AI-driven polymer composite material process optimization method

Abstract

The invention discloses an AI-driven polymer composite material process optimization method, and aims to solve the problems of data islands, process optimization lag and insufficient model timeliness in a polymer material production process. The method comprises the following steps: collecting full-link data according to a six-level customer product coding specification; the state parameters of the high-frequency equipment are safely stored in the sub-table 1 through encryption and identity authentication; constructing a structured database based on the production batch number association main table, the raw material sub-table set, the sub-table 2 and the sub-table 3; training a multi-model artificial intelligence system fusing gradient boosting regression, Bayesian optimization, a neural network and a random forest, and realizing a bidirectional linkage closed loop of formula recommendation, process optimization and performance prediction; and incremental learning is carried out by adopting a sliding window mechanism in combination with an online gradient descent and elastic weight consolidation strategy. According to the technical scheme, intelligent, efficient and safe optimization of the high polymer material process can be achieved, and the product quality and the production efficiency are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and polymer material processing technology, specifically relating to an AI-driven process optimization method for polymer composite materials. Background Technology

[0002] Manufacturing Execution Systems (MES) serve as a crucial hub connecting upper-level planning and management systems (ERP systems) with lower-level equipment control. Their data acquisition and collaboration capabilities directly impact the accuracy and response speed of process optimization. However, in the current polymer material production process, multi-dimensional data involving raw material parameters, formulation input, process conditions, and quality feedback have long been stored in a scattered manner, lacking unified coding standards and structured association mechanisms. This makes it difficult to effectively integrate end-to-end data, including ERP system data, hindering the construction and application of intelligent decision-making models.

[0003] While data-driven process optimization methods based on MES have been explored in the initial stages, existing technologies are still limited to improvements in single links or isolated dimensions. For example, some solutions focus on factory-level production early warning and data monitoring, but fail to establish dynamic correlations with core elements such as raw material batches, formula bidding, and customer feedback, resulting in significant prediction biases in data-driven models due to insufficient input dimensions. Other studies attempt to use machine learning to screen composite material components to optimize specific properties (such as flame retardancy), but ignore the impact of key equipment operating parameters on molding quality and lack a batch-level traceability system covering the entire cycle of "raw materials—formula—process—testing—feedback," thus failing to support closed-loop process optimization.

[0004] Existing technologies generally suffer from three major bottlenecks: data fragmentation, lack of security protection, and low degree of structuring. On the one hand, the lack of encryption and authentication mechanisms during cross-system data transmission makes sensitive information such as formulas and processes susceptible to leakage or tampering. On the other hand, the database design does not adopt a master-sub-table linkage architecture, making it difficult to efficiently organize heterogeneous data such as raw materials, actual formulas, equipment status, internal quality inspection, and customer feedback, resulting in sparse training samples and chaotic feature coupling in AI models. Especially in multi-process adaptation scenarios for thermoplastic and thermosetting materials, the models lack dynamic perception of molding constraints such as injection molding, compression molding, or pultrusion, making it difficult to accurately identify key factors affecting product performance, and thus failing to achieve cost-controllable formula recommendations and collaborative optimization of process parameters. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-driven process optimization method for polymer composite materials, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An AI-driven process optimization method for polymer composite materials includes the following specific steps:

[0008] Step S1: Based on the standardized customer product coding specifications, collect raw material data, formula input data, production process data, quality test data, and quality feedback data associated with the production batch number;

[0009] Step S2: The Manufacturing Execution System collects formula feeding data, batch average process parameters, process parameter fluctuation range, and key equipment status parameters during the production process. After encryption and authentication, the data is securely transmitted and stored in sub-table 1 of the production management database.

[0010] Step S3: Based on the end-to-end data composed of the main table and all sub-tables integrated through the production batch number, train a multi-model fusion artificial intelligence model; the main table stores quality test data, and the sub-tables include sub-tables storing formula bidding, actual bidding data and raw material sub-tables, sub-table 1 storing production process data, sub-table 2 storing internal quality inspection issues and sub-table 3 storing customer product feedback and customer product code data.

[0011] Step S4: Use the trained artificial intelligence model to perform the tasks of formula and equipment process recommendation and product performance prediction, and verify and optimize through a two-way linkage mechanism: when the performance prediction determines that the user's self-made formula does not meet the standards, the formula recommendation function is automatically triggered for optimization, forming a design-evaluation-optimization closed loop.

[0012] Step S5: The production end adjusts process parameters and conducts production verification based on the optimization results output by the artificial intelligence model;

[0013] Step S6: Based on the sliding window mechanism, incremental learning is performed on the artificial intelligence model periodically using newly added production data to maintain the model's prediction accuracy and timeliness.

[0014] Preferably, in step S1, the customer product coding specification adopts a six-digit hierarchical structure consisting of a two-digit usage scenario code, a two-digit main product code, and a two-digit component specification code. This structure enables standardized classification and rapid retrieval of product information, ensuring the consistency and integrity of data collection.

[0015] Preferably, the key equipment status parameters in step S2 include the current of the extruder main unit and each zone, the current of the loss-in-weight weighing device, etc., and the acquisition cycle is no more than 1 second. High-frequency data acquisition can reflect the equipment operating status in real time and provide an accurate data basis for process optimization.

[0016] Preferably, in step S2, the data transmission process is encrypted using an advanced encryption standard algorithm and combined with a digital certificate for identity authentication, ensuring the security and integrity of sensitive information such as formulas and process parameters during transmission.

[0017] Preferably, in step S3, the main table and the sub-table are associated through the production batch number to form a structured full-link data storage system. The main table and the raw material sub-table set are responsible for storing basic information of raw materials, formula standard deployment data and quality test results. Sub-table 1 stores real-time production process data, sub-table 2 records the problems found in the internal quality inspection process and their classification, and sub-table 3 collects customer feedback information after using the product and customer product code data.

[0018] Preferably, the task of recommending formulas and equipment processes in step S4 is achieved through the collaborative algorithm of gradient boosting regression, Bayesian optimization and random forest. The optimization objective not only considers product performance indicators, but also raw material cost factors, so as to achieve a balance between economic benefits and product quality.

[0019] Preferably, the performance prediction task in step S4 is implemented by a fusion model of backpropagation neural network and gradient boosting tree. This model introduces the internal quality inspection issues in sub-table 2 and the customer product feedback data in sub-table 3 as input features, and outputs the product quality risk level, which is divided into three levels: high, medium and low.

[0020] Preferably, the two-way linkage mechanism in step S4 is specifically manifested as follows: when the performance prediction model determines that the expected performance of the user-designed formula does not meet the preset standard, the system automatically calls the formula recommendation algorithm to generate an optimized formula scheme and feeds the recommendation result back to the design end, forming a complete closed-loop process from design to evaluation to optimization.

[0021] Preferably, the incremental learning in step S6 adopts a strategy that combines online gradient descent and elastic weight consolidation. Online gradient descent is used to quickly adapt to changes in data distribution, while elastic weight consolidation protects existing knowledge from being overwritten by new data by calculating the importance weights of parameters, thereby effectively avoiding catastrophic forgetting.

[0022] Preferably, the method is adapted to the production process of thermoplastic and thermosetting polymer materials, and dynamically adjusts the input feature dimensions and process constraints of the artificial intelligence model according to the specific type of injection molding, compression molding or pultrusion molding process to ensure the applicability and accuracy of the model in different process scenarios.

[0023] Preferably, the method also supports human-computer interaction via Chinese text commands. Users can input query or optimization commands through natural language, and the system automatically parses the command content and generates corresponding process parameter fluctuation curves and performance prediction comparison charts for visualization, assisting in decision analysis.

[0024] Preferably, the problem types, severity, and causes recorded in the quality feedback data are all stored using standardized codes. The problem type codes cover categories such as appearance defects, functional abnormalities, and insufficient durability. The severity is divided into three levels: urgent, important, and general. The problem cause codes include factors such as material formulation, process parameters, and equipment status.

[0025] Preferably, the structured database module adopts a relational database management system. The main table, the raw material sub-table set, and the three additional sub-tables maintain data consistency through foreign key constraints, and a composite index is established to improve the performance of multi-table join queries, ensuring efficient access under large data volumes.

[0026] Preferably, the data security module employs transport layer security protocols for encryption during the entire data transmission process and implements role-based access control policies in the data storage stage, setting different data access permissions for different users to ensure the security of sensitive process data.

[0027] Preferably, the multi-model fusion unit in the artificial intelligence optimization module adopts a stacked ensemble learning method, which uses the prediction results of base models such as gradient boosting regression, Bayesian optimization and random forest as meta-features and inputs them into the meta-learner for final decision-making, thereby improving the overall accuracy of recommendation and prediction.

[0028] Preferably, the sliding window mechanism in the model update module has a time window of 30 days. It automatically detects the amount of new data every 24 hours. When the amount of data reaches 80% of the window capacity, it triggers the incremental learning process to ensure that the model continuously adapts to changes in the production environment.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] By establishing standardized customer product coding specifications and a structured master-slave database architecture, effective integration and correlation of data across the entire polymer material production chain were achieved. A multi-model fusion artificial intelligence algorithm combined with a two-way linkage mechanism enabled intelligent recommendation and performance prediction of formulations and process parameters. Data security was ensured through encrypted transmission and identity authentication, and the timeliness of the model was maintained using an incremental learning strategy. This method supports various polymer materials and molding processes, possesses good adaptability and scalability, and significantly improves the accuracy and efficiency of process optimization. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall technical solution architecture of the polymer material process optimization method based on MES data proposed in this invention; Detailed Implementation

[0032] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0034] Currently, with the widespread application of polymer composite materials in aerospace, new energy vehicles, and high-end equipment, Manufacturing Execution Systems (MES), as the key hub connecting upper-level planning management and lower-level equipment control, directly impact the accuracy and response speed of process optimization through their data acquisition and collaboration capabilities. However, the multi-dimensional data involved in polymer material production processes, such as raw material parameters, formulation input, process conditions, and quality feedback, have long been stored in a scattered manner, lacking unified coding standards and structured association mechanisms. This makes it difficult to effectively integrate data across the entire process, hindering the construction and application of intelligent decision-making models. To address these technical problems, this invention proposes a closed-loop process optimization driven by data throughout the entire lifecycle of polymer materials. This is achieved by establishing standardized customer product coding standards, constructing a structured database with master-sub-table linkage, integrating multi-model artificial intelligence algorithms, and introducing a two-way linkage verification mechanism. The proposed method is applied to an AI-driven process optimization approach for polymer composite materials.

[0035] refer to Figure 1 The overall technical architecture of this invention includes a data acquisition layer, a secure transmission layer, a structured storage layer, an artificial intelligence optimization layer, and a human-computer interaction layer. Specifically, the data acquisition layer relies on the manufacturing execution system to acquire multi-source heterogeneous data covering the entire chain from raw materials to formulations, processes, testing, and feedback in real time; the secure transmission layer employs a dual protection mechanism of advanced encryption standard algorithms and digital certificates to ensure the integrity and confidentiality of sensitive information during cross-system transfer; the structured storage layer uses the production batch number as the core key to construct a relationship system between a main table, a raw material sub-table, and three sub-tables; the artificial intelligence optimization layer integrates various algorithms such as gradient boosting regression, Bayesian optimization, random forest, backpropagation neural network, and gradient boosting tree to form an intelligent decision-making engine with multi-model fusion and bidirectional linkage; and the human-computer interaction layer supports Chinese natural language command parsing and visualized result output to assist engineers in making process optimization decisions.

[0036] In the above-mentioned AI-driven process optimization method for polymer composite materials, step S1 involves collecting raw material data, formula feeding data, production process data, quality test data, and quality feedback data associated with the production batch number, based on standardized customer product coding specifications.

[0037] Specifically, the customer product coding specification adopts a 6-digit structure of "2-digit usage scenario code + 2-digit main product code + 2-digit component / specification code", with each segment clearly defined and adaptable to multiple industries.

[0038] The two-digit usage scenario code consists of uppercase letters, directly corresponding to the core application scenario of the product: "EC" represents electric vehicle-related products, "UA" represents the complete drone, and "RB" represents the complete robot. "XX" is reserved as an extension for new scenarios. The two-digit main product code consists of numbers, defining the complete finished product model for the corresponding scenario: In the "EC" scenario, "01" is the complete DC-DC charging gun, and "05" is the high-voltage connector; in the "UA" scenario, "01" is the single-rotor drone, and "02" is the multi-rotor drone; in the "RB" scenario, "01" is the humanoid bipedal robot, and "02" is the vehicle-shaped wheeled robot, ensuring that the main product level focuses on the complete finished product.

[0039] The 2-digit component / specification code is a number and is defined differently based on the main product type: If the main product is an assembled machine with multiple components (such as a DC charging gun, a single-rotor drone, or a humanoid bipedal robot), this code represents a specific functional component. For example, under "EC01" (integrated DC charging gun), "01" represents the gun body shell; under "UA01" (single-rotor drone), "02" represents the rotor bracket; and under "RB01" (humanoid bipedal robot), "03" represents the leg joint assembly. If the main product is a single-component core product (such as a high-voltage connector), this code represents a specification difference. For example, under "EC05" (high-voltage connector), "02" represents a current adaptation of 200A.

[0040] Typical coding examples: "EC0101" represents the overall DC charging gun housing for electric vehicles; "EC0502" represents the high-voltage connector for electric vehicles that adapts to 200A current; "UA0102" represents the rotor bracket for a single-rotor drone; and "RB0103" represents the leg joint component for a humanoid bipedal robot. Through unified coding, precise identification is provided for the association of data across the entire chain, ensuring that each production batch can be uniquely mapped to a specific product model in a specific application scenario, thereby achieving consistency and integrity of data collection at the source.

[0041] Raw material data includes supplier code, batch number, and basic physical properties (such as melt flow index, glass transition temperature, and tensile strength); formulation and feeding data covers the names of each component, theoretical feeding ratio, actual weighing value, and deviation rate; production process data records the molding method (injection molding, compression molding, or pultrusion), feeding speed, screw speed, die head pressure, temperature, current, and vacuum level of each zone of the extruder; quality testing data includes test reports on mechanical properties, thermal stability, and flame retardancy rating issued by the internal laboratory; quality feedback data includes internal quality inspection data in sub-Table 2 and customer product feedback data in sub-Table 3.

[0042] In the above-mentioned AI-driven process optimization method for polymer composite materials, in step S2, the manufacturing execution system collects the formula feeding data, batch average process parameters, process parameter fluctuation range, and key equipment status parameters during the production process, and after encryption and authentication, it securely transmits and stores them in sub-table 1 of the production management database.

[0043] Specifically, the key equipment status parameters include the current of the extruder main unit and each zone, the loss-in-weight weighing current, etc., with a sampling frequency of no more than 1 second. Taking a twin-screw extruder as an example, it is equipped with 4 independent temperature control zones, each equipped with a current sensor with a sampling period of 0.5 seconds; the die pressure is monitored in real time by a piezoelectric pressure sensor with a range of 0 to 50 MPa and an accuracy of ±0.1 MPa; the loss-in-weight weighing current is collected by a Hall current sensor with a dynamic range of 0 to 65 amperes. All raw signals are aggregated to the edge computing gateway via an industrial fieldbus (such as PROFINET). After data preprocessing (including filtering, noise reduction, and unit unification) is completed within the gateway, the secure transmission process is initiated.

[0044] The data transmission process employs Advanced Encryption Standard (AES) algorithms for encryption and utilizes digital certificates for authentication. Specifically, the Manufacturing Execution System (MES) client first requests a session key from the authentication server. After verifying the validity of the digital certificate (including the certificate authority's signature, validity period, and revocation status), the authentication server issues a 256-bit AES session key. All subsequent data packets are symmetrically encrypted using this key and appended with a Message Authentication Code (MAC) to prevent tampering. The encrypted data stream is then uploaded to the cloud-based production management database via a Transport Layer Security (TLS) protocol channel, ensuring the security and integrity of sensitive information such as formulas and process parameters during transmission.

[0045] The field structure of sub-table 1 includes: production batch number (foreign key), timestamp sequence, extruder current sequence, screw speed sequence, die pressure sequence, temperature sequence of each heating zone, average process parameters of the batch (e.g., average die pressure = 19.1 MPa, average screw speed = 120 rpm), and the fluctuation range of process parameters, i.e., the maximum and minimum values ​​of the batch process parameters (e.g., the maximum die pressure is 20.3 MPa, the minimum is 18.5 MPa, the maximum heating zone temperature is 314℃, and the minimum is 307℃).

[0046] In the above-mentioned AI-driven process optimization method for polymer composite materials, step S3 involves training a multi-model fusion artificial intelligence model based on the end-to-end data composed of the main table and all sub-tables integrated through the production batch number; the main table stores quality test data, and the sub-tables include a sub-table 1 storing formula bidding data, actual bidding data, and raw material sub-tables, a sub-table 1 storing production process data, a sub-table 2 storing internal quality inspection issues, and a sub-table 3 storing customer product feedback and customer product code data.

[0047] Specifically, the main table and sub-tables are linked through production batch numbers, forming a structured end-to-end data storage system. The main table fields include: production batch number (primary key), and quality test data. Raw material parameters are categorized into a series of sub-tables based on different raw material types (e.g., PA raw materials, thermal conductive powder, toughening powder, color powder, etc.). These sub-tables contain raw material supplier numbers, costs, theoretical and actual proportions of each component, and laboratory test results. Laboratory test items for raw materials are recorded based on the different functions of the raw materials (e.g., impact resistance testing for structural components, color difference testing for appearance components, durability testing for cover components), recording data on the mechanical and processing properties of different types of raw materials (e.g., PA raw materials, thermal conductive powder, toughening powder, color powder, etc.). All data in the main table and raw material sub-tables is automatically collected by the ERP system. Sub-table 2 records problems discovered during internal quality inspection and their classifications. Its fields include: production batch number (foreign key), problem type code, severity code, and problem cause code. Sub-table 3 collects customer feedback after product use. Its fields include: production batch number (foreign key), problem type code, severity code, and problem cause code. The problem type code uses a three-digit code; for example, "101" indicates appearance defects (such as color difference, flow marks), "202" indicates functional abnormalities (such as dimensional deviations, assembly failures), and "303" indicates insufficient durability (such as early cracking, performance degradation). Severity is divided into three levels: urgent (code "1"), important (code "2"), and minor (code "3"). The problem cause code includes factors such as material formulation (code "A"), process parameters (code "B"), and equipment status (code "C"). The structured database module uses a relational database management system. Foreign key constraints maintain data consistency between the main table, the raw material sub-table set, and the three supplementary sub-tables, and composite indexes are established to improve the performance of multi-table join queries. For example, B+ tree indexes are created on the "Production Batch Number" field of the main table and the "Production Batch Number" field of the raw material sub-tables and supplementary sub-tables 1, 2, and 3, respectively. A composite index is created on the "Problem Type Code" and "Severity Code" of sub-table 2. This allows the query response time to be controlled within 200 milliseconds when it is necessary to retrieve all batches that are "functionally abnormal" (problem type code "202") due to "excessive mold head pressure fluctuation" (problem cause code "B") and have a severity of "important" (code "2"). This ensures efficient access even with hundreds of millions of data points.

[0048] Based on this structured data, a multi-model fusion artificial intelligence model is trained. The multi-model fusion unit in the AI ​​optimization module employs a stacked ensemble learning method, using the prediction results of base models such as gradient boosting regression, Bayesian optimization, and random forest as meta-features, which are then input into the meta-learner for final decision-making. Base model training uses 5-fold cross-validation, with gradient boosting regression iterating 1000 times and a learning rate of 0.05. The meta-learner is a fully connected neural network with an input layer dimension matching the number of base models, hidden layers containing 64 neurons using the ReLU activation function, and an output layer using a linear activation function. Training employs the Adam optimizer, with a loss function of mean squared error, a batch size of 32, and 200 iterations.

[0049] The input feature vector dimensions of the base model are dynamically adjusted based on the material type and molding process: for injection molding of thermoplastic materials, the feature vector includes 48 dimensions such as raw material melt index, component ratio, average die pressure, maximum and minimum temperature, and screw speed; for compression molding of thermosetting materials, these are replaced with process constraint-related features such as curing time, pressure holding time, and heating rate, resulting in a total of 52 dimensions. The meta-learner uses a shallow fully connected neural network, containing one hidden layer (64 neurons) and a ReLU activation function. The output layer is a linear unit, used for regression prediction of target performance indicators or classification to determine risk levels.

[0050] In the above-mentioned AI-driven process optimization method for polymer composite materials, step S4 utilizes a trained artificial intelligence model to perform tasks such as formula and equipment process recommendation and product performance prediction, and verifies and optimizes them through a two-way linkage mechanism: when the performance prediction determines that the user's self-developed formula does not meet the standards, the formula recommendation function is automatically triggered for optimization, forming a design-evaluation-optimization closed loop.

[0051] Specifically, the task of recommending formulations and equipment processes is achieved through a collaborative approach using gradient boosting regression, Bayesian optimization, and random forest algorithms. The optimization objective considers not only product performance indicators but also raw material costs. Taking the optimization of a flame-retardant polycarbonate product for new energy vehicle battery casings as an example, the target performance includes a UL94 V-0 flame retardant rating, a heat distortion temperature ≥130 degrees Celsius, and a notched impact strength ≥60 kJ / m², while requiring the total raw material cost to not exceed 35 yuan / kg. The gradient boosting regression model first fits the performance-formulation mapping relationship based on historical data, outputting a preliminary recommended solution. The Bayesian optimization algorithm then constructs a Gaussian process surrogate model based on this, using the expected improvement amount as the acquisition function to iteratively search for the Pareto optimal solution set. The random forest model is used to evaluate the marginal impact of the proportions of each component on cost, ensuring that the recommended solution is within cost constraints. The outputs of these three models are used as meta-features input into a stacked ensemble model, ultimately generating an optimized formulation that balances performance and cost.

[0052] Meanwhile, the performance prediction task is achieved through a fusion model of a backpropagation neural network and a gradient boosting tree. This model incorporates internal quality inspection issues from sub-table 2 and customer feedback data from sub-table 3 as input features, and outputs the product quality risk level. The backpropagation neural network adopts a three-layer structure (128-dimensional input layer, 64-dimensional hidden layer, and 3-dimensional output layer). The output layer is normalized by the Softmax function to a probability distribution of high, medium, and low risk levels. The gradient boosting tree is an ensemble of 100 decision trees, each with a maximum depth of 6 and a learning rate of 0.1. The predicted probability vectors of the two are concatenated and then input into a logistic regression classifier for final determination. ,in, For the final risk probability, and These are the risk probability vectors output by the neural network and the gradient boosting tree, respectively. For learnable weights, For bias terms, This refers to the Softmax function. When the performance prediction model determines that the expected performance of a user-designed recipe does not meet preset standards (e.g., the predicted risk level is "high" or key performance indicators are below preset thresholds), the system automatically calls the recipe recommendation algorithm to generate an optimized recipe and feeds the recommendation results back to the design end, forming a complete closed-loop process from design to evaluation to optimization. This two-way linkage mechanism ensures that any unverified recipe changes must undergo rigorous evaluation by the artificial intelligence model, avoiding quality risks caused by subjective experience.

[0053] In the above-mentioned AI-driven process optimization method for polymer composite materials, in step S5, the production end adjusts the process parameters and conducts production verification based on the optimization results output by the artificial intelligence model.

[0054] Specifically, the optimization results are issued to the Manufacturing Execution System (MES) in the form of structured instructions. These instructions include the target formulation component ratios, recommended screw speed range (e.g., 115 to 125 rpm), die pressure setpoints (e.g., 33.0 ± 0.5 MPa), and target temperatures for each temperature zone. After parsing the instructions, the MES automatically updates the equipment control parameters, adjusts the formulation component ratios, and initiates a small-batch trial production process. During the trial production, sub-table 1 continuously records the actual operating parameters, while the main table synchronously collects the first-piece inspection results. If the quality test data of the trial production batch deviates from the model's predicted values ​​by less than 5%, the verification is considered successful, and the optimization scheme is formally incorporated into the standard process library; otherwise, the system feeds back the deviation data to the model training module, triggering a local retraining process.

[0055] In the above-mentioned AI-driven process optimization method for polymer composite materials, step S6 involves periodically using newly added production data to incrementally learn the artificial intelligence model based on a sliding window mechanism, in order to maintain the model's prediction accuracy and timeliness.

[0056] Specifically, the sliding window mechanism in the model update module has a time window of 30 days. It automatically detects new data every 24 hours, and triggers the incremental learning process when the data volume reaches 80% of the window capacity. Incremental learning employs a strategy combining online gradient descent and elastic weight consolidation. Online gradient descent is used to quickly adapt to changes in data distribution, with an initial learning rate of 0.01 that decays exponentially with each training epoch. Elastic weight consolidation protects existing knowledge from being overwritten by new data by calculating the importance weights of the parameters. (Parameter importance weights) The calculation formula is as follows: in, Indicates to from All possible sampling Find the average. This represents the distribution of old data (i.e., historical data within the sliding window). Here is the loss function. Indicates input data and parameters The loss value below, For the first There are [number] model parameters. During the incremental training phase, the total loss function is adjusted to:

[0057] ,in, For the total loss function, For the loss function on the new data, These are the model parameters before incremental learning. The regularization coefficient is set to 5000, and the importance weights are calculated based on batches of old data. The initial learning rate for online gradient descent is set to 0.01, and the decay coefficient per round is 0.995. This strategy effectively avoids catastrophic forgetting, ensuring that the model retains its ability to recognize historical conditions while absorbing new knowledge.

[0058] Furthermore, the method is adaptable to the production processes of thermoplastic and thermosetting polymer materials, and dynamically adjusts the input feature dimensions and process constraints of the artificial intelligence model according to the specific type of injection molding, compression molding, or pultrusion molding process. For example, in the pultrusion molding scenario, the system automatically shields features related to injection molding, such as "die pressure," and enables exclusive features such as "traction speed" and "curing oven temperature distribution," and incorporates process constraints such as fiber wetting and resin content uniformity into the optimization objectives. Simultaneously, the method supports human-computer interaction via Chinese text commands. Users can input query or optimization commands through natural language, such as "query the die pressure fluctuation curve for customer product code EC0101, batch 2024051201" or "recommend a low-cost flame-retardant formula for new energy vehicle battery casings." The system automatically parses the command content and generates corresponding process parameter fluctuation curves and performance prediction comparison charts for visualization, assisting in decision analysis.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-driven process optimization method for polymer composite materials, characterized in that... This includes the following specific steps: Step S1: Based on the standardized customer product coding specifications, collect raw material data, formula input data, production process data, quality test data, and quality feedback data associated with the production batch number; Step S2: The Manufacturing Execution System collects formula feeding data, batch average process parameters, process parameter fluctuation range, and key equipment status parameters during the production process. After encryption and authentication, the data is securely transmitted and stored in sub-table 1 of the production management database. Step S3: Based on the end-to-end data composed of the main table and all sub-tables integrated through the production batch number, train a multi-model fusion artificial intelligence model; the main table stores quality test data, and the sub-tables include sub-tables storing formula bidding, actual bidding data and raw material sub-tables, sub-table 1 storing production process data, sub-table 2 storing internal quality inspection issues and sub-table 3 storing customer product feedback and customer product code data. Step S4: Use the trained artificial intelligence model to perform the tasks of formula and equipment process recommendation and product performance prediction, and verify and optimize through a two-way linkage mechanism: when the performance prediction determines that the user's self-made formula does not meet the standards, the formula recommendation function is automatically triggered for optimization, forming a design-evaluation-optimization closed loop. Step S5: The production end adjusts process parameters and conducts production verification based on the optimization results output by the artificial intelligence model; Step S6: Based on the sliding window mechanism, incremental learning is performed on the artificial intelligence model periodically using newly added production data to maintain the model's prediction accuracy and timeliness.

2. The AI-driven process optimization method for polymer composite materials according to claim 1, characterized in that: The customer product coding standard adopts a six-digit hierarchical structure consisting of a two-digit usage scenario code, a two-digit main product code, and a two-digit component specification code.

3. The AI-driven process optimization method for polymer composite materials according to claim 1, characterized in that: The key equipment status parameters include the current of the extruder main unit and each zone, the current of the loss-in-weight scale, etc., and the acquisition cycle is no more than 1 second.

4. The AI-driven process optimization method for polymer composite materials according to claim 1, characterized in that: The data transmission process is encrypted using an advanced encryption standard algorithm and uses a digital certificate for authentication.

5. The AI-driven process optimization method for polymer composite materials according to claim 1, characterized in that: The main table and the sub-table are linked by a foreign key through the production batch number, and the structured database uses a relational database management system and establishes composite indexes to improve the performance of multi-table join queries.

6. The AI-driven process optimization method for polymer composite materials according to claim 1, characterized in that: The task of recommending formulas and equipment processes is achieved through the collaborative algorithms of gradient boosting regression, Bayesian optimization, and random forest. The optimization objective includes both product performance indicators and raw material cost factors.

7. The AI-driven process optimization method for polymer composite materials according to claim 1, characterized in that: The performance prediction task is implemented through a fusion model of backpropagation neural network and gradient boosting tree. This model takes the internal quality inspection issues in sub-table 2 and the customer feedback data in sub-table 3 as input features and outputs the product quality risk level, which is divided into three levels: high, medium and low.

8. The AI-driven process optimization method for polymer composite materials according to claim 1, characterized in that: The two-way linkage mechanism is specifically manifested as follows: when the performance prediction model determines that the expected performance of the user-designed formula does not meet the preset standard, the system automatically calls the formula recommendation algorithm to generate an optimized formula scheme and feeds the recommendation result back to the design end.

9. The AI-driven process optimization method for polymer composite materials according to claim 1, characterized in that: The incremental learning adopts a strategy that combines online gradient descent and elastic weight consolidation. Online gradient descent is used to quickly adapt to changes in data distribution, while elastic weight consolidation protects existing knowledge from being overwritten by new data by calculating the importance weights of parameters. The sliding window mechanism has a time window of 30 days, and the amount of new data is detected every 24 hours. When the amount of data reaches 80% of the window capacity, the incremental learning process is triggered.

10. The AI-driven process optimization method for polymer composite materials according to claim 1, characterized in that: The method is adaptable to the production processes of thermoplastic and thermosetting polymer materials, and dynamically adjusts the input feature dimensions and process constraints of the artificial intelligence model according to the specific type of injection molding, compression molding or pultrusion molding process. At the same time, it supports human-computer interaction through Chinese text commands, and the system automatically parses the command content and generates corresponding process parameter fluctuation curves and performance prediction comparison charts for visualization.

Images