being 14:00:00 on April 21, 2025, yielding a temperature of 600℃, a tensile rate of 3mm / s, and a cooling rate of 20℃ / min. Data cleaning techniques can employ median filtering to remove noise and eliminate outliers caused by equipment vibration. For instance, a record with a temperature abruptly changing to 650℃ is smoothed to 600℃, generating the first parameter dataset. This dataset contains 1000 records, ensuring no missing values ​​and meeting integrity requirements, thus forming a standardized parameter set.

[0169] Specifically, the standardized parameter set is normalized to map the temperature range of 0-1000℃ to 0-1, with 600℃ normalized to 0.6. Based on this dataset, forward propagation techniques are used to extract features, such as calculating the interaction effect between temperature and stretching rate using a multilayer perceptron, generating a first feature set containing 10 dimensions. If the feature dimension exceeds a preset threshold of 5, principal component analysis is used to retain 90% of the variance, reducing the dimension to 4, resulting in a second feature set. This process reduces computational complexity and ensures feature validity.

[0170] In one possible implementation, the second feature set is input into a logistic regression algorithm to train the mapping relationship between temperature, stretching rate, cooling rate and mechanical properties, transparency and chemical stability.

[0171] For example, a dataset contains normalized inputs of 0.6, 0.3, and 0.4, and outputs a transparency score of 85 and a chemical stability score of 0.9. After training, a first prediction dataset is generated. A consistency check reveals that the predicted transparency value for a certain record deviates by 3%, meeting the requirements, thus forming a performance prediction set. This set reliably reflects the relationship between parameters and performance.

[0172] Preferably, the performance prediction set extracts predicted values, such as temperature 650℃, tensile rate 4mm / s, cooling rate 25℃ / min, predicted transparency 90, and chemical stability 0.95. Data visualization technology formats the predicted values ​​into a line graph, showing the trend of transparency changing with temperature, intuitively presenting the impact of process parameters. After generating the prediction result dataset, it is verified whether the transparency is higher than the business target 88. If the requirement is confirmed to be met, the final prediction result is obtained. This result provides a basis for process optimization, such as adjusting the temperature to 620℃ to improve transparency.

[0173] Understandably, the above process forms a closed loop through denoising, feature extraction, dimensionality reduction, model training, and visualization. Each step ensures data reliability, and the prediction results provide accurate guidance for thin film preparation, improving process stability.

[0174] In step S40302, specifically, temperature, stretching rate, and cooling rate are obtained from real-time acquired process parameters. Data cleaning techniques are used to denoise the data, generating a first parameter dataset. The first parameter dataset is then determined to meet integrity requirements, resulting in a standardized parameter set. Based on the standardized parameter set and historical data, an Euclidean distance algorithm is used to calculate the deviation between the real-time process parameters and historical data, generating a first deviation dataset. The first deviation dataset is then determined to meet consistency requirements, resulting in a deviation feature set. If the deviation feature set shows a performance prediction result lower than the preset membrane quality threshold, a logistic regression algorithm is used to analyze the mapping relationship between the deviation feature set and mechanical properties, transparency, and chemical stability, generating a second deviation dataset. The second deviation dataset is then determined to meet quantization requirements, resulting in a deviation quantization set. Based on the deviation quantization set, data visualization techniques are used to format the deviation quantization values, generating a quantization result dataset. The quantization result dataset is then determined to meet business objectives, obtaining the final deviation quantization value.

[0175] For example, in thin film fabrication scenarios, real-time acquisition of temperature, stretching rate, and cooling rate is crucial for process monitoring. The temperature sensor records data once per second; for instance, on April 21, 2025, at 15:00:00, it recorded a temperature of 610℃, a stretching rate of 3.5 mm / s, and a cooling rate of 22℃ / min. Data cleaning employs a sliding window averaging method to remove noise and eliminate outliers caused by equipment vibration. For example, a record showing a sudden temperature change to 660℃ is smoothed to 610℃, generating the first parameter dataset. This dataset contains 800 records, which are checked to ensure no missing values ​​and meet integrity requirements, forming a standardized parameter set. Standardization is achieved through linear normalization, mapping temperatures from 0-1000℃ to 0-1. 610℃ is normalized to 0.61, and the stretching rate and cooling rate are processed similarly to ensure data consistency.

[0176] In one possible implementation, a standardized parameter set is compared with historical data, and the deviation is calculated using an Euclidean distance algorithm. The historical data contains 1000 records, such as process parameters with a transparency of 88 (e.g., temperature 600℃, stretching rate 3mm / s, cooling rate 20℃ / min). The Euclidean distance between the real-time data and the historical data generates a first deviation dataset, for example, with a deviation value of 0.05. After checking for consistency, outlier records with deviation values ​​greater than 0.1 are removed, resulting in a deviation feature set. This set reflects the differences between the real-time and historical processes, providing a basis for subsequent analysis.

[0177] It should be noted that if the deviation feature set shows that the predicted performance is lower than the membrane quality threshold, such as transparency below 85, then a logistic regression algorithm is used to analyze the mapping relationship between deviation and performance. Logistic regression is trained on a historical dataset. The input deviation feature set (e.g., a deviation value of 0.05 corresponds to transparency of 86 and chemical stability of 0.92) generates a second deviation dataset. After checking the quantification requirements and confirming that the predicted deviation is less than 5%, a quantified deviation set is formed. This set quantifies the specific impact of deviation on performance, facilitating process adjustments.

[0178] Specifically, the deviation quantification set is formatted into a bar chart using data visualization technology to show the relationship between deviation values ​​and transparency. For example, a deviation of 0.05 corresponds to a transparency of 86, and a deviation of 0.08 corresponds to a transparency of 84, intuitively presenting the effect of deviation on process parameters. After generating the quantification result dataset, it is verified whether it meets the business objectives. If the transparency is higher than 85, it is considered qualified, and the final deviation quantification value is obtained. For example, a deviation of 0.04 corresponds to a transparency of 87. This value guides process optimization, such as reducing the temperature to 605℃ to reduce deviation.

[0179] In one embodiment, the Euclidean distance algorithm quantifies the gap between real-time and optimal processes by comparing multi-dimensional parameter vectors, making it suitable for multi-parameter scenarios. Logistic regression, on the other hand, accurately maps the relationship between deviation and performance through a probabilistic model, making it suitable for non-linear scenarios. Visualization technology transforms complex data into intuitive graphics, facilitating rapid decision-making by process engineers.

[0180] Understandably, the above process provides precise guidance for thin film preparation and optimizes process stability through deviation analysis and visualization.

[0181] In step S40303, specifically, temperature, stretching rate, and cooling rate are obtained from real-time acquired process parameters. Mean filtering is used to denoise the data, generating a first process dataset. The first process dataset is then determined to meet integrity requirements, resulting in a standardized process set. Based on the standardized process set, principal component analysis is used to extract key features of temperature, stretching rate, and cooling rate, generating a first feature dataset. If the first feature dataset meets consistency requirements compared to a preset deviation quantization threshold, a feature quantization set is obtained. Gradient descent is used to iteratively calculate the feature quantization set, obtaining optimized values ​​for temperature, stretching rate, and cooling rate, generating a first optimized dataset. The first optimized dataset is determined to meet performance prediction requirements, resulting in an optimized parameter set. The optimized parameter set is formatted using data visualization technology to generate a quantized optimized dataset. The quantized optimized dataset is then determined to meet the membrane quality business objectives, obtaining the final optimized parameter combination.

[0182] For example, in thin film fabrication scenarios, real-time acquisition of temperature, stretching rate, and cooling rate is fundamental to process optimization. The temperature sensor records data once per second; for instance, on April 21, 2025, at 14:00:00, the data collected showed a temperature of 620°C, a stretching rate of 3.8 mm / s, and a cooling rate of 25°C / min. Mean filtering is used for noise reduction; its principle is to smooth out abnormal fluctuations by calculating the average value of the data within a fixed window.

[0183] For example, if the window size is set to 5 seconds, and the temperature suddenly changes to 670℃, it will be smoothed to 622℃ after filtering, generating the first process dataset. This dataset contains 1000 records, and its integrity is checked to ensure there are no missing values ​​or invalid records, thus ensuring data reliability.

[0184] In one possible implementation, the standardized process set is generated through min-max normalization, mapping temperatures from 0-1000℃ to 0-1, and normalizing 620℃ to 0.62. Tensile rate and cooling rate are treated similarly. Principal component analysis is used to extract key features; the principle is to reduce the dimensionality of high-dimensional data to principal components, preserving the core trends in temperature, tensile rate, and cooling rate.

[0185] For example, after analyzing 1000 records, temperature and cooling rate are identified as the principal components affecting membrane quality, generating the first feature dataset. This dataset is compared with the deviation quantization threshold; if the eigenvector variance is less than 0.08, it meets the consistency requirement and forms the feature quantization set.

[0186] It should be noted that the gradient descent algorithm finds the optimal values ​​of temperature, stretching rate, and cooling rate by iteratively optimizing the feature quantization set. The principle is based on minimizing the loss function and adjusting the parameters to approximate the membrane quality target.

[0187] For example, the initial temperature was adjusted from 620℃ to 615℃, and the stretching rate was optimized from 3.8 mm / s to 3.6 mm / s, generating the first optimized dataset. Inspection revealed that the optimized parameters achieved a prediction transparency of 87, meeting the performance requirements, thus forming an optimized parameter set.

[0188] Specifically, data visualization technology formats the optimized parameter set into a line graph, showing the relationship between temperature and transparency. For example, 615℃ corresponds to a transparency of 87, and 620℃ corresponds to 86, intuitively presenting the optimization effect. The quantified optimization dataset further validates the membrane quality target, such as a transparency higher than 85, generating the final optimized parameter combination, such as a temperature of 615℃, a stretching rate of 3.6 mm / s, and a cooling rate of 24℃ / min. This combination guides process adjustments to ensure stable quality.

[0189] In one embodiment, mean filtering is suitable for high-noise scenarios, principal component analysis reduces redundant information, gradient descent accurately optimizes parameters, and visualization facilitates rapid decision-making by process engineers.

[0190] Understandably, the above process provides efficient guidance for thin film preparation and improves process consistency through feature extraction and optimization.

[0191] Step S5: Using the optimized parameter combination, update the control system of the PET film preparation equipment and adjust the process parameter data.

[0192] Specifically, real-time data on temperature, stretching rate, and cooling rate are acquired from sensors on the production line. A first data set is generated using data acquisition technology, and this first data set is standardized to obtain a first standardized dataset. If the temperature or stretching rate in the first standardized dataset exceeds a preset threshold, logical judgment technology is used to separate abnormal data and generate a second abnormal dataset. Defect detection is performed on the second abnormal dataset to determine the defect type. Based on the defect type, a decision tree algorithm is used to calculate adjustment values ​​for temperature, stretching rate, and cooling rate to generate a first parameter adjustment set. This first parameter adjustment set is weighted to obtain a second optimized parameter set. Control commands are generated using the second optimized parameter set, and real-time control technology is used to update the production line control system, adjusting the temperature, stretching rate, and cooling rate to produce a PET film that meets performance requirements.

[0193] For example, in thin film fabrication scenarios, acquiring real-time data from production line sensors is fundamental to process optimization.

[0194] For example, sensors are deployed at key nodes in a PET film production line to collect parameters such as temperature, stretching rate, and cooling rate. Data acquisition technology uses a PLC system to acquire raw data at 1-second intervals, forming a first data set.

[0195] For example, a certain data collection recorded a temperature of 260°C, a tensile rate of 2.2 mm / s, and a cooling rate of 10°C / min.

[0196] It should be noted that the sensor data needs to be filtered to eliminate noise and ensure data accuracy. Standardization processing is applied to the first dataset.

[0197] Specifically, the z-score standardization method is used to convert temperature, stretching rate, and cooling rate into dimensionless values ​​with a mean of 0 and a standard deviation of 1.

[0198] In one embodiment, assuming a temperature range of 230-270°C, 260°C can be standardized to 0.8, and a tensile rate of 2.2 mm / s can be standardized to 0.6. This standardization facilitates subsequent anomaly detection and algorithm analysis, enhancing data consistency.

[0199] Preferably, for the first standardized dataset, logical judgment techniques are used to separate outlier data.

[0200] For example, the preset temperature threshold is ±1.5 standard deviations, and the stretching rate threshold is ±1.2 standard deviations. If the temperature value of 0.8 exceeds the threshold, but the stretching rate of 0.6 does not, the abnormal temperature record is separated into a second abnormal dataset.

[0201] Understandably, this separation ensures that subsequent analysis focuses on data that may lead to defects, thus improving detection efficiency.

[0202] In one possible implementation, defect detection for the second abnormal dataset determines the defect type through pattern recognition.

[0203] For example, abnormal temperatures may cause "surface wrinkles" or "uneven thickness." By comparing historical data, if the temperature exceeds 1.5 standard deviations and is concentrated at the front end of the production line, it is identified as "surface wrinkles." This method quickly locates the cause of defects through data features. For parameter adjustments based on the defect type, the decision tree algorithm calculates the adjustment value based on historical process data.

[0204] For example, regarding "surface wrinkles," decision tree analysis shows that reducing the temperature to 255°C and adjusting the stretching rate to 2.0 mm / s can reduce defects. The results form the first parameter adjustment set, and the adjustment direction is recorded.

[0205] It should be noted that decision trees ensure the targeted nature of adjustment values ​​through hierarchical judgment.

[0206] In one embodiment, the weighted processing of the first parameter adjustment set is optimized through weight allocation.

[0207] For example, temperature is weighted at 0.6, stretching rate at 0.3, and cooling rate at 0.1, generating a second set of optimized parameters, such as temperature 255°C, stretching rate 2.0 mm / s, and cooling rate 9.8°C / min. This weighted method balances the influence of multiple parameters and improves adjustment accuracy.

[0208] Specifically, the second set of optimized parameters generates control instructions, which are then updated on the production line through real-time control technology.

[0209] For example, after receiving instructions, the PLC system adjusts the heating module to 255°C, the stretching motor to 2.0 mm / s, and the cooling fan to 9.8°C / min. This real-time control ensures that the PET film has uniform thickness and a smooth surface, meeting performance requirements.

[0210] Understandably, rapid response in instruction generation and execution effectively reduces defects and improves production efficiency.

[0211] This invention discloses a data analysis-based control method for PET film fabrication. The method acquires real-time process parameter data to obtain a structured dataset, and uses a pre-established machine learning algorithm to train a parameter performance relationship model. It further acquires real-time images of the PET film surface, generating defect distribution data and quantified defect severity results. Then, it analyzes the interaction between the defect distribution data, the quantified defect severity results, and the process parameter data to obtain an optimized parameter combination. Finally, it uses this optimized parameter combination to update the PET film fabrication equipment control system and adjust the process parameter data. This invention achieves a high-performance PET film fabrication scheme by real-time acquisition of process parameters, optimization of parameter configuration using machine learning algorithms, and the combination of real-time defect detection and the interaction of process parameter data.

[0212] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0213] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They do not limit the scope of protection of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should fall within the scope of the claims of the present invention.

Claims

1. A data analysis-based method for controlling the preparation of PET films, characterized in that, Includes the following steps: S1: Acquire real-time process parameter data from sensors in the PET film preparation equipment and store it in the database to obtain a structured dataset; S2: Using a pre-established machine learning algorithm, a parameter performance relationship model is trained based on the structured dataset. The parameter performance relationship model is a model showing the relationship between the process parameter data and the performance of the PET film. S3: Obtain a real-time image of the PET film surface, process the image using a convolutional neural network algorithm, identify the type and location of surface defects, and generate defect distribution data and defect severity quantification results; S4: By analyzing the interaction between the defect distribution data, the defect severity quantification results, and the process parameter data through the parameter performance relationship model, an optimized parameter combination is obtained; S5: Using the optimized parameter combination, update the control system of the PET film preparation equipment and adjust the process parameter data; Step S4 includes the following steps: S401: Through the parameter performance relationship model, analyze the interaction between the defect distribution data, the defect severity quantification results, and the process parameter data to generate performance prediction adjustment values ​​and process parameter optimization boundary conditions; S402: Based on the performance prediction adjustment value and the process parameter optimization boundary conditions, the optimized value of the process parameter data is calculated using the gradient descent method of the machine learning algorithm; S403: Obtain the optimized parameter combination through the optimized value; Specifically, step S403 includes the following steps: S40301: By using the parameter performance relationship model and inputting the process parameter data, the membrane performance indicators, including mechanical properties, transparency, and chemical stability, are predicted to obtain the performance prediction results. S40302: If the performance prediction result is lower than the preset membrane quality threshold, the deviation between the process parameters and historical data is analyzed through the parameter performance relationship model to generate a quantitative value of performance deviation. S40303: Based on the quantified performance deviation value, the optimized value is recalculated using the gradient descent method of a machine learning algorithm to obtain the optimized parameter combination; Specifically, step S40302 includes: acquiring temperature, stretching rate, and cooling rate from real-time collected process parameters; performing noise reduction on the data using data cleaning technology to generate a first parameter dataset; determining that the first parameter dataset meets the integrity requirements to obtain a standardized parameter set; calculating the deviation between the real-time process parameters and historical data using an Euclidean distance algorithm based on the standardized parameter set and historical data to generate a first deviation dataset; determining that the first deviation dataset meets the consistency requirements to obtain a deviation feature set; if the deviation feature set shows a performance prediction result lower than the preset membrane quality threshold, then using a logistic regression algorithm to analyze the mapping relationship between the deviation feature set and mechanical properties, transparency, and chemical stability to generate a second deviation dataset; determining that the second deviation dataset meets the quantization requirements to obtain a deviation quantization set; and formatting the deviation quantization values ​​using data visualization technology based on the deviation quantization set to generate a quantization result dataset; determining that the quantization result dataset meets the business objectives to obtain the final deviation quantization value.

2. The data analysis-based PET film preparation control method according to claim 1, characterized in that, The process parameter data includes: temperature data, stretching rate data, and cooling rate data.

3. The data analysis-based PET film preparation control method according to claim 2, characterized in that, Step S1 includes the following steps: S101: Obtain real-time process parameter data from sensors in the PET film preparation equipment; S102: Determine whether the process parameter data meets the preset threshold range. If it does, mark it as valid data and obtain a valid dataset. S103: Based on the valid dataset, the temperature data, the stretching rate and the cooling rate are classified and stored according to timestamps using database storage technology, and a unique identifier is generated for each parameter; S104: Confirm that the stored procedure is complete and obtain the structured dataset.

4. The data analysis-based PET film preparation control method according to claim 1, characterized in that, The training of the parameter performance relationship model based on the structured dataset includes: The process parameter data is obtained from the structured dataset. The process parameter data is normalized using data preprocessing techniques to generate a first dataset. The first dataset is determined to meet the training requirements to obtain a standardized dataset. Based on the standardized dataset, a parameter performance relationship model is trained using a logistic regression algorithm.

5. The data analysis-based PET film preparation control method according to claim 4, characterized in that, Step S3 includes the following steps: S301: Acquire a real-time image of the PET film surface, process the image using a convolutional neural network algorithm, identify the type and location of surface defects, and generate defect distribution data; S302: Calculate the severity score of the defect using the defect distribution data to obtain a quantitative result of the defect severity.

6. The data analysis-based PET film preparation control method according to claim 5, characterized in that, S301 includes the following steps: S3011: Acquire a real-time image of the PET film surface, and use mean filtering technology to denoise the image to generate a first image set; S3012: If the resolution of the first image set meets the preset threshold, then a standardized image set is obtained; S3013: Based on the standardized image set, feature extraction is performed using a convolutional neural network algorithm to identify the defect types on the membrane surface and generate a first defect type set; S3014: If the first defect type set satisfies the preset defect classification threshold, then the defect feature set is obtained; S3015: The defect feature set is processed using image segmentation technology to determine the specific location of the defects on the membrane surface and generate a first defect location set; S3016: If the positioning accuracy of the first defect location set meets the preset threshold, then the defect location set is obtained; S3017: Integrate the defect feature set and the defect location set through data formatting technology to generate defect distribution data containing defect types and locations.

7. The data analysis-based PET film preparation control method according to claim 6, characterized in that, Step S302 includes the following steps: S3021: Obtain the defect type and location information from the defect distribution data, perform structured processing on the defect distribution data using data parsing technology, and generate a first structured dataset from the processed defect distribution data whose integrity meets a preset threshold; S3022: Based on the first structured dataset, feature extraction technology is used to separate the defect type and location information. If the separation degree of the separated feature set meets the preset classification threshold, a first feature set is generated. S3023: Using a pre-established scoring standard table, a weighted calculation method is used to calculate the severity score based on the defect type and location information in the first feature set. If the score consistency meets a preset threshold, a first score set is generated. S3024: The first scoring set is integrated using data formatting technology to generate a quantitative result dataset containing the defect type, location information, and severity score; S3025: If the format of the quantification result dataset meets the preset business objectives, then the final defect severity quantification result is generated.

8. The data analysis-based PET film preparation control method according to claim 1, characterized in that, Step S403 specifically includes the following steps: S4031: Using a pre-established defect solution database, based on the defect distribution data, the quantification results of the severity of prime number defects, and the performance prediction adjustment value, match them with the defect type index and historical data in the database to determine the defect processing parameters; S4032: The optimized parameter combination is obtained by combining the optimized value with the defect handling parameters.

Images