A polymer composite process AI model capacity prediction method

Abstract

The application belongs to the field of laboratory preparation and industrial amplification, and particularly relates to a capacity prediction method for a high polymer composite process AI model, which comprises the following steps: generating a unique small test batch number, guaranteeing data integrity through data standardization and Merkle tree verification; constructing a small test and mass production paired sample set, extracting process characteristics such as thermal history and shear mixing strength, and obtaining an input feature vector by fusing equipment scale descriptors; constructing a multi-task cross-scale prediction model based on a Transformer and an MLP, and outputting mass production performance and capacity indexes; constructing an amplification risk index and grading early warning, and generating optimal process parameters of a large machine through a multi-objective genetic algorithm. The application realizes capacity prediction from small test to mass production, reduces cross-scale amplification risk, optimizes mass production process, and improves production efficiency and product quality.

Description

Technical Field

[0001] This invention pertains to both laboratory preparation and industrial scale-up, specifically a method for predicting the production capacity of a polymer composite material process AI model. Background Technology

[0002] With its advantages of being lightweight, high-strength, and environmentally resistant, polymer composite materials have become core materials in high-end manufacturing fields such as automobiles and aerospace. The key bottleneck in their industrialization lies in the problem of distortion across scale from laboratory pilot-scale testing to industrial mass production.

[0003] The industry currently faces a common dilemma: successful pilot testing but unsuccessful mass production. While precise control during pilot testing can achieve stable product performance and process parameters, scaling up to large-scale mass production often results in issues such as insufficient capacity, excessive product performance fluctuations, or soaring energy consumption – a phenomenon known as scale-up distortion. The core causes of this pain point include: traditional methods only extract simple process parameters, failing to explore core dynamic characteristics such as thermal history and shear mixing intensity, and ignoring the fundamental differences between pilot testing and large-scale production in terms of geometric scale, heating capacity, and material conveying efficiency, thus failing to characterize cross-scale transfer patterns; reliance on empirical judgment or simple linear scaling models fails to predict the coupling relationship between mass production performance and capacity, making it impossible to mitigate distortion risks in advance; and the lack of quantitative risk assessment and process optimization methods leads to blind parameter adjustments during pilot production, resulting in high production costs, long cycles, and low industrialization success rates.

[0004] Therefore, a capacity prediction method based on an AI model for polymer composite material processes is needed to solve the above problems. Summary of the Invention

[0005] To address the technical problems mentioned in the background, this invention provides a method for predicting the production capacity of a polymer composite material process AI model.

[0006] The objective of this invention can be achieved through the following technical solutions: This invention provides a method for predicting production capacity using an AI model for polymer composite material processes, comprising the following steps: Please refer to Figure 1 As shown, this invention provides a method for predicting the production capacity of a polymer composite material process AI model, comprising the following steps: Step 1: Data Acquisition and Batch Modeling: Generate a unique batch number for each pilot experiment, collect various types of heterogeneous data, process the heterogeneous data through a data standardization algorithm, and complete batch integrity verification. Step 2, Data Association and Feature Construction: Paired sample sets are constructed from heterogeneous data, and thermal history features, shear and mixing intensity features, energy input features, and stability features of pilot process parameters are extracted. The input feature vector is obtained by fusing the equipment scale descriptor.

[0007] Step 3: Model Training: Based on the paired sample set, a multi-task cross-scale prediction model is constructed using Transformer and MLP. The input feature vector is input into the multi-task cross-scale prediction model, and the output mass production performance prediction value and capacity index prediction value are output. Step 4: Solution Generation: Construct an amplified risk index based on the predicted mass production performance and capacity indicators, and generate the optimal process parameters for the large machine through a multi-objective genetic algorithm; In step five, re-optimization: when the cross-scale prediction model determines that the current formulation does not meet the standards, the system automatically triggers a small-scale formulation recommendation based on Bayesian optimization, generates optimized component ratios, and reassesses the risks of mass production.

[0008] As a preferred technical solution, based on step one, a unique batch number is generated for each pilot experiment, and various types of heterogeneous data are collected. The heterogeneous data is then processed using a data standardization algorithm to complete batch integrity verification. The specific steps are as follows: A unique batch number is generated for each pilot experiment, and a record of raw material parameters and formulation inputs corresponding to that batch number is established, including resin system information, filler and additive types and their mass fractions, raw material suppliers and incoming batch numbers, formulation version numbers, target performance indicators and testing standards. Furthermore, pilot process parameters, pilot sample performance parameters, and mass production verification data are collected in real time. Pilot process parameters include screw extruder technical parameters, screw configuration parameters, and control setpoints and sensor feedback values, including screw speed, heating zone temperature, melt temperature, heating zone barrel volume, material density, feeding rate, and vacuum degree. Pilot sample performance parameters include tensile strength, flexural strength, and impact strength. Mass production verification data includes the mass production verification batch number matched with the pilot batch number, mass production process parameters, mass production sample performance, and mass production capacity indicators. The pilot-scale experimental data was divided into numerical, character, and time-series types. The numerical data was mapped to the 0-1 range using the Z-score normalization algorithm to obtain standard numerical data. Character data was mapped to integer codes by constructing an industry-specific classification dictionary. One-hot encoding was performed on the integer codes to generate sparse character vectors. Using the pilot-scale experimental start time as zero, the timestamps of all time-series data were converted to relative times. A linear interpolation algorithm was used to resample time-series data from different sampling frequencies to a unified frequency, resulting in standard time-series data. Each field value of the normalized time-series data was traversed, and the field name and value were concatenated into a string. An MD5 hash value (HM) was generated and set as the leaf node of the current Merkle tree. The field-level hash values ​​of all fields within a single record were concatenated in ascending order of field name to generate a record-level hash. The record-level hash values ​​of all records within the collected data were concatenated in ascending order of primary key to generate a table-level hash. Finally, the data was concatenated in ascending order of table-level hashes to generate a batch-level Merkle root hash. The calculation logic is as follows: ,in , , and These are table-level hashes for formula feeding records, process parameters, sample performance, and mass production verification parameters, respectively. The algorithm is a hash algorithm. The batch-level Merkle root hash is compared with the preset batch hash baseline value. If the two are completely consistent, it is determined that the data has not been tampered with. If they are inconsistent, the tampered node is located by the hash backtracking algorithm of the Merkle tree and corrected to obtain the corrected data.

[0009] As a preferred technical solution, based on step two, paired samples are constructed from various types of data, and the thermal history characteristics, intensity characteristics, energy input characteristics, and stability characteristics of the pilot-scale process parameters are extracted. These are then fused with the equipment scale descriptor to obtain the input feature vector. The specific steps are as follows: By using the unique connection key of the batch number in the pilot test, and by using a multi-table internal join algorithm to connect various parameters of raw materials and formula feeding records, pilot test process parameters and pilot test sample performance parameters, structured data is obtained. The structured data is then converted into a pilot test full-link input vector. The mass production verification results are set as supervision labels and paired with the corresponding pilot test full-link input vectors to generate a paired sample set. Arrange the pilot-scale process parameters in chronological order to obtain time series curves. Perform several layers of wavelet decomposition on a single time series curve to obtain approximation coefficients and detail coefficients. Calculate an adaptive threshold based on the detail coefficients. Reconstruct the detail coefficients that are greater than the adaptive threshold to obtain denoised time series curves. Extract the feeding rate FW, material density PL, barrel volume VH of each heating zone, and screw channel filling coefficient JS from the denoised timing curve, and calculate the material residence time of each heating zone. The calculation logic is as follows: , where k is the number of the heating zone cylinder; Multiply the average temperature of each temperature zone by the corresponding residence time, and sum them to obtain the total heat exposure intensity. This index directly reflects the cumulative heat absorbed by the material during processing. Then, calculate the standard deviation and mean of the product of temperature and residence time for each temperature zone. Subtract the ratio of the standard deviation to the mean from 1 to obtain the heat exposure uniformity. The closer this ratio is to 1, the better. Query the thermal decomposition temperature of the current resin system and compare the highest temperature of each temperature zone with the thermal decomposition temperature. If the highest temperature of any temperature zone exceeds the thermal decomposition temperature, the overheating risk index is set to 1; otherwise, it is set to 0. This is used to determine whether the material has a risk of thermal degradation. The total heat exposure intensity, heat exposure uniformity, and thermal degradation risk are integrated into a thermal history feature. Extract the screw diameter HJ, screw speed NH, and screw groove depth MC from the denoised time series curve, and calculate the instantaneous shear rate of the material in the screw groove. The calculation logic is as follows: ,in Pi is used to convert the circular motion of the screw rotation speed into linear velocity calculation, obtain the current head pressure, and perform time integration based on the instantaneous shear rate and head pressure to obtain the mixing intensity FX. Its calculation logic is as follows: Where yc is the instantaneous shear rate and P is the head pressure. The symbol for definite integral indicates the cumulative calculation of instantaneous products over the interval [t1, t2]. This indicator comprehensively reflects the material mixing effect under the combined influence of shearing and flow resistance. It integrates instantaneous shear rate and mixing intensity to obtain intensity characteristics. Instantaneous power time-series data from the denoised time-series curve is extracted, and the cumulative energy input is obtained by time integration of the instantaneous power time-series data. The total feed mass in the stable section is obtained by time integration of the feed rate time-series data within the time period. The energy input per unit mass of material is obtained by dividing the cumulative energy input by the total feed mass. The cumulative energy input and the energy input per unit mass of material are integrated into the energy input characteristic. The torque curve is decomposed into multiple scales by wavelet transform to obtain wavelet coefficients at different scales. The probability distribution of the wavelet coefficients is calculated. The process trajectory of a high-quality batch with the same formula and target performance in historical small-scale tests is selected as the reference trajectory. The alignment distance between the current batch trajectory and the reference trajectory is calculated by dynamic time warping algorithm. The alignment distance and texture entropy are integrated into the stability characteristic. The geometric scale ratio is obtained by calculating the ratio of the product of the screw diameter and the barrel length of the pilot-scale equipment to that of the large-scale equipment. The heating zone capacity difference is obtained by calculating the ratio of the difference in the number of heating zones between the large-scale equipment and the pilot-scale equipment to the total number of heating zones in the large-scale equipment. The feeding capacity difference is calculated by calculating the ratio of the difference in the maximum feeding rate between the large-scale equipment and the pilot-scale equipment to the maximum feeding rate of the large-scale equipment. The vacuum capacity difference is obtained by calculating the ratio of the difference in the maximum vacuum degree between the large-scale equipment and the pilot-scale equipment to the maximum vacuum degree of the large-scale equipment. The two ratios are added together to obtain the vacuuming capacity difference. The geometric scale ratio, heating zone capacity difference, feeding capacity difference, and vacuuming capacity difference are integrated into an equipment scale descriptor. The input feature vector is obtained by fusing the small-scale full-link input vector, thermal history features, intensity features, energy input features, stability features, and equipment scale descriptors.

[0010] As a preferred technical solution, based on step three, which involves constructing a cross-scale prediction model using Transformer and MLP based on a paired sample set, and outputting predicted values ​​for mass production performance and capacity indicators, the specific steps are as follows: The paired sample set is divided into training, validation, and test sets proportionally. The training set samples are input into the model, which outputs mass production performance predictions and capacity index predictions. The loss value of the current batch is calculated based on the composite objective function. The gradient of each layer parameter is calculated using the gradient descent method, and all parameters of the shared feature extraction layer and the task-specific branch are updated. After each round of training, the model performance is evaluated using the validation set. If the validation set loss reaches the current optimal level, the model weights are saved. When the maximum number of iterations is reached or the early stopping strategy is triggered, the optimal cross-scale prediction model is output. The cross-scale prediction model consists of an input layer, a shared feature extraction layer, and task-specific branches. The shared feature extraction layer includes a temporal feature subnetwork, a static feature subnetwork, and a fusion layer. The task-specific branches include a mass production sample performance mapping branch and a mass production capacity index prediction branch. The input feature vector is sent to the input layer. The temporal feature subnetwork uses a single-layer bidirectional LSTM to analyze the dynamic changes of thermal history features, intensity features, energy input features, and stability features. The static feature subnetwork layer is a fully connected layer that analyzes the static features of the pilot-scale full-link and equipment scale descriptors. The fusion layer concatenates the dynamic and static features into a shared feature vector. The mass production sample performance mapping branch outputs the mass production performance prediction value, and the mass production capacity index prediction branch outputs the capacity index prediction value.

[0011] As a preferred technical solution, based on step four, an amplified risk index is constructed, and optimal process parameters for the large-scale machine are generated using a multi-objective genetic algorithm. The specific steps are as follows: Set the target specification range for mass production performance indicators and obtain the predicted mass production performance values. and the standard deviation of the predicted value And calculate the performance deviation probability, the calculation logic is as follows: ,in Let be the cumulative distribution function of the standard normal distribution. and Given the minimum and maximum values ​​of the target specification range, based on the uncertainty and stability characteristics of the capacity indicators output by the model, historical data is fitted through a logistic regression model to obtain the probability of process instability. The difference between the predicted capacity indicator and the standard capacity indicator is calculated to obtain the capacity deviation value. The risk index is obtained by weighted summation of the performance deviation probability, the capacity deviation value, and the process instability probability. The risk index is mapped to a four-level early warning system. If the risk index is less than the first risk threshold, it is judged as low risk. If the first risk threshold is less than the risk index and less than the second risk threshold, it is judged as medium risk, indicating that there is a single dimension of risk that is too high. The corresponding process parameters need to be adjusted. If the second risk threshold is less than the risk index and less than the third risk threshold, it is judged as high risk. After supplementing the pilot test, the prediction is re-evaluated. If the risk index is greater than or equal to the third risk threshold, it is judged as the highest risk. Pilot process optimization needs to be carried out again, and mass production is not started for the time being. The process parameters of each major machine are converted into chromosomes using real-number encoding, with each gene locus corresponding to the value of a process parameter. Based on the historical mass production process parameter database and equipment constraints, an initial population is randomly generated. Then, through selection, crossover, and mutation operations, the top few individuals with the highest fitness are retained to form a new generation population. The fitness calculation logic is as follows: H1 and H2 are the target weights. and The maximum expected qualified output and maximum energy consumption per unit of output from historical data are used to normalize the target value. and To determine the actual qualified output and actual energy consumption under the corresponding process parameter A; stop iterating when the maximum number of iterations is reached, output the optimal process parameters of the mainframe, and send them to the process execution terminal for control.

[0012] Compared with the prior art, the beneficial effects of the present invention are: By denoising process time-series data through wavelet decomposition, process features such as thermal history, shear and mixing intensity, energy input, and stability are deeply extracted to characterize the dynamic process of material processing. The equipment scale descriptor is integrated to quantify the cross-scale differences between pilot and large-scale machines in terms of geometry, heating capacity, and material conveying, thus compensating for the shortcomings of traditional models that ignore equipment differences. A multi-task model is built based on Transformer and MLP to jointly learn mass production performance and capacity indicators, share cross-scale mapping rules, improve the model's generalization ability and prediction consistency, and effectively reduce prediction distortion caused by cross-scale amplification. The recommended process parameters for large machines predicted by this invention can be directly imported into the production end-to-end optimization system of an AI-driven polymer composite material process optimization method as the initial process setting, thereby achieving accurate, rapid start-up and continuous optimization of the mass production process. The re-optimized step five can improve the efficiency of pilot-scale R&D and form a two-level linkage system from pilot-scale to mass production with an AI-driven process optimization method for polymer composite materials. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to illustrate the main idea of ​​the present invention.

[0014] Figure 1 This is a diagram illustrating the method steps of the present invention. Detailed Implementation

[0015] The technical solutions in 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 also within the scope of protection of the present invention.

[0016] Please refer to Figure 1 As shown, this invention provides a method for predicting the production capacity of a polymer composite material process AI model, comprising the following steps: Step 1: Data Acquisition and Batch Modeling: Generate a unique batch number for each pilot experiment, collect various types of heterogeneous data, process the heterogeneous data through a data standardization algorithm, and complete batch integrity verification. Step 2, Data Association and Feature Construction: Paired sample sets are constructed from heterogeneous data, and thermal history features, shear and mixing intensity features, energy input features, and stability features of pilot process parameters are extracted. The input feature vector is obtained by fusing the equipment scale descriptor.

[0017] Step 3: Model Training: Based on the paired sample set, a multi-task cross-scale prediction model is constructed using Transformer and MLP. The input feature vector is input into the multi-task cross-scale prediction model, and the output mass production performance prediction value and capacity index prediction value are output. Step 4: Solution Generation: Construct an amplified risk index based on the predicted mass production performance and capacity indicators, and generate the optimal process parameters for the large machine through a multi-objective genetic algorithm; In step five, re-optimization: when the cross-scale prediction model determines that the current formulation does not meet the standards, the system automatically triggers a small-scale formulation recommendation based on Bayesian optimization, generates optimized component ratios, and reassesses the risks of mass production.

[0018] As a preferred technical solution, based on step one, a unique batch number is generated for each pilot experiment, and various types of heterogeneous data are collected. The heterogeneous data is then processed using a data standardization algorithm to complete batch integrity verification. The specific steps are as follows: A unique batch number is generated for each pilot test, and a record of raw material parameters and formulation inputs corresponding to that batch number is established, including resin system information, filler and additive types and their mass fractions, raw material suppliers and incoming batch numbers, formulation version numbers, target performance indicators and testing standards. Real-time data is collected on pilot process parameters, pilot sample performance parameters, and mass production verification data. Pilot process parameters include screw extruder technical parameters (extruder model, motor power, screw length-to-diameter ratio, center distance, and rated torque, etc.), screw configuration parameters (details of screw elements and their arrangement, the ratio of conveying section to kneading section length, etc.), and control setpoints and sensor feedback values, including screw speed, heating zone temperature, melt temperature, heating zone barrel volume, material density, feeding rate, and vacuum degree. Pilot sample performance parameters include tensile strength, flexural strength, and impact strength. Mass production verification data includes the mass production verification batch number matching the pilot batch number, mass production process parameters, mass production sample performance, and mass production capacity indicators. The pilot-scale experimental data was divided into numerical, character, and time-series types. The numerical data was mapped to the 0-1 range using the Z-score normalization algorithm to obtain standard numerical data. Character data was mapped to integer codes by constructing an industry-specific classification dictionary. One-hot encoding was performed on the integer codes to generate sparse character vectors. Using the pilot-scale experimental start time as zero, the timestamps of all time-series data were converted to relative times. A linear interpolation algorithm was used to resample time-series data from different sampling frequencies to a unified frequency, resulting in standard time-series data. Each field value of the normalized time-series data was traversed, and the field name and value were concatenated into a string. An MD5 hash value (HM) was generated and set as the leaf node of the current Merkle tree. The field-level hash values ​​of all fields within a single record were concatenated in ascending order of field name to generate a record-level hash. The record-level hash values ​​of all records within the collected data were concatenated in ascending order of primary key to generate a table-level hash. Finally, the data was concatenated in ascending order of table-level hashes to generate a batch-level Merkle root hash. The calculation logic is as follows: ,in , , and These are table-level hashes for formula feeding records, process parameters, sample performance, and mass production verification parameters, respectively. The algorithm is a hash algorithm. The batch-level Merkle root hash is compared with the preset batch hash baseline value. If the two are completely consistent, it is determined that the data has not been tampered with. If they are inconsistent, the tampered node is located by the hash backtracking algorithm of the Merkle tree and corrected to obtain the corrected data.

[0019] As a preferred technical solution, based on step two, the heterogeneous data is used to construct paired samples, and the thermal history features, intensity features, energy input features, and stability features of the pilot-scale process parameters are extracted. These are then fused with the equipment scale descriptor to obtain the input feature vector. The specific steps are as follows: By using the unique connection key of the batch number in the pilot test, and by using a multi-table internal join algorithm to connect various parameters of raw materials, formula feeding records, pilot test process parameters, and pilot test sample performance parameters, structured data is obtained. Structured data is converted into small-scale full-link input vectors, and the mass production verification results of large-scale mass production verification data are set as supervision labels and paired with the corresponding small-scale full-link input vectors to generate a paired sample set. It should be noted that the supervision label is essentially a combination of mass production sample performance sub-vectors and mass production capacity index sub-vectors. Arrange the pilot-scale process parameters in chronological order to obtain time series curves. Perform several layers of wavelet decomposition on a single time series curve to obtain approximation coefficients and detail coefficients. Calculate an adaptive threshold based on the detail coefficients. Reconstruct the detail coefficients that are greater than the adaptive threshold to obtain denoised time series curves. Extract the feeding rate FW, material density PL, barrel volume VH of each heating zone, and screw channel filling coefficient JS from the denoised timing curve, and calculate the material residence time of each heating zone. The calculation logic is as follows: , where k is the number of the heating zone cylinder; Multiply the average temperature of each temperature zone by the corresponding residence time, and sum them to obtain the total heat exposure intensity. This index directly reflects the cumulative heat absorbed by the material during processing. Then, calculate the standard deviation and mean of the product of temperature and residence time for each temperature zone. Subtract the ratio of the standard deviation to the mean from 1 to obtain the heat exposure uniformity. The closer this ratio is to 1, the more uniform the heat exposure of each part of the material. Pre-query the thermal decomposition temperature of the current resin system and compare the highest temperature of each temperature zone with the thermal decomposition temperature. If the highest temperature of any temperature zone exceeds the thermal decomposition temperature, the thermal risk index is set to 1; otherwise, it is set to 0. This is used to determine whether the material has a risk of thermal degradation. The total heat exposure intensity, heat exposure uniformity, and thermal degradation risk are integrated into a thermal history feature. It should be noted that the screw groove filling coefficient is the ratio of the actual volume of material filling the screw groove to the total volume of the screw groove in the small-scale processing of polymer composite materials. The value ranges from 0 to 1. It is a core process parameter for quantifying the filling state of the material in the screw groove and directly affects the shearing action and heat transfer efficiency. Extract the screw diameter HJ, screw speed NH, and screw groove depth MC from the denoised time series curve, and calculate the instantaneous shear rate of the material in the screw groove. The calculation logic is as follows: ,in Pi is used to convert the circular motion of the screw rotation speed into linear velocity calculation, obtain the current head pressure, and perform time integration based on the instantaneous shear rate and head pressure to obtain the mixing intensity FX. Its calculation logic is as follows: Where yc is the instantaneous shear rate and P is the head pressure. The symbol for definite integral indicates the cumulative calculation of instantaneous products over the interval [t1, t2]. This indicator comprehensively reflects the material mixing effect under the combined influence of shearing and flow resistance. It integrates instantaneous shear rate and mixing intensity to obtain the intensity feature. Instantaneous power time series data of the denoised time series curve are extracted, and the cumulative energy input is obtained by time integration of the instantaneous power time series data. The total feed mass of the stable section is obtained by time integration of the feed rate time series data within the time period. The cumulative energy input is divided by the total feed mass to obtain the energy input per unit mass of material. This indicator reflects the density of energy absorbed by the material and is directly related to the melting and mixing effect of the material. The cumulative energy input and the energy input per unit mass of material are integrated into the energy input feature. The torque curve is decomposed into multiple scales by wavelet transform to obtain wavelet coefficients at different scales. The probability distribution of the wavelet coefficients is calculated, and the texture entropy is calculated based on the probability distribution. The smaller the texture entropy, the lower the irregularity of the torque curve. The process trajectory of the same batch with the same formula and target performance in the historical small test is selected as the reference trajectory. The alignment distance between the current batch trajectory and the reference trajectory is calculated by the dynamic time warping algorithm. The smaller the distance, the higher the consistency between the current process trajectory and the high-quality trajectory. The alignment distance and texture entropy are integrated into the stability feature. The purpose of the equipment scale descriptor is to quantify the cross-scale differences between pilot-scale equipment and large-scale mass production equipment, providing key equipment-dimensional information support for the model to learn the process and performance mapping rules from pilot-scale to mass production, and avoiding prediction distortion caused by equipment differences. In the geometric scale ratio dimension, focusing on the size matching relationship of the core working area of ​​the equipment, the ratio of the screw diameter product of the pilot-scale equipment and the barrel length of the large-scale equipment is calculated to obtain the geometric scale ratio. This ratio directly reflects the scale differences in the internal material holding space and heat and mass transfer paths of the equipment. In the heating zone capacity difference dimension, the number of heating zones is used as the core indicator, calculating the difference in the number of heating zones between the large-scale equipment and the pilot-scale equipment as a percentage of the number of heating zones in the large-scale equipment. The ratio of quantities is used to characterize the difference in the precision of temperature control between the two. The larger the value, the more precise the temperature zone control of the large-scale equipment and the more significant the difference in heating capacity between it and the pilot-scale equipment. In terms of the difference in feeding and vacuuming capabilities, the differences in both material conveying capacity and environmental control capability are comprehensively considered. The difference in feeding capacity is calculated by the ratio of the difference in the maximum feeding rate between the large-scale and pilot-scale equipment to the maximum feeding rate of the large-scale equipment. The difference in vacuuming capability is calculated by the ratio of the difference in the maximum vacuum degree between the large-scale and pilot-scale equipment to the maximum vacuum degree of the large-scale equipment. The two ratios are added together to obtain the difference in vacuuming capability. The geometric scale ratio, heating zone capability difference, feeding capability difference, and vacuuming capability difference are integrated into an equipment scale descriptor. The input feature vector is obtained by fusing the small-scale full-link input vector, thermal history features, intensity features, energy input features, stability features, and equipment scale descriptors.

[0020] As a preferred technical solution, based on step three, a cross-scale prediction model is constructed using Transformer and MLP based on the paired sample set, outputting predicted values ​​for mass production performance and production capacity indicators. The specific steps are as follows: The paired sample set is divided into training, validation, and test sets proportionally. The training set samples are input into the model, which outputs mass production performance predictions and capacity index predictions. The loss value of the current batch is calculated based on the composite objective function. The gradient of each layer parameter is calculated using the gradient descent method, and all parameters of the shared feature extraction layer and the task-specific branch are updated. After each round of training, the model performance is evaluated using the validation set. If the validation set loss reaches the current optimal level, the model weights are saved. When the maximum number of iterations is reached or the early stopping strategy is triggered, the optimal cross-scale prediction model is output. The cross-scale prediction model consists of an input layer, a shared feature extraction layer, and task-specific branches. The shared feature extraction layer includes a temporal feature subnetwork, a static feature subnetwork, and a fusion layer. The task-specific branches include a mass production sample performance mapping branch and a mass production capacity index prediction branch. The input feature vector is sent to the input layer. The temporal feature subnetwork uses a single-layer bidirectional LSTM to analyze the dynamic changes of thermal history features, intensity features, energy input features, and stability features. The static feature subnetwork layer is a fully connected layer that analyzes the static features of the pilot-scale full-link and equipment scale descriptors. The fusion layer concatenates the dynamic and static features into a shared feature vector. The mass production sample performance mapping branch outputs the mass production performance prediction value, and the mass production capacity index prediction branch outputs the capacity index prediction value.

[0021] As a preferred technical solution, based on step four, an amplified risk index is constructed, and optimal process parameters for the large-scale machine are generated using a multi-objective genetic algorithm. The specific steps are as follows: Set the target specification range for mass production performance indicators and obtain the predicted mass production performance values. and the standard deviation of the predicted value And calculate the performance deviation probability, the calculation logic is as follows: ,in Let be the cumulative distribution function of the standard normal distribution. and Given the minimum and maximum values ​​of the target specification range, based on the uncertainty and stability characteristics of the capacity indicators output by the model, historical data is fitted through a logistic regression model to obtain the probability of process instability. The difference between the predicted capacity indicator and the standard capacity indicator is calculated to obtain the capacity deviation value. The risk index is obtained by weighted summation of the performance deviation probability, the capacity deviation value, and the process instability probability. The risk index is mapped to a four-level early warning system. If the risk index is less than the first risk threshold, it is judged as low risk. If the first risk threshold is less than the risk index and less than the second risk threshold, it is judged as medium risk, indicating that there is a single dimension of risk that is too high. The corresponding process parameters need to be adjusted. If the second risk threshold is less than the risk index and less than the third risk threshold, it is judged as high risk. After supplementing the pilot test, the prediction is re-evaluated. If the risk index is greater than or equal to the third risk threshold, it is judged as the highest risk. Pilot process optimization needs to be carried out again, and mass production is not started for the time being. The process parameters of each major machine are converted into chromosomes using real-number encoding, with each gene locus corresponding to the value of a process parameter. Based on the historical mass production process parameter database and equipment constraints, an initial population is randomly generated. Then, through selection, crossover, and mutation operations, the top few individuals with the highest fitness are retained to form a new generation population. The fitness calculation logic is as follows: H1 and H2 are the target weights. and The maximum expected qualified output and maximum energy consumption per unit of output from historical data are used to normalize the target value. and To determine the actual qualified output and actual energy consumption under the corresponding process parameter A; stop iterating when the maximum number of iterations is reached, output the optimal process parameters of the mainframe, and send them to the process execution terminal for control.

[0022] As a preferred technical solution, based on step five, when the cross-scale prediction model determines that the current formulation does not meet the standards, the system automatically triggers a small-scale formulation recommendation based on Bayesian optimization, generates optimized component ratios, and reassesses mass production risks. The specific steps are as follows: It should be noted that this step, as an optional function of the cross-scale prediction model, is linked with thermal history features and equipment scale descriptors to form a local closed loop from "small-scale formulation → performance prediction → formulation optimization → re-prediction". The system determines that the current formula is substandard and triggers the formula optimization process when the mass production prediction result output by the cross-scale prediction model meets any of the following conditions: If the predicted mass production performance value exceeds the target specification range, the predicted capacity index is lower than the equipment design standard value, and the scale-up risk index is greater than or equal to the third risk threshold, it is judged as an extremely high risk and the formula needs to be re-implemented for small-scale process optimization. Once triggered, the system automatically extracts the core information of the current formulation, including the resin system type, mass fraction of each filler and additive, raw material batch number, small-scale process characteristics and equipment scale descriptor, and sets it as the initial input for Bayesian optimization. Gaussian process regression is set as a surrogate model for Bayesian optimization. Taking the current substandard formula as the center, several initial formula samples that meet the constraints are randomly generated in the optimization space. All samples are input into the cross-scale prediction model to obtain the corresponding objective function values ​​and form the initial sample set. Using the initial sample set as training data, a Gaussian process regression surrogate model is constructed, with the kernel function set as the radial basis function, and the posterior probability distribution of the objective function is obtained by fitting. The expected improvement is used as the acquisition function to calculate the expected improvement value of all candidate points in the optimization space, and the point with the largest improvement value is selected as the next formulation sample to be evaluated. The formula sample is input into the cross-scale prediction model and the objective function value is calculated. The calculation logic is as follows: Where Q1, Q2, and Q3 are preset proportional weights, PH is the probability of the optimized formula achieving the core performance indicators in mass production, HG is the predicted value of the qualified output per unit time of the optimized formula, and JS is the predicted value of the energy consumption per unit output of the optimized formula. and These are the corresponding equipment design standard values; Add the formula sample and the objective function value to the sample set to obtain the updated sample set. Repeat the iteration until the formula sample with the largest objective function value is output as the optimal recommended formula. The system standardizes the optimal recommended formula obtained by Bayesian optimization and outputs the component ratios of the optimized formula.

[0023] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0024] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0025] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0026] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0027] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0028] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0029] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0030] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0031] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting production capacity using an AI model for polymer composite material processes, characterized in that, Includes the following steps: Step 1: Generate a unique batch number for each pilot experiment, collect heterogeneous data of various types, process the heterogeneous data through a data standardization algorithm, and complete the batch integrity verification. Step 2: Construct a paired sample set from the heterogeneous data, and extract the thermal history features, shear and mixing intensity features, energy input features, and stability features of the pilot process parameters. Then, fuse the equipment scale descriptor to obtain the input feature vector. Step 3: Based on the paired sample set, construct a multi-task cross-scale prediction model using Transformer and MLP. Input the input feature vector into the multi-task cross-scale prediction model and output the mass production performance prediction value and the capacity index prediction value. Step 4: Construct an amplified risk index based on the predicted mass production performance and capacity indicators, and generate the optimal process parameters for the large-scale machine using a multi-objective genetic algorithm; In step five, when the cross-scale prediction model determines that the current formulation does not meet the standards, the system automatically triggers a small-scale formulation recommendation based on Bayesian optimization, generates optimized component ratios, and reassesses the risks of mass production.

2. The method for predicting production capacity using an AI model for polymer composite material processes according to claim 1, characterized in that, A unique batch number is generated for each pilot experiment, and various types of heterogeneous data are collected. The specific steps are as follows: A unique batch number is generated for each small-scale test, and a record of raw material parameters and formulation feeding is established corresponding to the batch number, including resin system information, filler and additive types and their mass fractions, raw material suppliers and incoming batch numbers, formulation version number, target performance indicators and test standards. Then, real-time data collection is performed on the small-scale process parameters, small-scale sample performance parameters, and large-scale production verification data. The small-scale process parameters include the screw extruder technical parameters, screw configuration parameters, and control setpoints and sensor feedback values, including parameters such as screw speed, heating zone temperature, melt temperature, heating zone barrel volume, material density, feeding rate, and vacuum degree. The small-scale sample performance parameters include tensile strength, flexural strength, and impact strength. The large-scale production verification data includes the production verification batch number matching the small-scale test batch number, production process parameters, production sample performance, and production capacity indicators.

3. The method for predicting production capacity using an AI model for polymer composite material processes according to claim 2, characterized in that, Heterogeneous data is processed using data standardization algorithms to complete batch integrity verification. The specific process is as follows: The pilot-scale experimental data was divided into numerical, character, and time-series types. Numerical data was mapped to the 0-1 range using Z-score normalization to obtain standard numerical data. Character data was mapped to integer codes using an industry-specific classification dictionary. One-hot encoding was performed on the integer codes to generate sparse character vectors. Using the pilot-scale experiment start time as zero, the timestamps of all time-series data were converted to relative times. Linear interpolation was used to resample time-series data from different sampling frequencies to a unified frequency, resulting in standard time-series data. Each field value of the standardized time-series data was iterated through, and the field name and value were concatenated into a string. This string was then processed using MID (Meanwhile, the MID data was processed.) The algorithm generates a field-level hash value HM and sets it as the leaf node of the current Merkle tree. It then concatenates the field-level hash values ​​of all fields within a single record in ascending order of field name to generate a record-level hash. Next, it concatenates the record-level hash values ​​of all records within the collected data in ascending order of primary key to generate a table-level hash. Finally, it concatenates the data in ascending order of table-level hashes to generate a batch-level Merkle root hash. The batch-level Merkle root hash is compared with a preset batch hash baseline value. If they match exactly, the data is determined not to have been tampered with. If they do not match, the Merkle tree's hash backtracking algorithm is used to locate the tampered node and correct it to obtain corrected data.

4. The method for predicting production capacity using an AI model for polymer composite material processes according to claim 1, characterized in that, The heterogeneous data are used to construct paired sample sets, and the thermal history characteristics, intensity characteristics, energy input characteristics, and stability characteristics of the pilot-scale process parameters are extracted. The specific steps are as follows: By using the unique connection key of the batch number of the pilot test, the raw material parameters, formula feeding records, pilot process parameters and pilot sample performance parameters are connected through a multi-table internal join algorithm to obtain structured data. The structured data is then converted into a pilot full-link input vector. The mass production verification results are set as supervision labels and paired with the corresponding pilot full-link input vector to generate a paired sample set. Arrange the pilot-scale process parameters in chronological order to obtain time series curves. Perform several layers of wavelet decomposition on a single time series curve to obtain approximation coefficients and detail coefficients. Calculate an adaptive threshold based on the detail coefficients. Reconstruct the detail coefficients that are greater than the adaptive threshold to obtain denoised time series curves. Extract the feeding rate, material density, barrel volume of each heating zone, and screw channel filling coefficient from the noise-reduced time-series curve, and calculate the material residence time of each heating zone; Multiply the average temperature of each temperature zone by the corresponding residence time, and sum them to obtain the total heat exposure intensity. This index directly reflects the cumulative heat absorbed by the material during processing. Then, calculate the standard deviation and mean of the product of temperature and residence time for each temperature zone. Subtract the ratio of the standard deviation to the mean from 1 to obtain the heat exposure uniformity. The closer this ratio is to 1, the better. Query the thermal decomposition temperature of the current resin system and compare the highest temperature of each temperature zone with the thermal decomposition temperature. If the highest temperature of any temperature zone exceeds the thermal decomposition temperature, the overheating risk index is set to 1; otherwise, it is set to 0. This is used to determine whether the material has a risk of thermal degradation. The total heat exposure intensity, heat exposure uniformity, and thermal degradation risk are integrated into a thermal history feature. Extract the screw diameter, screw speed, and screw groove depth from the denoised time-series curve, calculate the instantaneous shear rate of the material in the screw groove, obtain the current die head pressure, and perform time integration based on the instantaneous shear rate and die head pressure to obtain the mixing intensity; Instantaneous shear rate and mixing intensity are integrated to obtain intensity characteristics; instantaneous power time series data of denoised time series curves are extracted, and cumulative energy input is obtained by time integration of instantaneous power time series data; total feed mass in the steady segment is obtained by time integration of feed rate time series data within the time period; energy input per unit mass of material is obtained by dividing cumulative energy input by total feed mass; and cumulative energy input and energy input per unit mass of material are integrated into energy input characteristics. The torque curve is decomposed into multiple scales by wavelet transform to obtain wavelet coefficients at different scales. The probability distribution of the wavelet coefficients is calculated. The process trajectory of a high-quality batch with the same formula and target performance in historical small-scale tests is selected as the reference trajectory. The alignment distance between the current batch trajectory and the reference trajectory is calculated by dynamic time warping algorithm. The alignment distance and texture entropy are integrated into a stability feature.

5. The method for predicting production capacity using an AI model for polymer composite material processes according to claim 1, characterized in that, The specific steps for obtaining the input feature vector by fusing device scale descriptors are as follows: The geometric scale ratio is obtained by calculating the ratio of the product of the screw diameter and the barrel length of the pilot-scale equipment to that of the large-scale equipment. The heating zone capacity difference is obtained by calculating the ratio of the difference in the number of heating zones between the large-scale equipment and the pilot-scale equipment to the total number of heating zones in the large-scale equipment. The feeding capacity difference is calculated by calculating the ratio of the difference in the maximum feeding rate between the large-scale equipment and the pilot-scale equipment to the maximum feeding rate of the large-scale equipment. The vacuum capacity difference is obtained by calculating the ratio of the difference in the maximum vacuum degree between the large-scale equipment and the pilot-scale equipment to the maximum vacuum degree of the large-scale equipment. The two ratios are added together to obtain the vacuuming capacity difference. The geometric scale ratio, heating zone capacity difference, feeding capacity difference, and vacuuming capacity difference are integrated into an equipment scale descriptor. The input feature vector is obtained by fusing the small-scale full-link input vector, thermal history features, intensity features, energy input features, stability features, and equipment scale descriptors.

6. The method for predicting production capacity using an AI model for polymer composite material processes according to claim 1, characterized in that, Based on a paired sample set, a cross-scale prediction model is constructed using Transformer and MLP to output predicted values ​​for mass production performance and capacity indicators. The specific steps are as follows: The paired sample set is divided into training, validation, and test sets proportionally. The training set samples are input into the model, which outputs mass production performance predictions and capacity index predictions. The loss value of the current batch is calculated based on the composite objective function. The gradient of each layer parameter is calculated using the gradient descent method, and all parameters of the shared feature extraction layer and the task-specific branch are updated. After each round of training, the model performance is evaluated using the validation set. If the validation set loss reaches the current optimal level, the model weights are saved. When the maximum number of iterations is reached or the early stopping strategy is triggered, the optimal cross-scale prediction model is output. The cross-scale prediction model consists of an input layer, a shared feature extraction layer, and task-specific branches. The shared feature extraction layer includes a temporal feature subnetwork, a static feature subnetwork, and a fusion layer. The task-specific branches include a mass production sample performance mapping branch and a mass production capacity index prediction branch. The input feature vector is sent to the input layer. The temporal feature subnetwork uses a single-layer bidirectional LSTM to analyze the dynamic changes of thermal history features, intensity features, energy input features, and stability features. The static feature subnetwork layer is a fully connected layer that analyzes the static features of the pilot-scale full-link and equipment scale descriptors. The fusion layer concatenates the dynamic and static features into a shared feature vector. The mass production sample performance mapping branch outputs the mass production performance prediction value, and the mass production capacity index prediction branch outputs the capacity index prediction value.

7. The method for predicting production capacity using an AI model for polymer composite material processes according to claim 1, characterized in that, The specific steps for constructing an amplified risk index and mapping it to a four-level early warning system are as follows: Set the target specification range for mass production performance indicators, obtain the predicted mass production performance value and the standard deviation of the predicted value, and calculate the performance deviation probability. Based on the uncertainty and stability characteristics of the capacity indicators output by the model, fit historical data through a logistic regression model to obtain the process instability probability. Subtract the predicted capacity indicator value from the standard capacity indicator to obtain the capacity deviation value. Calculate the risk index by weighting and summing the performance deviation probability, the capacity deviation value, and the process instability probability. The risk index is mapped to a four-level early warning system. Specifically, if the risk index is less than the first risk threshold, it is judged as low risk; if the first risk threshold is less than or equal to the risk index and less than the second risk threshold, it is judged as medium risk, indicating that there is a single dimension of risk that is too high, and the corresponding process parameters need to be adjusted. If the second risk threshold is less than or equal to the risk index and less than the third risk threshold, then the risk is considered high and a new prediction is made after a supplementary small-scale experiment. If the risk index is greater than or equal to the third risk threshold, then the risk is considered the highest, and a small-scale process optimization needs to be carried out again, and mass production will not be started for the time being.

8. The method for predicting production capacity using an AI model of polymer composite material processing according to claim 7, characterized in that, The specific steps for generating optimal process parameters for large-scale machines using a multi-objective genetic algorithm are as follows: The process parameters of each machine are converted into chromosomes by real number encoding. Each gene position corresponds to the value of a process parameter. Based on the historical mass production process parameter database and equipment constraints, an initial population is randomly generated. Through selection, crossover and mutation operations, the top few individuals with the highest fitness are retained to form a new generation population. The iteration stops when the maximum number of iterations is reached, the optimal process parameters of the mainframe are output, and sent to the process execution terminal for control.

Images