Multi-source data driven intelligent optimization method for composite material pultrusion process

Abstract

The invention discloses a multi-source data-driven intelligent optimization method for a composite material pultrusion process. The method comprises the following steps: forming a multi-source process database by collecting process parameters and multi-working-condition simulation data obtained by simulation in real time; constructing a technological process LSTM prediction model based on a long short-term memory network, predicting quality, energy consumption and efficiency parameters of the product, and performing multi-target autonomous optimization to obtain a collaborative optimal solution set of the quality, energy consumption and efficiency of the product; according to the collaborative optimal solution set, a process scheme is generated, and an instruction is formed and issued; and according to the data of the staged multi-source process database, constructing a continuously learned data sample, updating and optimizing the process LSTM prediction model, and updating and optimizing the process decision-making scheme library. Compared with the prior art, the method has the advantages that autonomous decision, accurate control and continuous optimization of pultrusion process parameters are realized, so that the production efficiency is remarkably improved and the energy consumption is reduced on the premise of ensuring stable and reliable product quality.

Description

Technical Field

[0001] This invention belongs to the field of composite material pultrusion process, and in particular, it is a multi-source data-driven intelligent optimization method for composite material pultrusion process. Background Technology

[0002] like Figure 1 , Figure 2 The composite material pultrusion process shown is mainly used to produce products reinforced with glass fiber, carbon fiber, and other materials. It boasts advantages such as high production efficiency, stable product quality, and low cost, and has become one of the core processes for the industrial production of fiber-reinforced polymers (FRP). Currently, pultruded FRP profiles, due to their lightweight, high strength, and corrosion resistance, are widely used in aerospace, automotive, construction, electrical equipment, corrosion-resistant components, petrochemical, and electronics industries.

[0003] Currently, the pultrusion process for composite materials mainly employs two methods: manual parameter adjustment based on experience and local control with single-variable feedback. The former heavily relies on the operator's experience, manually setting process parameters such as mold temperature and traction speed, and using fixed thresholds or simple PID control for local adjustments. This approach suffers from problems such as parameter adjustment lag and inability to quantify the coupling relationship of multiple parameters, easily leading to uneven curing and fluctuations in yield. The latter achieves local closed-loop control through a single sensor (such as a temperature sensor), improving some process stability, but it fails to consider the nonlinear interference of multi-source data, and the degree of curing detection relies on offline sampling, making it difficult to provide real-time quality information.

[0004] Both of the above methods lack data modeling capabilities and dynamic optimization mechanisms. Static empirical models cannot adjust process operation parameters according to the differences in curing characteristics of different matrix resins such as epoxy resin and polyurethane, which can easily lead to shrinkage and the formation of cracks or pores. Control models based on fixed thresholds or preset mapping tables are too rigid and lack dynamic response to external disturbances such as ambient temperature and tension, resulting in insufficient resin wetting or curing mismatch. It is difficult to balance energy consumption and efficiency. The extensive control mode can easily lead to mold overheating or slow curing, making it difficult to achieve high performance, low energy consumption and high yield production. It is difficult to achieve synergistic optimization among performance, energy consumption and efficiency, resulting in insufficient product performance consistency, high residual stress and difficulty in meeting the requirements of modern high-precision pultrusion molding processes. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-source data-driven intelligent optimization method for composite material pultrusion processes, so as to achieve autonomous decision-making, precise control and continuous optimization of pultrusion process parameters, thereby significantly improving production efficiency and reducing energy consumption while ensuring stable and reliable product quality.

[0006] The technical solution to achieve the purpose of this invention is a multi-source data-driven intelligent optimization method for composite material pultrusion forming process, the specific steps of which are as follows:

[0007] S1. Construct a multi-source process database: Real-time acquisition of process parameters and preprocessing of the data; Numerical simulation of the processed data and process setting parameters as input to obtain multi-condition simulation data; The real-time acquired process parameters and multi-condition simulation data constitute a multi-source process database.

[0008] S2. Constructing an intelligent prediction model: Based on a long short-term memory network, a process LSTM prediction model is constructed, and data from a multi-source process database is input to predict the quality, energy consumption, and efficiency parameters of the product; at the same time, the predicted quality, energy consumption, and efficiency parameters of the product are autonomously optimized through multiple objectives to obtain the optimal solution set for the quality, energy consumption, and efficiency of the product.

[0009] S3. Process Execution: Based on the optimal solution set of product quality, energy consumption and efficiency, a process plan is generated and instructions are issued; the zoned temperature control mold is adjusted according to the zoned temperature parameters in the instructions, and the adaptive traction system is executed according to the traction speed in the instructions;

[0010] S4. Adaptive Learning: Based on data from a multi-source process database at different stages, construct data samples for continuous learning, update and optimize the LSTM prediction model for the process, and update and optimize the process scheme decision base.

[0011] The significant advantages of this invention compared to existing technologies are:

[0012] (1) By constructing a multi-source process database, we can break through the limitations of traditional control and realize full-process data-driven operation: By integrating physical sensors and numerical simulation, we can construct a multi-source process database, overcome the shortcomings of existing technologies such as single data source and inability to quantify the coupling relationship of multiple parameters, and lay a solid data foundation for intelligent optimization.

[0013] (2) By constructing an integrated intelligent closed loop of “prediction-optimization-execution”: The LSTM time series prediction model is innovatively combined with the “energy efficiency matching-multi-objective genetic” progressive optimization algorithm, realizing end-to-end automation from intelligent decision-making to precise execution, solving the fundamental problems of traditional methods relying on human experience, lagging regulation, and inability to coordinate and optimize multiple objectives.

[0014] (3) Through multi-objective optimization, the system can significantly reduce energy consumption per ton of product and improve production efficiency while ensuring high curing degree (>99%) and low defect rate (<2%) of the product. It effectively solves problems such as insufficient product performance consistency, high residual stress and fluctuation in yield.

[0015] (4) Through the adaptive learning mechanism, it can continuously learn from production data, update the prediction model and optimization strategy, and thus adapt to changes in raw materials, equipment aging and environmental disturbances in the long term. It has the robustness and long life cycle that traditional static models do not have. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the continuous pultrusion process.

[0017] Figure 2 Schematic diagram of a pultrusion heating die;

[0018] Figure 3 This is the technical approach of the present invention;

[0019] Figure 4 The cross-sectional shape and dimensions of the product;

[0020] Figure 5 The mesh generation results for the product cross-section;

[0021] Figure 6 Numerical simulation results for this invention: cross-sectional temperature and degree of curing (at 2.88 s, 35.72 cm from the inlet).

[0022] Figure 7 The numerical simulation results show the changes in temperature and degree of curing at the surface and center over time; where (a) represents the temperature change at the surface and center, and (b) represents the degree of curing change at the surface and center.

[0023] Figure 8 To improve the accuracy of model predictions;

[0024] Figure 9 A schematic diagram of the adaptive learning closed-loop optimization process.

[0025] Figure descriptions: 1. Fiber yarn frame, 2. Impregnation tank, 3. Preforming tank, 4. Molding mold and heating system, 5. Traction machine, 6. Resin bath, 1-1 Impregnating fiber, 1-2 Impregnating fiber liquid stage, 4-1 Heating plate, 1-3 Impregnating fiber gel stage, 4-2 Molding mold, 1-4 Impregnating fiber solidification stage. Detailed Implementation

[0026] The terminology used in this invention is for illustrative purposes only and is not intended to limit the invention. The following description, in conjunction with the appendix, further clarifies this concept. Figure 1-9 The following is a detailed description of some embodiments of the present invention.

[0027] A multi-source data-driven intelligent optimization method for composite material pultrusion processes is proposed. This method deploys a multi-source sensor network and combines it with numerical simulation to construct a multi-source database covering process parameters, environmental parameters, and online quality indicators. Based on this database, a long short-term memory network is used to build an intelligent prediction model that accurately maps the complex relationships between process parameters and product quality, energy consumption, and efficiency. Furthermore, a progressive optimization strategy consisting of an energy efficiency matching algorithm and a multi-objective genetic algorithm is employed to automatically find the optimal set of process parameters that satisfy multiple objectives. Finally, a closed-loop execution system accurately distributes the optimized parameters to the production line controller, driving the temperature control and traction units. An adaptive learning mechanism is introduced to continuously update the prediction model and optimization strategy, ultimately achieving autonomous optimization and continuous improvement of the pultrusion process in terms of high quality, low energy consumption, and high efficiency.

[0028] The specific steps are as follows:

[0029] S1. Constructing a Multi-Source Process Database: The multi-source process database includes real-time acquisition of process parameters and multi-condition simulation data during the pultrusion process of composite materials. The specific steps are as follows:

[0030] S1.1 Data Acquisition: Set up various sensors to collect process parameters in real time during the pultrusion process of composite materials. The process parameters mainly include zone temperature and traction speed.

[0031] Platinum resistance temperature sensors are arranged on the surface of each independent heating zone of the mold to collect the heating temperature along the line. Miniature or armored platinum resistance sensors are arranged inside the mold cavity to collect the resin temperature inside the mold. Platinum resistance temperature sensors are arranged in the resin storage tank, the outlet of the mixer, or the resin preheater to collect the resin temperature before it enters the mold. Platinum resistance temperature sensors are arranged in the workshop environment or cooling water system to collect the process environment temperature.

[0032] A miniature silicon piezoresistive pressure transmitter is embedded in the inner wall of the mold curing section, with its sensitive end face flush with the flow channel, for direct monitoring of the in-mold pressure during the resin curing process.

[0033] A laser velocity sensor is deployed on the surface of the product between the mold exit and the traction mechanism to measure the actual pultrusion speed of the product in real time without contact.

[0034] Vibratory viscometers are deployed in resin tanks or glue supply lines to monitor resin viscosity changes online and in real time.

[0035] Miniature interdigitated electrode dielectric sensors are deployed in the inlet gel zone, middle curing zone, and outlet of the pultrusion die to detect the degree of curing in real time.

[0036] Low-angle dark field lighting is deployed at the mold exit, along with a line scan camera with over 8 megapixels, to acquire real-time images of the products exiting the mold.

[0037] S1.2 Preprocessing of real-time acquired process parameter data:

[0038] S1.2.1 Cleaning: Taking ambient temperature pretreatment as an example:

[0039] The parameter fluctuation range is set according to the process flow; if it exceeds the range, it is discarded: if (T am (t) < T am_min ) or (T am (t) > T am_max If T is removed, then T is removed. am (t); the T am (t) represents the workshop ambient temperature measured at time t, where T is the temperature at time t. am_min and T am_max These represent the minimum and maximum possible workshop ambient temperatures, respectively.

[0040] Calculate the difference between adjacent data points; if the difference exceeds a reasonable range, discard the data: if |T am (t) - T am (t- t) |> If T), then remove T. am (t), the T is Within time step t am The maximum possible change.

[0041] S1.2.2 Denoising: A Butterworth low-pass filter is used to denoise and filter the data.

[0042] S1.2.3 The product images collected at the mold exit are analyzed using Halcon inspection software, a machine view library in an industrial computer with an embedded GPU accelerator, to obtain information on surface cracks and bubble defects.

[0043] S1.3. Using the processed data and process setting parameters as input, numerical simulation is performed to obtain multi-condition simulation data; the specific method is as follows:

[0044] S1.3.1 Geometric Modeling and Mesh Generation: (e.g.) Figure 4 As shown, the cross-sectional shape of the pultruded product is flat, 120 mm long, 5 mm high, with a corner radius of 1 mm and an edge angle of 85°. The total length of the pultrusion region is 105 cm, divided into three regions, each 35 cm long. A geometric model was created and meshed on the Ansys platform based on the product shape, as shown below. Figure 5 As shown.

[0045] S1.3.2 Solution Setup: PAM-RTM software was used to numerically simulate the pultrusion curing process based on curing reaction kinetics and heat transfer.

[0046] ;

[0047] In the formula, α represents the degree of curing, t represents time, and A represents the frequency factor. Let A represent the activation energy of the curing process, R represent the gas constant, T represent the temperature, and m and n both represent the reaction order. And A = 2.92 × 10⁻⁶. 8 , The parameters are: m=0.28, n=0.82. Epoxy resin (model SWANCOR 2208) and carbon fiber (model Formosa Plastics T70048K) were selected. The temperatures in the three zones were: Zone 1 140-160℃, Zone 2 160-180℃, and Zone 3 180-200℃, with the temperature in each zone increasing uniformly with distance. The pultrusion speed was a constant 0.5 m / min (i.e., 0.833 cm / s). The resin volume fraction in the composite material was 0.27%, and the fiber volume fraction was 0.73%. Detailed input parameters are shown in Table 1. Furthermore, for different products, the input data can include the processed data (pultruded product temperature, degree of cure) and process setting parameters (die wall temperature and pultrusion speed in different zones).

[0048] Table 1 Input Parameter List

[0049]

[0050] S1.3.3 Result Processing: Simulation of the pultrusion process yields multi-condition simulation data, which is then processed to achieve visualization, such as... Figure 6 ;

[0051] Obtain the changing trends of key parameters in the process, such as Figure 7 The temperature and curing degree of the surface and center were measured over time to obtain data on curing degree, generation efficiency and energy consumption. The simulation database is shown in Table 2.

[0052] Table 2 Multi-condition simulation database (partial case studies)

[0053]

[0054] S2. Construct an intelligent prediction model. Based on a Long Short-Term Memory (LSTM) network, construct a prediction model for the process and input data from a multi-source process database to predict the quality, energy consumption, and efficiency parameters of the finished product. Simultaneously, perform multi-objective autonomous optimization on the predicted quality, energy consumption, and efficiency parameters of the finished product to obtain a collaborative optimal solution set for the quality, energy consumption, and efficiency of the finished product. The specific method is as follows:

[0055] S2.1 Constructing a predictive model for the process based on a Long Short-Term Memory (LSTM) network:

[0056] To accurately capture the complex nonlinear temporal relationships between pultrusion process parameters (see Table 2) and product quality, production efficiency, and energy consumption, this invention employs a Long Short-Term Memory (LSTM) network as the core prediction model. Through its internal gating mechanism, LSTM can effectively learn and memorize long-term dynamic dependencies in the process, overcoming the shortcomings of traditional models in modeling multi-parameter coupling and time lag effects.

[0057] S2.1.1 Internal Structure and Mathematical Description of an LSTM Unit: An LSTM unit processes input data at each time step t, and its core calculation steps are as follows:

[0058] 1. Forgotten Gate f t : Determines how much information to discard from the previous cell state.

[0059]

[0060] 2. Input gate i t and candidate cell status : Determines how much new information will be stored in the cell state.

[0061]

[0062]

[0063] 3. Cell state renewal C t : Combine information from the forget gate and the input gate to update the cell state.

[0064]

[0065] 4. Output gate O t and h t The output of this time step is determined based on the updated cell state.

[0066]

[0067]

[0068] in, It is the Sigmoid activation function. This represents the final output vector generated by the LSTM unit at time t after completing all internal computations (through the forget gate, input gate, and output gate). This represents the observed parameter data at time t. It is the hyperbolic tangent activation function. W represents element-wise multiplication. * and b * These are the weight matrix and bias term for the corresponding gate, respectively.

[0069] S2.1.2 Input, output, and parameter settings of the LSTM prediction model for the process:

[0070] (1) Model input The input at each time step is a 7-dimensional vector, with the following parameters:

[0071] T1_set: Temperature setting (°C, e.g., 150) for the first heating zone of the mold (referring to the molding die in the background art).

[0072] T2_set: Temperature setting for the second heating zone of the mold (°C, e.g., 170)

[0073] T3_set: Temperature setting for the third heating zone of the mold (°C, e.g., 190)

[0074] pull_speed: Traction speed (m / min, e.g., 0.5)

[0075] resin_viscosity: The online monitored resin viscosity (mPa·s, e.g., 450).

[0076] fiber_volume_fraction: Fiber volume fraction (dimensionless, e.g., 0.65)

[0077] ambient_temp: Workshop ambient temperature (°C, e.g., 25)

[0078] (2) Model output y t The predicted output at each time step is a 5-dimensional vector representing key process quality and performance indicators.

[0079] Final_cure_degree: Predicted degree of curing at the product export point (%, target > 99%);

[0080] max_temperature_diff: Predicted maximum temperature difference (°C) across the product cross-section, target < 2°C;

[0081] surface_defect_prob: The predicted probability (%) of surface cracks / bubbles appearing (target < 2%);

[0082] energy_consumption_per_ton: Predicts energy consumption per ton of product (kJ / t, e.g., 95000);

[0083] production_efficiency: Predicted production efficiency (kg / min, e.g., 0.465);

[0084] (3) Network structure and hyperparameters: Input sequence length: The model uses process data from the past 60 time steps (i.e., [x]). t-59 , x t-58 , ..., x t The LSTM prediction model for the process is used as input to predict the output y at the current time step. t This is equivalent to about 5-15 minutes of historical data (adjusted according to the data acquisition frequency), which is sufficient to capture the dynamic process of the curing reaction.

[0085] (4) Network layers:

[0086] Input layer: Receives an input sequence of shape (60, 7).

[0087] LSTM layer 1: Contains 128 LSTM units, set return_sequences=True to output the hidden state of all time steps to the next layer.

[0088] LSTM layer 2: Contains 64 LSTM units, set return_sequences=False, only outputs the hidden state of the last time step to the next layer.

[0089] Fully connected layer: A Dense layer with 32 neurons, using the ReLU activation function.

[0090] Output layer: A Dense layer with 5 neurons, using linear activation functions, corresponding to 5 output targets.

[0091] (5) Training configuration:

[0092] Loss function: mean squared error.

[0093] Optimizer: Adam optimizer, with an initial learning rate set to 0.001.

[0094] Batch size: 32.

[0095] Training rounds: 200, with early stopping mechanism, automatically terminating training when the validation set loss no longer decreases within 15 consecutive rounds.

[0096] In actual deployment, the system continuously retrieves the latest time-series process data from the multi-source process database built by S1. After preprocessing, this data is input into a pre-trained LSTM prediction model for the process. The model, through its learned nonlinear mapping relationships, predicts in real time the quality (degree of curing, defects), energy consumption, and efficiency of the product to be produced under the current parameter settings. Figure 8 As shown, the LSTM prediction for this process has a prediction error of less than 2% for curing time, energy consumption and efficiency on the test set.

[0097] S2.2 Multi-objective autonomous optimization of the predicted product quality, energy consumption and efficiency parameters: The global process optimization is decomposed into two stages: "energy efficiency benchmark optimization" and "multi-objective fine trade-off", realizing progressive optimization from energy efficiency baseline constraints to optimal comprehensive performance.

[0098] S2.2.1 Based on the energy efficiency matching algorithm, an energy efficiency query model is constructed to optimize the energy efficiency benchmark of the process parameter range in the composite material pultrusion process and obtain a high-energy-efficiency initial population.

[0099] Based on the predicted quality, energy consumption, and efficiency parameters of the product, and under the premise of meeting the basic quality requirements of the product, the optimal range of process parameters for energy efficiency is quickly identified, providing a high-quality initial solution set for subsequent fine optimization and avoiding blind searching in a vast and inefficient parameter space by multi-objective algorithms. Based on a multi-condition database of real-time detection and simulation, a rapid energy efficiency query model is constructed using key process parameters such as zone temperature, traction speed, resin type, and fiber volume fraction as input. The process of constructing the rapid energy efficiency query model based on the energy efficiency matching algorithm is as follows:

[0100] 3. Data-driven: Extract stable production data and corresponding energy consumption per ton of product from the multi-source process database built by S1 for different material systems (resin, fiber) and traction speeds.

[0101] 4. Benchmark Establishment: Through K-means cluster analysis and regression fitting, a benchmark energy consumption value E is established for each typical material-velocity combination. B and its corresponding core temperature range [T min ,T max ].

[0102] 5. Online Matching: Once the production task is determined (i.e., the materials and target speed are known), the system invokes the energy efficiency matching algorithm. The algorithm first matches the corresponding baseline energy consumption E based on the current materials and target speed. B and core temperature range [T min ,Tmax ].

[0103] 6. Generate initial solutions: Within the core temperature range, using methods such as Latin hypercube sampling, generate a set (50) of solutions that satisfy... process parameter combination {x e},in For the allowable relative error, 5%, {x e} is the initial population with high energy efficiency.

[0104] S2.2.2 Starting with the high-energy-efficiency initial population output by the energy efficiency matching algorithm, a synergistic optimal solution set of quality, energy consumption and efficiency is achieved based on the multi-objective genetic algorithm.

[0105] Starting with a high-energy-efficiency initial population output by the energy efficiency matching algorithm, a non-dominated sorting genetic algorithm with an elite retention strategy is used for fine-grained search to achieve the optimal balance between quality, energy consumption, and efficiency.

[0106] The optimization problem format is as follows:

[0107] (1) Define decision variables: high-energy-efficiency initial population {x} e}:[T1_set, T2_set, T3_set, pull_speed];

[0108] (2) Construct the objective function {y} e}: Find the minimum value of the total mass loss, energy consumption per ton of product, and the reciprocal of production efficiency.

[0109] a. Overall quality loss:

[0110] ;

[0111] in, To predict the degree of cure, and Corresponding to the defect probability values Maximum temperature difference of the cross section after normalization The weighting coefficients.

[0112] b. Energy consumption per ton of product

[0113] c. Reciprocal of production efficiency .

[0114] (3) Constraints: Curing degree > 99%, defect probability < 2%, maximum temperature difference < 2℃, and physical upper and lower limits of each parameter, such as the lower and upper limits of the mold heating zone temperature of 120℃ and 250℃ respectively.

[0115] (4) Calculation process:

[0116] Initialization: The output of the energy efficiency matching algorithm {x} e} serves as the initial population for a non-dominated sorting genetic algorithm with an elite retention strategy.

[0117] Evaluation: For each individual x in the population e It calls the LSTM process prediction of S2.1 to quickly predict all objective function values ​​and constraint violations.

[0118] Evolution: The algorithm performs selection, crossover, and mutation operations. Individuals are stratified according to the objective function value, with higher-ranking solutions being retained first. Within the same non-dominated layer, individuals located in sparse regions are preferentially selected to maintain population diversity.

[0119] Convergence and Output: After iterating to the preset algebra or reaching convergence, output a set of co-optimal solutions in terms of quality, energy consumption, and efficiency.

[0120] S3. Execute the process plan: Based on the quality of the product. Energy consumption With efficiency The collaborative optimal solution set x d (T1_set, T2_set, T3_set, pull_speed) generates a process plan and issues instructions; the zoned temperature control mold adjusts according to the zoned temperature parameters (T1_set, T2_set, T3_set) in the instructions, and the adaptive traction system executes according to the traction speed pull_speed in the instructions;

[0121] This invention achieves end-to-end automation from intelligent decision-making to device execution. For example... Figure 3 As shown, after the optimal solution set is output in S2.2, the composite material pultrusion process system completes the automatic deployment and closed-loop control of process parameters according to the following process:

[0122] S3.1 Automatic decision-making and instruction issuance for the optimal solution:

[0123] 1. Strategy matching:

[0124] a. "Quality First" Strategy: Automatically select the solution from the optimal solution set that has the highest predicted degree of curing at the product outlet, the smallest predicted maximum temperature difference at the product cross-section, and the lowest predicted probability of surface cracks / bubbles.

[0125] b. "Energy Efficiency First" Strategy: Automatically selects the option with the lowest predicted energy consumption per ton of product.

[0126] c. "Balanced production" strategy: Select the option with the highest weighted composite score among the three objectives.

[0127] 2. Instruction Generation and Verification: Select the chosen scheme x dThe optimal process parameters (zone temperatures T1_set, T2_set, T3_set and pull_speed) are encapsulated into a standardized control command package. Before being issued, boundary checks are performed to ensure that the command values ​​are within the physical upper and lower limits of each parameter defined in S2.2, preventing illegal commands from damaging the equipment.

[0128] 3. Command issuance: Through the unified communication architecture of the open platform, control command packages are issued to the programmable logic controllers on the production line in real time.

[0129] S3.2, Precise temperature control of mold zones:

[0130] The partitioned temperature control mold is the execution terminal for achieving precise temperature control. The specific steps are as follows:

[0131] S3.2.1 Instruction Reception: The intelligent temperature controllers of each zone of the mold receive new temperature setpoints from the programmable logic controller (e.g., Zone 1 T1_set→150℃, Zone 2 T2_set→170℃, Zone 3 T3_set→190℃).

[0132] S3.2.2, Precise Control:

[0133] Heating: The temperature controller drives a solid-state relay to control the power of the resistance heating element embedded in the mold, enabling rapid heating.

[0134] Cooling: When the temperature exceeds the set value due to the exothermic reaction, the temperature controller will open the solenoid valve of the zone and start the compressed air cooling duct to force-cool the product.

[0135] S3.2.3 Closed-loop feedback: The platinum resistance temperature sensor built into each zone detects the actual temperature of the mold in real time and feeds it back to the temperature controller to form a high-precision temperature PID closed-loop control loop, ensuring that the mold temperature is stable within ±1℃ of the set value.

[0136] S3.3 Dynamic Response of Adaptive Traction Machine: The adaptive traction machine is responsible for accurately executing the pull_speed traction command and ensuring product quality.

[0137] S3.3.1 Speed ​​Execution: The programmable logic controller converts the received traction speed setting (e.g., pull_speed=0.5 m / min) into a frequency command for the servo driver. Upon receiving the command, the servo motor responds rapidly within 0.5 seconds, driving the traction chain to clamp the product and perform uniform pultrusion at the new set speed.

[0138] S3.3.2 Tension Adaptive Mechanism: Tension sensors deployed on the traction mechanism monitor the tension on the product in real time. When an abnormal tension is detected (such as increased tension due to changes in friction, posing a risk of damaging the product), the traction machine does not change the optimized main line speed. Instead, it fine-tunes the torque output of the servo motor through the programmable logic controller or adjusts the air pressure to dynamically adjust the clamping pressure of the traction track, thereby ensuring the smoothness and safety of the traction process while maintaining the optimal process speed.

[0139] S3.4 Real-time data acquisition enables closed-loop monitoring and real-time feedback to achieve adaptive and autonomous optimization of the LSTM prediction model for the process: The end of this optimization is the starting point for the next optimization, forming a continuously improving intelligent closed loop.

[0140] S3.4.1 Effect Monitoring: The system continuously collects data from the sensor network (temperature, viscosity, dielectric, vision, etc.) deployed in S1.1 to monitor the actual production status under the new parameters.

[0141] S3.4.2 Performance Evaluation: Compare actual production data (such as measured degree of curing and energy consumption) with the predicted values ​​of the LSTM prediction model of the process to evaluate the optimization effect.

[0142] S3.4.3 Adaptive Trigger: If the following situations occur, the system will automatically trigger the autonomous optimization process of S2.2 and start a new round of optimization: (1) The actual index continues to deviate from the predicted value; (2) The raw material batch is changed (the resin viscosity baseline changes abruptly); (3) The preset periodic optimization time point is reached.

[0143] S4. Based on the phased data from the multi-source process database, construct continuous learning data samples, update and optimize the LSTM prediction model for the process, and update and optimize the process scheme decision base. To achieve the system's self-evolution, this invention designs a three-in-one adaptive learning closed-loop optimization method: continuous learning data samples, periodic incremental learning and retraining of the model, and policy base updates. Figure 9 This mechanism ensures that the system can learn from each production practice and continuously improve its intelligence.

[0144] S4.1 Constructing data samples for continuous learning:

[0145] S4.1.1, Data Lake Construction: This involves combining the multi-source data (process parameters, sensor readings, visual defect images, etc.) collected in real-time and preprocessed in S1 with the co-optimal solution set x from each optimization decision. d (T1_set, T2_set, T3_set, pull_speed) (the final process parameters) and their corresponding objective function y d (Final quality inspection results, total energy consumption, i.e., quality) Energy consumption The actual production results are linked together, timestamped, and stored in a unified industrial data lake.

[0146] S4.1.2, Effect-Label Association: Generate a unique "production batch" for each batch (or each time period) of products, and assign the collaborative optimal solution set x used by that batch. d (T1_set, T2_set, T3_set, pull_speed) and the objective function y d (Actual quality of final inspection) Energy consumption with energy efficiency indicators Precise binding is performed to form high-quality sample pairs. <x d , y d >

[0147] S4.2 Online Update of the Process LSTM Prediction Model: This system adopts a strategy that combines "periodic incremental learning" and "performance-triggered retraining" to continuously improve the process LSTM prediction model.

[0148] S4.2.1 Incremental Learning: The system automatically initiates model fine-tuning every N new valid production batches of data (e.g., N=100), or at fixed intervals (e.g., once a month). The newly accumulated data... <x e , y e The sample data was mixed with some historical data, and the existing S2.1 process LSTM prediction model was retrained a few times with a low learning rate.

[0149] S4.2.2 Performance-Triggered Retraining: The system continuously monitors the predictive performance of the LSTM prediction model for the process. When it finds that the mean squared error of the model's prediction on the latest data is consistently higher than the threshold, or when there is a systematic drift in the deviation between the predicted and actual values, it automatically triggers a deep retraining of the model. A complete and thorough model retraining is performed using all available historical data in the data lake to regenerate the prediction model.

[0150] S4.3 Update the process decision database;

[0151] S4.3.1 Dynamic Update of Energy Efficiency Benchmark Library: The benchmark energy consumption value E on which the energy efficiency matching algorithm in S2.2 depends. B The database is not static. The system periodically recalculates and updates the baseline energy consumption values ​​for each "material-rate" combination based on the latest high-quality and high-yield data, ensuring that they always match the current production system capacity.

[0152] S4.3.2, Accumulation of Optimal Solution Set Experience: Each successfully optimized optimal solution set is saved and its corresponding production conditions (such as resin batch, environmental season) are marked. When encountering similar production conditions, the system can prioritize matching and initialization from the historical solution set, thereby significantly accelerating the optimization convergence process. This is equivalent to the system accumulating a "optimal process case library".

[0153] S4.3.3 Adaptive Adjustment of Optimized Algorithm Parameters: The system records the convergence of the algorithm under different production conditions. Based on this data, a meta-model can be built for dynamically recommending algorithm parameters (such as population size and crossover / mutation probability) to maintain efficient search performance in different scenarios.

[0154] This invention discloses a multi-source data-driven intelligent optimization system for composite material pultrusion processes, including...

[0155] Data Acquisition and Processing Module: This module is used to build a multi-source process database. Through a multi-source sensor network deployed on the pultrusion production line, it collects real-time data on mold temperature, in-mold pressure, traction speed, resin viscosity, dielectric curing degree, and product surface images. It also performs data cleaning, noise reduction, and defect analysis. Simultaneously, it utilizes numerical simulation technology to build an extended database covering multiple operating conditions, providing a data foundation for intelligent modeling.

[0156] Intelligent Prediction Module: Used to build and run process indicator prediction models. This module uses a long short-term memory network as its core, receives time-series data from the data acquisition and processing module, and learns the complex nonlinear mapping relationship between process parameters and the final curing degree of the product, cross-sectional temperature difference, defect probability, energy consumption per ton of product, and production efficiency, so as to achieve advanced and high-precision prediction of various key indicators.

[0157] Multi-objective parameter optimization module: This module is used to achieve autonomous and collaborative optimization of process parameters. Based on the output of the intelligent prediction module, it adopts a progressive optimization architecture consisting of an energy efficiency matching algorithm and a multi-objective genetic algorithm. The energy efficiency matching algorithm first quickly locks down the range of high-energy-efficiency process parameters based on materials and speed. Then, the multi-objective genetic algorithm performs a fine search within this range to find the optimal solution set in terms of quality, energy consumption, and efficiency objectives.

[0158] Intelligent Execution and Control Module: This module accurately translates optimization decisions into equipment actions. Based on the real-time production strategy, it automatically selects the optimal solution from the solution set output by the multi-objective parameter optimization module. It then transmits the process parameters to the production line's programmable logic controller (PLC) via industrial communication protocols, driving the zoned temperature-controlled molds and adaptive traction system to execute, forming a high-precision closed-loop control.

[0159] Adaptive Learning and Optimization Module: This module enables continuous system evolution. It constructs a continuous learning data pipeline to fine-tune and retrain the LSTM process predictions in the intelligent prediction module online, and dynamically updates the energy efficiency benchmark library and optimal process case library in the multi-objective parameter optimization module, giving the system the ability to continuously optimize in response to changes in production conditions.

[0160] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A multi-source data driven intelligent optimization method for composite pultrusion process, characterized in that, The specific steps are: S1, constructing a multi-source process database: real-time acquisition of process parameters, preprocessing of data; the processed data and process setting parameters are used as input for numerical simulation to obtain multi-condition simulation data; real-time acquisition of process parameters and multi-condition simulation data form a multi-source process database; S2, constructing an intelligent prediction model: based on a long short-term memory network, an LSTM model is constructed to predict the process, and the data of the multi-source process database is input to predict the quality, energy consumption and efficiency of the product; at the same time, the predicted quality, energy consumption and efficiency of the product are multi-objective self-optimized to obtain the collaborative optimal solution set of the quality, energy consumption and efficiency of the product; S3, executing the process scheme: according to the collaborative optimal solution set of the quality, energy consumption and efficiency of the product, a process scheme is generated and a command is issued; The partition temperature control mold adjusts according to the partition temperature parameters in the command, and the self-adaptive traction system executes according to the traction speed in the command; S4, according to the data of the phased multi-source process database, construct a continuous learning data sample, update the LSTM prediction model of the optimized process, and update the process decision scheme library.

2. The multi-source data-driven intelligent optimization method for composite pultrusion process according to claim 1, wherein, The specific steps of numerical simulation in S1 are: S1.3.1, geometric modeling and meshing; S1.3.2, solution setting: using PAM-RTM software, based on solidification reaction kinetics and heat transfer, numerical simulation of the pultrusion solidification process is carried out: ； wherein a represents the degree of cure, t represents time, A represents a frequency factor, represents the activation energy of the curing process, R represents the gas constant, T represents the temperature, and m and n both represent the reaction order; S1.3.3, obtain the simulation results of the pultrusion forming process, establish a multi-condition simulation database, and perform visual processing.

3. The multi-source data-driven intelligent optimization method for composite pultrusion process according to claim 1, wherein, In S2, the input, output and parameter setting of the process LSTM prediction model are as follows: (1) Model input : The input of each time step is a 7-dimensional vector, with the following parameters: T1_set: set temperature of the first heating zone of the mold; T2_set: set temperature of the second heating zone of the mold; T3_set: set temperature of the third heating zone of the mold; pull_speed: traction speed; resin_viscosity: online monitoring of resin viscosity; fiber_volume_fraction: fiber volume fraction; ambient_temp: workshop environment temperature; (2) Model output y t : The predicted output at each time step is a 5-dimensional vector representing key process quality and performance metrics: Final_cure_degree: predicted degree of cure at the product outlet; max_temperature_diff: predicted maximum temperature difference of the product cross section; surface_defect_prob: predicted probability of surface cracks / bubbles; energy_consumption_per_ton: predicted energy consumption per ton of product; production_efficiency: predicted production efficiency; (3) Network architecture and hyperparameters: input sequence length: the model uses past time steps of process data to predict the output y at the current time step t ; sufficient to capture the dynamic process of the curing reaction; (4) Network hierarchy: Input layer: received input sequence; LSTM layer 1: contains 128 LSTM units, set return_sequences=True, output the hidden state of all time steps to the next layer; LSTM layer 2: contains 64 LSTM units, set return_sequences=False, only output the hidden state of the last time step to the next layer; Fully connected layer: a Dense layer with 32 neurons using ReLU activation function; Output layer: a Dense layer with 5 neurons using linear activation function, corresponding to 5 output targets respectively.

4. The multi-source data-driven intelligent optimization method for composite pultrusion process according to claim 1, wherein, In S2, the specific steps of multi-objective autonomous optimization of the predicted product quality, energy consumption and efficiency parameters are as follows: S2.2.1, based on the energy efficiency matching algorithm, an efficiency query model is constructed to optimize the energy efficiency benchmark of the process parameter interval in the composite material pultrusion forming process, and a high energy efficiency initial population is obtained; S2.2.2, taking the high energy efficiency initial population output by the energy efficiency matching algorithm as the starting point, based on the multi-objective genetic algorithm, the collaborative optimization of quality, energy consumption and efficiency is realized.

5. The multi-source data-driven intelligent optimization method for composite pultrusion process according to claim 4, wherein, In S2.2.1, based on the energy efficiency matching algorithm, the construction of the efficiency query model is as follows: Data-driven: from the multi-source process database, the stable production data and corresponding ton product energy consumption under different material systems and pulling speeds are extracted; Benchmarking: Establish benchmark energy consumption value E for each material-velocity combination by K-means clustering analysis and regression fitting B and the corresponding core temperature interval [T min , T max ], and establish an energy efficiency benchmark library; Online matching: when the production task is determined, i.e. the material and target speed are known, the system calls the energy efficiency matching algorithm to match the corresponding reference energy consumption E according to the current material and target speed B and the core temperature interval [T min ,T max ] Generating initial solutions: in the core temperature interval, a set of high-energy-efficient initial populations {x e} that satisfy are generated by using the Latin hypercube sampling method, where, is the predicted energy consumption, is the allowed relative error.

6. The multi-source data-driven intelligent optimization method for composite pultrusion process according to claim 5, wherein, In S2.2.2, (1) Define decision variables: high energy efficiency initial group species {x e}: [T1_set, T2_set, T3_set, pull_speed]; In the formula, T1_set is the set temperature of the first heating zone of the mold; T2_set is the set temperature of the second heating zone of the mold; T3_set is the set temperature of the third heating zone of the mold; pull_speed is the pulling speed; (2) Construct the objective function {y e}: Solve the minimum value of the comprehensive quality loss, energy consumption per ton of product, and the reciprocal of production efficiency; wherein the overall quality loss : ； wherein x is an input parameter corresponding to the predicted case; for predicting the curing degree, for the defect probability value a weight coefficient of the defect probability value, a weight coefficient of the normalized maximum temperature difference of the cross section; and a weight coefficient of the normalized maximum temperature difference of the cross section. Objective function {y e The constraint conditions of the objective function {y The output parameters corresponding to the predicted case x, i.e., the solidification degree of the product at the mold outlet > 99%, the defect probability of the product < 2%, the maximum temperature difference of the product cross section < 2℃, and the physical upper and lower limits of each process parameter.

7. The multi-source data-driven intelligent optimization method for composite pultrusion process according to claim 1, wherein, The specific steps of step 3 are as follows: S3.1 Optimal solution automatic decision and instruction issuing; optimal solution selection: select the solution with the highest predicted solidification degree at the product outlet, the smallest predicted maximum temperature difference of the product cross section, the lowest predicted probability of surface cracks / bubbles, the lowest predicted ton product energy consumption, and the highest weighted comprehensive score of the three targets from the optimal solution set; S3.2, precise partition temperature control of the mold; according to the temperature instruction, the partition temperature control is carried out, and the real-time detection and feedback of the partition temperature are used to ensure the stability of the mold temperature; S3.3, dynamic response of the adaptive pulling machine: the adaptive pulling machine executes the composite material pultrusion forming process according to the speed instruction and ensures the product quality; S3.4, real-time collection of effect monitoring data and performance evaluation data in the composite material pultrusion forming process, realization of closed-loop monitoring and performance evaluation, and real-time feedback to realize adaptive autonomous optimization of the process LSTM prediction model.

8. The multi-source data-driven intelligent optimization method for composite pultrusion process according to claim 6, wherein, In S4, the strategy of periodic incremental learning combined with performance triggered retraining is adopted to update and improve the process LSTM prediction model, and the specific steps are as follows: (1) Incremental learning: the system automatically starts a model fine-tuning once every N new valid production batch data is accumulated, or every fixed time; mix the newly accumulated <x e , y e > samples with the historical data in the data samples of continuous learning, and retrain the existing S2 process LSTM prediction for a few rounds; (2) Model pruning: the system automatically prunes the model once every N new valid production batch data is accumulated, or every fixed time; (2) Performance triggered retraining: the system continuously monitors the prediction performance of the process LSTM prediction, and when it is found that the prediction mean square error of the model on the latest data is continuously higher than the threshold value, or the deviation between the predicted value and the actual value appears systematic drift, the model is automatically triggered for deep retraining; using all available data in the continuous learning data sample, a complete and sufficient model retraining is carried out to regenerate the process LSTM prediction model.

9. The multi-source data-driven intelligent optimization method for composite pultrusion process according to claim 8, wherein, In S4, the specific steps of updating the process decision case library are as follows: (1) Dynamic updating of the energy efficiency benchmark library: periodically, according to the latest data, the benchmark energy consumption value E under each material-speed combination is recalculated and updated B so that the benchmark energy consumption value always meets the current production system capacity; (2) Construction of solution set library of collaborative optimal solution set: each time the collaborative optimal solution set is successfully optimized, it will be saved and marked with the corresponding production conditions, and when similar production conditions are encountered, the solution set library of the historical collaborative optimal solution set is matched and initialized in priority, thereby accelerating the optimization convergence process; (3) Adaptive adjustment of algorithm parameters: the system will record the convergence of the algorithm under different production conditions; based on these data, a meta-model is established to recommend algorithm parameter settings that adapt to the new prediction model in a timely manner, so as to maintain efficient updating of the prediction model in different scenarios.

10. A multi-source data-driven intelligent optimization system for composite material pultrusion processes, characterized in that, Comprise: Data acquisition and processing module: for building a multi-source process database; Intelligent prediction module: for building and running a process LSTM prediction model; Multi-objective parameter optimization module: for realizing self-cooperative optimization of process parameters: Intelligent execution and control module: for accurately converting optimization decisions into equipment execution actions.

Images