Preparation method of high-toughness glass fiber reinforced PA material

Abstract

The application provides a preparation method of high-toughness glass fiber reinforced PA material, and belongs to the technical field of the preparation of high-toughness glass fiber reinforced PA material.The application determines coupling agent coating process parameters by drying and pretreating polyamide chips and adopting a Fick diffusion and Arrhenius reaction kinetics coupling back calculation method, puts the surface-treated glass fiber and the polyamide chips into a double-screw extruder with a fiber length protection type screw configuration, synchronously starts an online near-infrared Raman combined spectrum system, and realizes real-time output of optimal screw parameters by a multi-scale space-time graph convolution and physical embedding fusion prediction model, predicts a fiber three-dimensional orientation tensor, and adjusts an injection molding process, and finally, returns detection data to the fusion prediction model for incremental training, so that the technical problem of unstable material toughness caused by uncontrollable glass fiber length distribution in the preparation process of glass fiber reinforced polyamide material is solved.

Description

Technical Field

[0001] This invention belongs to the field of preparation technology of high-toughness glass fiber reinforced PA materials, and specifically, it relates to a method for preparing high-toughness glass fiber reinforced PA materials. Background Technology

[0002] Glass fiber reinforced polyamide materials are widely used in automotive structural components, electronic and electrical housings, and engineering machinery parts due to their high strength and lightweight properties. In existing technologies, the composite of glass fiber and polyamide matrix mainly relies on twin-screw extrusion processes. Traditional processes rely on manual experience to set screw speed, temperature at each stage, and coupling agent dosage, and use offline detection methods for post-processing evaluation of fiber length distribution in the final product. Some improved solutions introduce a single online sensor to monitor the extruded melt state, or use simplified statistical regression models to make limited adjustments to process parameters.

[0003] However, due to the strong coupling of material transfer relationships between the functional sections of the screw, a single sensor signal cannot fully characterize the collaborative dynamics between multiple sections. The simplified regression model also lacks physical constraints on viscosity evolution and fiber breakage mechanisms, resulting in a lack of mechanistic basis for process adjustment and insufficient accuracy in predicting and controlling fiber length distribution.

[0004] In current twin-screw extrusion processes, the lack of predictive models that can integrate multi-source temporal signals and embed physical mechanisms means that process parameter optimization remains at the level of empirical trial and error, failing to achieve quantitative prediction and closed-loop control of glass fiber length distribution. In other words, existing technologies suffer from the technical problem of uncontrollable glass fiber length distribution during the preparation of glass fiber reinforced polyamide materials, leading to unstable material toughness. Summary of the Invention

[0005] In view of this, the present invention provides a method for preparing high-toughness glass fiber reinforced PA material, which can solve the technical problem in the prior art where the uncontrollable distribution of glass fiber length during the preparation of glass fiber reinforced polyamide material leads to unstable material toughness.

[0006] This invention is achieved as follows: This invention provides a method for preparing a high-toughness glass fiber reinforced PA material, comprising the following steps:

[0007] The polyamide chips were dried, and the glass fibers were surface-treated with silane coupling agent. The coating temperature curve and dosage of the coupling agent were determined by the inverse algorithm of interfacial coupling reaction kinetics coupled with Fick diffusion and Arrhenius reaction rate equation.

[0008] Surface-treated glass fibers and dried polyamide chips are fed into a twin-screw extruder with a fiber length protection screw configuration. Simultaneously, an online near-infrared pull-and-trade spectroscopy system is started to collect melt data, and the optimal screw parameters are output in real time by a multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model.

[0009] Based on the glass fiber length distribution index and melt degradation index output by the online near-infrared Raman spectroscopy system, the screw speed, temperature of each section and residence time are adjusted in a multivariate coordinated manner using a model predictive control algorithm;

[0010] After the extruded melt passes through the die, the Kalman filter soft measurement module is used to estimate the melt molecular weight in real time. When the estimated melt molecular weight is lower than the set threshold, vacuum exhaust treatment is added and the drying time is increased.

[0011] The composite granules obtained by extrusion granulation were injection molded to prepare standard specimens. The Reduced StrainClosure and machine learning correction models were used to predict the three-dimensional orientation tensor of glass fibers in the full flow field of injection molding. The holding pressure and injection rate were adjusted according to the prediction results.

[0012] Impact strength and elongation at break of the spline are tested. If the test value is lower than the target threshold, the test data is sent back to the multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model for incremental training, and then returned to the aforementioned extrusion step to readjust the process parameters.

[0013] Specifically, the interface coupling reaction kinetics back-calculation algorithm that couples Fick diffusion and Arrhenius reaction rate equations specifically models the adsorption process of silane coupling agents on the glass fiber surface as a set of coupled partial differential equations of two-dimensional Fick diffusion equations and Arrhenius reaction rate equations. Using measured adsorption data as constraints, the diffusion coefficient, activation energy, and pre-exponential factor are back-calculated using the quasi-Newton method.

[0014] The convergence criterion for the inverse estimation is that the Frobenius norm of the parameter change between two adjacent iterations is lower than a convergence threshold, and the reference value for the convergence threshold is [value missing]. .

[0015] The measured adsorption data were obtained by taking samples at multiple coating time points under multiple temperature gradients. The number and range of each temperature gradient and coating time point constituted the experimental design matrix. The mean of the parameter when the ratio of the standard deviation to the mean of the parameter estimation result was lower than the stability threshold was used as the reference value. The reference value of the stability threshold was 5%.

[0016] The multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model uses a graph structure to represent each functional segment of the twin-screw extruder. The bottom layer is a temporal convolutional network module, the middle layer is a GraphSAGE variant graph convolutional module, and the top layer is a physical constraint layer. The three layers are fused using residual connections.

[0017] The receptive field width of the temporal convolutional network module is dynamically determined based on the screw rotation speed parameter. When the screw rotation speed is higher than the rotation speed threshold, the receptive field width takes the first width value, and when it is lower than or equal to the rotation speed threshold, the receptive field width takes the second width value. The reference value of the rotation speed threshold is 150 r / min, the reference value of the first width value is 32 time steps, and the reference value of the second width value is 16 time steps.

[0018] The GraphSAGE variant graph convolution module employs a message-passing iterative mechanism, using the hidden state vectors of each node... Norm change is used as a convergence criterion; when all nodes... Iteration stops when all changes are below the iteration convergence threshold. The reference value for the iteration convergence threshold is [value missing]. The reference value for the maximum number of iterations is 6.

[0019] The physical constraint layer embeds the analytical expressions of the Carreau viscosity equation and the Weibull fiber fracture probability model into the network in a differentiable form. The output of the physical layer serves as the residual basis, and the output of the data-driven layer serves as the superposition of correction terms.

[0020] The multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model supports condition generation functionality. Given the target glass fiber length distribution, it obtains parameter suggestions for the optimal screw speed and temperature curve by backpropagating to the input layer.

[0021] The Kalman filter soft measurement module uses online near-infrared spectral signals as observations and the polyamide hydrolysis degradation kinetic equation as the equation of state to recursively estimate the molecular weight of the melt. The molecular weight threshold for initiating vacuum degassing is determined by comparing offline end-group titration data with the estimated value.

[0022] The glass fiber length distribution index is the ratio of number-average fiber length to weight-average fiber length, and its value ranges from 0 to 1. The closer the value is to 1, the more concentrated the distribution. Its threshold is obtained by statistical analysis of multiple sets of offline fiber length measurement data.

[0023] The improved Reduced Strain Closure and machine learning correction model introduces the Reduced Strain Closure closure approximation on the basis of the Folgar-Tucker fiber orientation constitutive equation, uses a multilayer perceptron to perform data-driven correction of the fiber interaction coefficients, and uses a GPU parallel finite element solver to solve the three-dimensional tensor field of the entire flow field.

[0024] The fiber length protection screw configuration refers to a screw arrangement in the design of a twin-screw extruder that reduces the number of meshing blocks and extends the length of the conveying section to reduce the mechanical breakage of glass fibers in the high-shear zone.

[0025] The online near-infrared Raman combined spectroscopy system simultaneously installs a near-infrared spectroscopy probe and a Raman spectroscopy probe at the extruder head. Through a pre-established partial least squares regression or support vector regression correction model, the spectral signal is converted into the glass fiber length distribution index and melt degradation index in real time.

[0026] The drying temperature range for polyamide chips is 80–120℃, the drying time range is 4–8 h, and the moisture content control threshold is below 100 ppm; the glass fiber content in the mass ratio of glass fiber to polyamide chips ranges from 30% to 40%; the extrusion temperature range is 240–285℃, and the screw speed range is 250–280 r / min.

[0027] This invention constructs a multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model, which integrates multi-source heterogeneous signals such as temperature time series, torque time series, and online near-infrared spectrum of each functional segment of the screw into a graph structure for collaborative modeling. The bottom-layer temporal convolutional network extracts local dynamic features, the middle-layer graph convolutional module transmits material state information between segments, and the top-layer physical constraint layer uses the Carreau viscosity equation and Weibull fiber fracture model as prior constraints. The three-layer structure forms an organic whole through residual connection, thereby realizing the quantitative prediction of glass fiber length distribution.

[0028] Traditional processes rely on single sensor signals and empirical regression models, which cannot describe the material evolution laws of multi-segment coupling. The fusion prediction model of this invention ensures that the material state of each node in the graph reaches self-consistent convergence through a message passing iteration mechanism, so that the prediction results reflect the global material evolution law of the screw. At the same time, the introduction of a physical constraint layer ensures that the model can maintain physical consistency even when training data is insufficient, avoiding the problem of pure data-driven models failing in the extrapolation interval.

[0029] In summary, the present invention solves the technical problem mentioned in the background art of unstable material toughness caused by uncontrollable glass fiber length distribution during the preparation of glass fiber reinforced polyamide materials. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 The diagram shows the convergence process of parameters derived from the inverse dynamics of Fick diffusion and Arrhenius coupling.

[0032] Figure 3 The graph shows the change of glass fiber length distribution index with extrusion time as output by the multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0034] like Figure 1 The diagram shown is a flowchart of a method for preparing a high-toughness glass fiber reinforced PA material provided by the present invention. This method includes the following steps:

[0035] S01. Dry the polyamide chips at 80-120℃ for 4-8 h to reduce the moisture content to below 100 ppm. At the same time, treat the glass fiber with silane coupling agent and determine the coating temperature curve and dosage of coupling agent using the Fick diffusion-Arrhenius coupling interface coupling reaction kinetics back-calculation algorithm.

[0036] S02. Surface-treated glass fibers and dried polyamide chips are fed into a twin-screw extruder with fiber length protection screw configuration at a mass ratio. Simultaneously, an online near-infrared-La Mans combined spectroscopy system is started to collect melt data, and the optimal screw parameters are output in real time by a multi-scale spatiotemporal graph convolution-physical embedding fusion prediction model.

[0037] S03. Based on the glass fiber length distribution index and melt degradation index output by the online near-infrared-Laman combined spectroscopy system, the screw speed, temperature of each section and residence time are adjusted in a multivariate coordinated manner using a model predictive control algorithm.

[0038] S04. After the extruded melt passes through the die, the Kalman filter soft measurement module is used to estimate the melt molecular weight in real time. When the estimated melt molecular weight is lower than the set threshold, vacuum exhaust treatment is added and the drying time of the drying process is increased.

[0039] S05. The composite granules obtained by extrusion granulation are injection molded to prepare standard specimens. The Reduced StrainClosure-machine learning correction model is used to predict the three-dimensional orientation tensor of glass fiber in the full flow field of injection molding, and the holding pressure and injection rate in the injection molding process are adjusted according to the prediction results.

[0040] S06. Test the impact strength and elongation at break of the obtained sample. If the test value is lower than the target threshold, the test data is sent back to the multi-scale spatiotemporal graph convolution-physical embedding fusion prediction model for incremental training, and then returned to S02 to readjust the formula process parameters.

[0041] The principle and implementation of the Fick diffusion-Arrhenius coupled interface coupling reaction kinetics inverse estimation algorithm are as follows. This algorithm models the chemical adsorption process of silane coupling agent molecules on the glass fiber surface as a coupled partial differential equation system of two-dimensional Fick diffusion equation and Arrhenius reaction rate equation. Using measured data of coupling agent adsorption on the glass fiber surface at different temperatures and coating times as constraints, the algorithm employs the quasi-Newton method (L-BFGS-B) to inversely estimate three kinetic parameters: diffusion coefficient, activation energy, and pre-exponential factor. Specifically, the continuous partial differential equation is discretized into a finite difference scheme, and a Jacobian matrix is ​​constructed. The parameters are iteratively updated along the gradient descent direction in the parameter space. After each iteration, the mass conservation residual is checked. The convergence criterion is that the Frobenius norm of the parameter change is lower than... The algorithm ultimately outputs an optimal set of kinetic parameters to determine the temperature profile and upper limit of the coupling agent coating dosage. The technical advantages of this algorithm are as follows: Traditional coupling agent dosage optimization relies on numerous repeated experiments and lacks mechanistic guidance. This algorithm, based on atomic-scale adsorption kinetic equations, obtains physically meaningful kinetic parameters through reverse deduction. This transforms the design of the temperature profile for the coupling agent coating process from empirical exploration to parameter-supported quantitative design, fundamentally solving the thermodynamic competition problem of weak interfacial bonding due to insufficient coupling agent and embrittlement of the interfacial layer due to excessive coupling agent. Simultaneously, it significantly reduces the number of screening experiments and improves the repeatability and traceability of the interfacial treatment process.

[0042] The Fick diffusion equation is a partial differential equation describing the migration of substances to lower concentration regions driven by a concentration gradient. The spatiotemporal evolution formula of the coupling agent concentration field in two-dimensional form is expressed as follows: ;in The concentration of the coupling agent ( ), Reference concentration ( ), For time (s), Reference time (s) The diffusion coefficient ( ), Reference diffusion coefficient ( ), , For spatial coordinates (m) The reference length is (m).

[0043] The Arrhenius rate equation describes the relationship between the chemical reaction rate constant and temperature, and is expressed as follows: ;in The reaction rate constant ( ), The reference rate constant ( ), Pre-exponential factor ( ), For reference pre-exponential factor ( ), Activation energy ( ), molar gas constant ( ), The absolute temperature is K. Reference temperature (K).

[0044] The Frobenius norm is the arithmetic square root of the sum of the squares of the matrix elements, used to measure the overall change of the parameter matrix between two adjacent iterations. When this value is lower than... The convergence of the reverse iteration is determined at the time.

[0045] The diffusion coefficient, activation energy, pre-exponential factor, inverse threshold, and reference values ​​were obtained as follows: Six temperature gradients were set at 10°C intervals within the range of 20–80°C. At each temperature, the amount of coupling agent adsorbed on the glass fiber surface was measured at coating times of 30 s, 60 s, 120 s, and 300 s, resulting in approximately 30 measured data points. Using these data points as constraints, the inverse algorithm was iteratively run multiple times to calculate the mean and standard deviation of the parameter estimation results. The parameter mean corresponding to a ratio of standard deviation to mean of less than 5% was used as the reference value, and the design range of the coupling agent coating temperature curve was determined based on this reference value.

[0046] The specific structure of the multi-scale spatiotemporal graph convolution-physical embedding fusion prediction model (hereinafter referred to as the fusion prediction model) is as follows: This model uses a graph structure to represent the functional sections of a twin-screw extruder. Each node in the graph corresponds to a screw functional section (including the feeding section, melting section, mixing section, and homogenizing section). The directed edges between nodes represent the flow connection relationship of materials along the screw axis. The input feature vector of each node is composed of the temperature time-series data, torque time-series data, and online near-infrared spectral data of the corresponding functional section, with a dimension of [missing information]. The bottom layer is a temporal convolutional network module, which performs causal convolution on the input feature vector of each node along the time axis to extract local temporal dynamic features. The receptive field width of the temporal convolutional network module is dynamically determined according to the screw speed parameter. Specifically, when the screw speed is higher than 150 r / min, the receptive field width is set to 32 time steps, and when the screw speed is lower than or equal to 150 r / min, it is set to 16 time steps. This threshold is obtained by performing autocorrelation analysis on 50 sets of time-series data at different speeds, taking the average of the number of time steps corresponding to the autocorrelation coefficient dropping below 0.2 and rounding it down. The middle layer is a graph convolutional module, which uses a variant of GraphSAGE to perform neighborhood aggregation on the hidden states of each node in the graph, passing the material state information of adjacent functional sections to the current node. After each graph convolution, the node features are normalized. The middle layer introduces a message passing iteration mechanism, which performs a maximum of 6 iterations. After each iteration, the hidden state vector of each node is used as the result. Norm change is used as a convergence criterion; when all nodes... The changes were all lower than The iteration stops when the time is right. The top layer is the physical constraint layer, which embeds the analytical expressions of the Carreau viscosity equation and the fiber fracture Weibull probability model in a differentiable form into the network as a physical prior layer. This layer forces the viscosity prediction value output by the model to satisfy the physical conservation relationship with the fiber fracture probability distribution. The physical constraint layer and the data-driven layer are fused using a residual connection method. The output of the physical layer serves as the residual basis, and the output of the data-driven layer serves as the superposition of correction terms. This ensures that when the data-driven layer produces systematic errors due to insufficient training samples, the physical layer can still provide stable prior constraints. The final output of the model is the predicted value of the glass fiber length distribution index and the material state vector of each functional segment. At the same time, the model supports the condition generation function. Given the target glass fiber length distribution, the optimal screw speed and temperature curve parameter suggestions are obtained by backpropagation to the input layer.

[0047] The technical effects of the fusion prediction model are as follows. This model integrates heterogeneous time-series signals such as temperature, torque, and online spectra from various screw sections into a graph structure for collaborative modeling. The bottom-layer temporal convolutional network extracts local dynamics, the middle-layer graph convolutional network transmits material information between sections, and the top-layer physical constraint layer embeds the viscosity equation and fiber fracture model. These three layers are connected by residuals to form an organic whole. The introduction of the physical constraint layer ensures that the model maintains physical consistency even with insufficient training data, avoiding the risk of pure data-driven models failing in the extrapolation range. The message-passing iteration mechanism ensures that the material state of each node in the graph reaches self-consistent convergence before output, allowing the prediction results to reflect the global material evolution law of the screw rather than local approximations. The condition generation function enables reverse optimization by deriving the optimal screw parameters from the target fiber length distribution, transforming the traditional empirically-dependent forward trial-and-error process into a target-driven quantitative optimization process. This fundamentally solves the problem of uncontrollable glass fiber length distribution during extrusion and significantly compresses the process development cycle.

[0048] The steps for establishing the training dataset for the fusion prediction model specifically include: collecting at least 50 sets of twin-screw extrusion experimental data under different screw configurations, speeds, and temperature curves. Each set of data includes full-process temperature time-series data, torque time-series data, and online near-infrared spectral data. After extrusion, offline sampling is performed to determine the glass fiber length distribution and the impact strength and elongation at break of the final product. The above multi-source data are divided according to the screw functional segments, and a node feature matrix and a graph adjacency matrix are constructed to form a graph sample. The glass fiber length distribution index, impact strength, and elongation at break are used as supervision labels. The training set, validation set, and test set are divided in a 7:2:1 ratio.

[0049] The specific steps for training the fusion prediction model include: using the weighted sum of the mean squared error loss function and the physical residual penalty term as the total loss function, where the physical residual penalty term consists of the equation residuals of the Carreau viscosity equation and the Weibull fiber fracture model; and using the Adam optimizer to update the parameters, with an initial learning rate set to... If the validation set loss does not decrease for 10 consecutive rounds, the learning rate is reduced by a factor of 0.5. After training until the validation set loss stabilizes, the generalization performance is evaluated using the test set. After training, the model weights are saved for online inference in S02 and incremental training in S06.

[0050] The improved Reduced Strain Closure-Machine Learning Correction Model is based on the Folgar-Tucker fiber orientation constitutive equation. It introduces the Reduced Strain Closure closed approximation to reduce the tensor solution dimension. At the same time, a machine learning regression model (mainly based on a multilayer perceptron) is used to perform data-driven correction on the fiber interaction coefficients in the Folgar-Tucker equation, so that it can still maintain high accuracy when the fiber volume fraction is higher than 20%. The correction model uses a GPU parallel finite element solver to solve the three-dimensional tensor field of the entire flow field, compressing the simulation time of a single injection molding to an acceptable range.

[0051] Among them, the fiber length protection screw configuration refers to a screw arrangement in a twin-screw extruder where the number of meshing blocks is reduced and the length of the conveying section is extended, thereby reducing the mechanical breaking effect of the high-shear zone on the glass fibers and protecting the fibers to maintain a longer length distribution.

[0052] Among them, the online near-infrared-Raman combined spectroscopy system refers to the system in which near-infrared spectroscopy probes and Raman spectroscopy probes are installed at the extruder head to collect melt spectral data in real time. Through a pre-established partial least squares regression or support vector regression correction model, the spectral signal is converted into an online estimation system for glass fiber length distribution index and melt degradation index in real time.

[0053] The glass fiber length distribution index is a dimensionless value that characterizes the uniformity of glass fiber distribution by the ratio of number-average fiber length to weight-average fiber length after statistical analysis of the glass fiber length distribution in the extruded melt. The closer the value is to 1, the more concentrated the distribution. The value ranges from 0 to 1. The threshold of this index is obtained by offline fiber length measurement of no less than 20 samples under different process conditions, statistical analysis of the lower limit of the glass fiber length distribution index corresponding to the condition that the impact strength and elongation at break both meet the target values, and averaging the results of multiple experiments as the threshold.

[0054] The melt degradation index is a dimensionless value that estimates the relative decrease in melt molecular weight by comparing the intensity ratio of characteristic peaks in the online near-infrared spectrum. The calculation formula is as follows: ;in The intensity of the degradation characteristic peak (in any unit). Reference peak intensity (any unit). For online estimation of molecular weight ( ), The initial molecular weight of the raw material ( The threshold for the melt degradation index was obtained by performing end-group titration and gel permeation chromatography offline detection on no less than 15 groups of samples under different drying conditions and extrusion temperatures. The ratio of the intensity of the characteristic peaks in the online near-infrared spectrum corresponding to the detected molecular weight decrease of more than 10% was used as the threshold benchmark. The average value was determined after more than 3 repeated experiments.

[0055] The Kalman filter soft measurement module refers to a software measurement system that uses online near-infrared spectral signals as observations, polyamide hydrolysis degradation kinetics equations as state equations, and recursively estimates the melt molecular weight using a Kalman filter algorithm to output real-time molecular weight estimates. The threshold is obtained by comparing the offline end-group titration data of no less than 20 extrusion experiments with the Kalman filter estimates, statistically analyzing the lowest estimated molecular weight when the deviation is within 5%, and using this as the threshold for initiating vacuum degassing. This threshold is determined after verification through no less than 3 rounds of iterative experiments.

[0056] Among them, model predictive control algorithm refers to using process mechanism model as prediction equation, rolling prediction of process output in the future finite time domain in each control cycle, solving for the optimal control input at the current moment in the rolling optimization way, and applying the obtained optimal control input to the actual process, and solving again in the next cycle to achieve coordinated control under multivariable constraints.

[0057] The Carreau viscosity equation is a constitutive equation that describes the change of viscosity of non-Newtonian polymer melts with shear rate. It treats the melt as a shear-thinning fluid, where the viscosity gradually decreases from zero shear viscosity to infinite shear viscosity as the shear rate increases.

[0058] The Weibull fiber fracture probability model describes the probability of glass fiber fracture under a given shear stress using the Weibull probability distribution function, where the scale and shape parameters are obtained by fitting experimental data of glass fiber mechanical properties.

[0059] Among them, the GraphSAGE variant is an improved form based on the GraphSAGE graph convolution algorithm. GraphSAGE updates the hidden state of a node by sampling the neighborhood and aggregating the neighborhood features. In this scheme, attention weighting is introduced into its aggregation function so that the influence weight of different screw functional segments on the current node is dynamically adjusted according to the material flow rate and temperature difference.

[0060] Among them, node features Convergence threshold of norm change The following methods were used to obtain the data: Ablation experiments were conducted on the fusion prediction model on the training dataset, and the changes in prediction accuracy at different iteration numbers were recorded. The node features corresponding to a prediction accuracy improvement of no more than 0.1% were statistically analyzed. The amount of change was determined by taking the average of multiple experiments.

[0061] Three example recipes are given below.

[0062] Formula 1: 65% polyamide 66 chips, 30% glass fiber (6 mm filament length), 0.8% silane coupling agent (aminosilane type), 0.3% antioxidant 1098 (hindered phenol type), 0.5% lubricant (ethylene bis-stearamide), 0.4% heat stabilizer (hindered amine type), and the balance is processing aids; extrusion temperature range 260-280℃, screw speed 280 r / min, drying temperature 110℃, drying time 8 h, and moisture content controlled below 80 ppm.

[0063] Formula 2: 60% polyamide 6 chips, 35% glass fiber (6 mm filament length), 1.0% silane coupling agent (epoxysilane type), 0.3% antioxidant 168 (phosphite type), 0.4% lubricant (calcium stearate), 0.5% heat stabilizer (copper salt composite type), 1.5% compatibilizer (maleic anhydride grafted polyamide), and the balance being processing aids; extrusion temperature range 240-260℃, screw speed 260 r / min, drying temperature 90℃, drying time 6 h, and moisture content controlled below 100 ppm.

[0064] Formula 3: 55% polyamide 66 chips, 40% glass fiber (6 mm filament length), 1.2% silane coupling agent (aminosilane type), 0.4% antioxidant 1098 (hindered phenol type), 0.2% antioxidant 168 (phosphite type), 0.6% lubricant (ethylene bis-stearamide), 0.5% heat stabilizer (hindered amine type), 2.0% toughening agent (maleic anhydride-grafted EPDM rubber), with the balance being processing aids; extrusion temperature range 265–285℃, screw speed 250 r / min, drying temperature 120℃, drying time 8 h, and moisture content controlled below 80 ppm.

[0065] The specific implementation of step S01 is as follows: Polyamide chips are placed in a forced-air drying oven, with the drying temperature set between 80 and 120°C, and the drying time between 4 and 8 hours, reducing the moisture content of the chips to below 100 ppm to prevent water-induced hydrolysis and degradation of the polyamide during extrusion. Simultaneously, the glass fiber undergoes a silane coupling agent surface treatment. The treatment process parameters are determined by an inverse algorithm based on the interfacial coupling reaction kinetics coupled with the Fick diffusion and Arrhenius reaction rate equations. This algorithm models the chemical adsorption process of coupling agent molecules on the glass fiber surface as a coupled partial differential equation system consisting of a two-dimensional Fick diffusion equation and an Arrhenius reaction rate equation. Using measured coupling agent adsorption data at different temperatures and coating times as constraints, a quasi-Newton method is used to inversely estimate the diffusion coefficient, activation energy, and pre-exponential factor. The convergence criterion is that the Frobenius norm of the parameter change in two adjacent iterations is lower than... The final output is the optimal set of kinetic parameters, which is used to determine the temperature curve and upper limit of the amount of coupling agent to be coated, thereby avoiding the problems of weak interfacial bonding due to too little coupling agent or embrittlement of the interfacial layer due to too much coupling agent.

[0066] The specific implementation of step S02 is as follows: Surface-treated glass fibers and dried polyamide chips are fed into a twin-screw extruder with a fiber length protection screw configuration through the side feed port according to the formula mass ratio. This screw configuration reduces the mechanical breakage effect of the high-shear zone on the glass fibers by reducing the number of meshing blocks and extending the conveying section length. After the extruder starts, an online near-infrared Raman combined spectroscopy system is started simultaneously. Near-infrared spectroscopy probes and Raman spectroscopy probes are installed at the die head position to collect melt spectral data in real time. The spectral signals are converted into glass fiber length distribution index and melt degradation index through a pre-established partial least squares regression or support vector regression correction model. A multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model represents each screw functional segment with a graph structure. The bottom-level temporal convolutional network extracts the local temporal dynamic features of each node, and the receptive field width is dynamically determined according to the screw speed. The middle-level GraphSAGE variant graph convolutional module transmits the material state between segments through attention-weighted neighborhood aggregation, and message passing iterates to each node. Norm change is lower than The model converges in real time; the top-level physical constraint layer embeds the Carreau viscosity equation and the Weibull fiber fracture probability model in a differentiable form, and merges them with the data-driven layer through residual connections, so that the model can output optimal screw parameter suggestions in real time.

[0067] The specific implementation of step S03 is as follows: Using the glass fiber length distribution index and melt degradation index output by the online near-infrared La Mans combined spectroscopy system in S02 as process variables, a model predictive control algorithm is used to perform rolling predictions of the process output within a finite time domain in each control cycle. The optimal control input at the current moment is solved by rolling optimization, and multi-variable coordinated adjustments are made to the screw speed, temperature of each section, and material residence time. When the glass fiber length distribution index is lower than a preset threshold, the model predictive control algorithm prioritizes reducing the screw speed or lowering the temperature of the high-shear section to weaken the mechanical and thermal degradation effect on the fiber. When the melt degradation index exceeds the threshold, the model predictive control algorithm adjusts the uniformity constraint of the temperature of each section in conjunction to prevent local overheating from causing a further decrease in molecular weight, thereby achieving coordinated control under multi-variable constraints.

[0068] The specific implementation of step S04 is as follows: After the extruded melt is formed by the die, the Kalman filter soft measurement module uses the online near-infrared spectral signal as the observation and the polyamide hydrolysis degradation kinetic equation as the equation of state to estimate the melt molecular weight in real time through the Kalman filter recursive algorithm. When the estimated value is lower than the set molecular weight threshold, the system determines that the melt has an excessive risk of hydrolysis degradation, automatically triggers vacuum degassing to remove residual moisture, and simultaneously adjusts the drying time of the drying process to no less than the upper limit to reduce the amount of moisture introduced from the source. The molecular weight threshold is determined by comparing and statistically analyzing the end-group titration offline data of no less than 20 sets of extrusion experiments with the Kalman filter estimated value, taking the lowest estimated molecular weight mean with a deviation within 5%, and verifying it through no less than 3 rounds of iterative experiments.

[0069] The specific implementation of step S05 is as follows: The composite granules obtained in S04 are molded into standard splines using an injection molding machine. During the injection molding process, an improved Reduced Strain Closure and machine learning correction model are used to predict the three-dimensional orientation tensor of the glass fibers in the entire flow field. This model introduces the Reduced Strain Closure approximation to reduce the dimensionality of the tensor solution based on the Folgar-Tucker fiber orientation constitutive equation. At the same time, a multilayer perceptron is used to perform data-driven correction on the interaction coefficients between fibers, so that it can maintain high accuracy even when the fiber volume fraction is higher than 20%. A GPU parallel finite element solver is used to solve the three-dimensional tensor field of the entire flow field, compressing the simulation time to an acceptable range. Based on the predicted three-dimensional orientation tensor distribution, the holding pressure and injection rate in the injection molding process are adjusted accordingly to make the fiber orientation distribution tend to the target state and improve the anisotropic mechanical uniformity of the product.

[0070] The specific implementation of step S06 is as follows: The standard sample obtained in S05 is tested for impact strength and elongation at break according to national standards. If the test value is lower than the target threshold, the screw parameters, spectral data, fiber length distribution and test results of this experiment are used as new samples and fed back to the multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model for incremental training. The weighted sum of the mean square error loss function and the physical residual penalty term is used as the total loss function. The Adam optimizer is used to update the model weights. After the validation set loss tends to stabilize, the updated model weights are saved. Then, the process parameters are readjusted in S02 to form a closed-loop iterative optimization process until the test value meets the target threshold.

[0071] It should be noted that the key technologies of this invention include: First, a kinetic inverse calculation algorithm coupling Fick diffusion and the Arrhenius reaction rate equation transforms the coupling agent adsorption process from empirical exploration to quantitative design based on atomic-scale kinetic equations. By inversely calculating, physically meaningful kinetic parameters are obtained, enabling a quantitative solution at the mechanistic level to the thermodynamic competition problem of coupling agent dosage, fundamentally changing the passive situation of traditional repeated trial-and-error screening processes. Second, a multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model. This model uses a unified graph structure to model the multi-segment coupled material evolution process. The bottom-layer temporal convolutional network captures local dynamics, the middle-layer graph convolution transmits inter-segment information, and the top-layer physical constraint layer uses the Carreau viscosity equation and the Weibull fiber fracture model as prior constraints, ensuring that the prediction results mathematically satisfy physical conservation. This compensates for the failure of purely data-driven models in the extrapolation range and achieves inverse optimization of the optimal screw parameters by inversely calculating the target fiber length distribution through a condition generation function. The two key technologies mentioned above work together, with the former ensuring the precision of the interface treatment process and the latter realizing intelligent closed-loop control of the extrusion process. Together, they form a complete technology chain from interface design to process optimization, which fundamentally ensures the controllability of glass fiber length distribution and ultimately endows the material with stable high toughness.

[0072] It should be noted that this invention also solves the following technical problem: In the injection molding stage, the distribution of the three-dimensional orientation tensor of glass fibers directly affects the mechanical properties of the product in all directions. Traditional injection molding processes rely on experience to set the holding pressure and injection rate, which cannot quantitatively predict the three-dimensional distribution of the fiber orientation field, resulting in anisotropic mechanical fluctuations in the product. This invention employs an improved Reduced Strain Closure and machine learning correction model. Based on the Folgar-Tucker constitutive equation, it introduces the Reduced Strain Closure closed approximation to reduce the tensor solution dimension, and uses a multilayer perceptron to perform data-driven correction of the fiber interaction coefficients. A GPU parallel finite element solver is used to achieve rapid solution of the three-dimensional tensor field of the entire flow field. Based on the predicted orientation tensor distribution, the injection molding process parameters are quantitatively adjusted to make the fiber orientation distribution tend towards the target state, thereby solving the technical problem of anisotropic mechanical fluctuations in the product caused by the unpredictable three-dimensional orientation distribution of glass fibers during injection molding.

[0073] Specifically, the principle of this invention is as follows: The reason this invention can solve the technical problem of uncontrollable glass fiber length distribution lies in its solution, which simultaneously addresses the issue from two dimensions: signal fusion and physical mechanisms. Graph structures are naturally suitable for characterizing the directed material flow relationships between different functional segments of the screw. The GraphSAGE variant, through attention-weighted neighborhood aggregation, allows the temperature and material flow differences between functional segments to quantitatively affect the hidden state updates of adjacent nodes, thereby capturing the strong coupling dynamics between multiple segments. The receptive field width of the temporal convolutional network is dynamically determined based on the screw speed, ensuring effective extraction of local temporal dynamics under different operating conditions. The physical constraint layer embeds the Carreau viscosity equation and the Weibull fiber fracture probability model in a differentiable form, making the network output mathematically constrained by physical conservation relationships—a generalization guarantee not provided by purely data-driven models. Furthermore, the condition generation function, through backpropagation to the input layer, achieves reverse optimization by inferring the optimal screw parameters from the target fiber length distribution, transforming forward trial-and-error parameter tuning into target-driven quantitative optimization. Logically, this constitutes a complete closed-loop control system, conforming to the basic principles of engineering cybernetics.

[0074] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0075] The specific implementation of step S01 is as follows: Polyamide chips are placed in an oven and dried at 80–120°C for 4–8 hours until the moisture content drops below 100 ppm. When performing silane coupling agent surface treatment on glass fibers, the Fick diffusion-Arrhenius coupled interface coupling reaction kinetics inverse algorithm is used to determine the coupling agent coating temperature curve and dosage. This algorithm models the chemical adsorption process of coupling agent molecules on the glass fiber surface as a coupled partial differential equation system of two-dimensional Fick diffusion equation and Arrhenius reaction rate equation. The spatiotemporal evolution formula of the coupling agent concentration field is expressed as follows:

[0076] ;

[0077] In the formula, This refers to the concentration of the coupling agent, in units of... For reference concentration, the initial coating concentration is used, in units of... For time, in seconds. For reference time, 30 seconds is used, unit: seconds. The diffusion coefficient is expressed in units of... For reference diffusion coefficient, units , Spatial coordinates, unit: meters For reference length, the characteristic dimension of the fiber surface treatment area is taken, in meters. The left side of the formula is the partial derivative of dimensionless concentration with respect to dimensionless time, and the right side... The value is the dimensionless diffusion coefficient, multiplied by the sum of the second-order partial derivatives of the dimensionless concentration with respect to the dimensionless spatial coordinates; all terms are dimensionless. The relationship between the reaction rate constant and temperature is described by the Arrhenius equation, expressed as follows:

[0078] ;

[0079] In the formula, The reaction rate constant is expressed in units of 1 / 3 and 2 / 3. The reference rate constant is expressed in units of... Pre-exponential factor, unit For reference to the pre-exponential factor, the unit For activation energy, unit Let be the molar gas constant, taken as 8.314. Absolute temperature, unit K. For reference temperature, take 293.15 K. (Unit: K) Left side of the formula. The ratio of dimensionless rate constants is shown on the right. For dimensionless pre-exponential factor ratios, the exponential term... Both the numerator and denominator are ratios with energy dimensions, and the whole is dimensionless. The coupling of these two equations forms a system of partial differential equations. The continuous partial differential equations are discretized into a finite difference scheme, and a Jacobian matrix is ​​constructed. Six temperature gradients are set at 10℃ intervals within the range of 20–80℃. At each temperature, samples are taken at coating times of 30s, 60s, 120s, and 300s to measure the adsorption amount of coupling agent on the glass fiber surface, resulting in approximately 30 measured data points. Using these data points as constraints, a bounded finite memory quasi-Newton method is employed to measure the diffusion coefficient. ,activation energy Pre-exponential factors Three dynamic parameters are estimated inversely. After each iteration, the change in the Frobenius norm of the parameter column vector is used as the convergence criterion. The definition is as follows:

[0080] ;

[0081] In the formula, For the first The estimated value of the diffusion coefficient at the next iteration, in units. For the first The estimated activation energy at the next iteration, in units. The activation energy reference value was obtained from preliminary fitting of experimental data, and the unit is... For the first The estimated value of the pre-exponential factor at the next iteration, in units. All three terms, when divided by their respective reference values, are dimensionless quantities, making... All components have the same dimensions. The convergence criterion formula is expressed as follows:

[0082] ;

[0083] In the formula, The Frobenius norm, defined as the arithmetic square root of the sum of the squares of the elements of a matrix or vector, is applied here to the difference of the dimensionless parameter column vectors, resulting in a dimensionless quantity. When the above convergence condition is met, the optimal set of kinetic parameters is output. The mean of the parameters corresponding to a standard deviation to mean ratio of the parameter estimation results being less than 5% is used as a reference value to determine the temperature profile and upper limit of the coupling agent coating.

[0084] The specific implementation of step S02 is as follows: Surface-treated glass fibers and dried polyamide chips are fed into a twin-screw extruder with a fiber length protection screw configuration at a specific mass ratio. This screw configuration reduces the number of meshing blocks and extends the conveying section length, reducing mechanical breakage of the glass fibers in the high-shear zone. Simultaneously, an online near-infrared-La Mansell-Robbins combined spectroscopy system is activated to acquire melt spectral data in real time. The spectral signals are then converted into online estimates of the glass fiber length distribution index and melt degradation index using a partial least squares regression or support vector regression correction model. The glass fiber length distribution index is defined as the ratio of number-average fiber length to weight-average fiber length, expressed by the following formula:

[0085] ;

[0086] In the formula, The glass fiber length distribution index is dimensionless and ranges from 0 to 1. Number-average fiber length, in meters. Weight-average fiber length, in meters (m) The closer the value is to 1, the more concentrated the distribution. and Dimensionless quantities are obtained by dividing quantities that have the same dimensions. The formula for the melt degradation index is as follows:

[0087] ;

[0088] In the formula, The intensity of the degradation characteristic peak, in any unit. For reference peak intensity, any unit, and Same unit For online molecular weight estimation, units The initial molecular weight of the raw material, in units ; left side of the formula The ratio of dimensionless spectral intensities is shown on the right. middle The molecular weight ratio is dimensionless, and all terms are dimensionless. The multi-scale spatiotemporal graph convolution-physical embedding fusion prediction model represents the functional sections of a twin-screw extruder using a graph structure. Each node in the graph corresponds to a screw functional section, and the directed edges between nodes represent the flow connections of material along the screw axis. The input feature vector of each node is composed of the temperature time-series data, torque time-series data, and online near-infrared spectral data of the corresponding functional section, with a dimension of [missing information]. , The value is a positive integer, determined by the sum of the feature dimensions of the three data types. The underlying temporal convolutional network module performs causal convolution along the time axis on the input feature vector of each node to extract local temporal dynamic features. The receptive field width is determined based on the screw rotation speed. (unit Dynamically determined: when The receptive field width is set to 32 time steps. The time step is set to 16. This threshold is obtained by taking the average of the number of time steps corresponding to the autocorrelation coefficient falling below 0.2 from the autocorrelation analysis of 50 sets of time-series data at different rotation speeds, and then rounding it down. The middle-layer graph convolution module uses a directed graph aggregation algorithm with attention weighting to perform neighborhood aggregation of the hidden states of each node. The aggregation weight is dynamically adjusted according to the differences in material flow rate and temperature. After each graph convolution, the node features are normalized, and a message-passing iteration mechanism is introduced, with a maximum of 6 iterations. The convergence criterion formula is expressed as follows:

[0089] ;

[0090] In the formula, For the first During the nth iteration The hidden state vector of each node has a dimension of . For vectors norm For numerically stable terms, the empirical value is To prevent the denominator from being zero; the numerator on the left side of the formula is the difference between the hidden state vectors of two adjacent iterations. Norm, the denominator is the current iteration hidden state vector The norm plus the stabilization term results in a dimensionless quantity. The criterion requires that iteration stop when all nodes in the graph satisfy the above inequality. This threshold is determined by the node characteristics corresponding to a statistical prediction accuracy improvement of no more than 0.1% in ablation experiments. The variation is determined by taking the mean. The top-level physical constraint layer embeds the analytical expressions of the Carlow viscosity equation and the Weibull fiber fracture probability model into the network in a differentiable form, forcing the viscosity prediction value output by the model to satisfy the physical conservation relationship with the fiber fracture probability distribution. The physical constraint layer and the data-driven layer are fused using residual connections, with the output of the physical layer serving as the residual basis and the output of the data-driven layer serving as the superposition of correction terms. The model ultimately outputs optimal screw parameter suggestions in real time.

[0091] The specific implementation of step S03 is as follows: based on the glass fiber length distribution index output by the online near-infrared-Laman combined spectroscopy system. With melt degradation index The model predictive control algorithm is used to coordinate and adjust the screw speed, temperature at each stage, and residence time using multiple variables. The model predictive control algorithm uses the process mechanism model as the prediction equation and performs rolling prediction of the process output in the future finite time domain within each control cycle. It solves for the optimal control input at the current moment in a rolling optimization manner, and then solves again in the next cycle to achieve coordinated control under multivariable constraints.

[0092] The specific implementation of step S04 is as follows: After the extruded melt passes through the die, the Kalman filter soft measurement module uses the online near-infrared spectral signal as the observation and the polyamide hydrolysis degradation kinetic equation as the equation of state to recursively estimate the melt molecular weight using the Kalman filter algorithm. It outputs real-time estimates. A threshold is set by comparing offline end-group titration data from at least 20 extrusion experiments with Kalman filter estimates, statistically analyzing the lowest estimated molecular weight when the deviation is within 5%, and verifying this through at least three rounds of iterative experiments. If the estimated value is lower than the set threshold, vacuum exhaust treatment is added, and the drying time of the drying process is increased.

[0093] The specific implementation of step S05 is as follows: The composite granules obtained from extrusion granulation are injection molded to prepare standard specimens. An improved simplified strain closure-machine learning correction model is used to predict the three-dimensional orientation tensor of glass fibers in the entire injection molding flow field. This model introduces a simplified strain closure approximation to reduce the dimensionality of the tensor solution based on the Fogg-Tucker fiber orientation constitutive equation. Simultaneously, a multilayer perceptron is used to perform data-driven correction on the fiber interaction coefficients in the Fogg-Tucker equation, ensuring high accuracy even when the fiber volume fraction is higher than 20%. A parallel finite element solver using a graphics processor is employed to solve the three-dimensional tensor field of the entire flow field, compressing the simulation time for a single injection molding cycle to an acceptable range. The holding pressure and injection rate in the injection molding process are adjusted based on the prediction results.

[0094] The specific implementation of step S06 is as follows: Impact strength and elongation at break are tested on the obtained sample. If the tested values ​​are lower than the target threshold, the test data is fed back to the multi-scale spatiotemporal graph convolutional-physical embedding fusion prediction model for incremental training, and then returned to S02 to readjust the formulation and process parameters. The total loss function formula for incremental training is expressed as follows:

[0095] ;

[0096] In the formula, The total loss function value, and Same dimensions As a reference loss value, the mean squared error loss value from the initial training epochs is used, and compared with... , Same dimensions The mean squared error loss term is calculated from the squared mean of the differences between the predicted value and the supervised label. This is a physical residual penalty term, consisting of the equation residuals from the Carlow viscosity equation and the Weibull fiber fracture model. , These are weighting coefficients, with empirical values ​​of 0.8 and 0.2 respectively, satisfying... Divide each term in the formula by All values ​​thereafter are dimensionless. Incremental training uses an adaptive moment estimator optimizer to update parameters, with an initial learning rate set to... If the validation set loss does not decrease for 10 consecutive rounds, the learning rate is decayed by a factor of 0.5. After training, the model weights are saved for subsequent online inference.

[0097] To better understand and implement this invention, the following is an example 2 of a specific application scenario: Based on Formula 3, the technicians built a test scenario, selecting 55% polyamide 66 chips, 40% glass fiber, 1.2% aminosilane coupling agent, 0.4% antioxidant 1098, 0.2% antioxidant 168, 0.6% ethylene bis-stearamide lubricant, 0.5% hindered amine heat stabilizer, 2.0% maleic anhydride-grafted EPDM toughening agent, and the balance being processing aids. The extrusion temperature range was 265-285℃, the screw speed was 250 r / min, the drying temperature was 120℃, the drying time was 8 h, and the moisture content was controlled below 80 ppm.

[0098] In stage S01, technicians set six temperature gradients at 10°C intervals within the range of 20–80°C. At each temperature, samples were taken at coating times of 30 s, 60 s, 120 s, and 300 s, resulting in 30 measured adsorption data points. These data were then input into a kinetic inverse estimation algorithm coupling the Fick diffusion and Arrhenius reaction rate equations for parameter estimation. After multiple iterations, the Frobenius norm of the parameter changes converged to [value missing]. The following reference values ​​for diffusion coefficient, activation energy, and pre-exponential factor were obtained. The ratio of the standard deviation to the mean of the parameter estimation results was less than 5%, confirming that the parameter stability met the requirements. Based on this, the coating temperature curve and the upper limit of dosage of aminosilane coupling agent were determined to be 1.2%.

[0099] In stage S02, technicians feed surface-treated glass fibers into a twin-screw extruder with a fiber length protection screw configuration through the side feed port, and acquire melt data in real time using an online near-infrared pull-and-trade spectroscopy system. A multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model represents four functional nodes—feeding section, melting section, mixing section, and homogenization section—using a graph structure. The receptive field width of the bottom-layer temporal convolutional network is set to 16 time steps based on a rotation speed of 250 r / min. The middle-layer GraphSAGE variant graph convolutional module undergoes 6 message-passing iterations to refine each node. The change is lower than Convergence occurs, and the top-level physical constraint layer outputs prediction results that satisfy the constraints of the Carreau viscosity equation and the Weibull fiber fracture model, such as... Figure 3As shown, the glass fiber length distribution index output by the model in real time remains at a high level during the steady-state stage of extrusion, indicating that the fiber length distribution is concentrated. The predicted values ​​of the glass fiber length distribution index output by each functional segment under different process conditions are shown in Table 1.

[0100] Table 1. Predicted values ​​of glass fiber length distribution index for each functional segment under different process conditions.

[0101]

[0102] In stage S03, the model predictive control algorithm adjusts the screw speed, temperature of each section and residence time in a multivariate manner based on the glass fiber length distribution index and melt degradation index output by the online spectral system, so as to keep the glass fiber length distribution index above the target threshold and the melt degradation index below the threshold.

[0103] In stage S04, the Kalman filter soft measurement module estimates the molecular weight of the melt in real time. In this test scenario, the estimated molecular weight value was always higher than the set threshold, and the vacuum exhaust process was not triggered. The drying time remained unchanged at 8 hours.

[0104] In stage S05, technicians improved the injection molding standard spline by using the Reduced Strain Closure and machine learning correction model to predict the three-dimensional orientation tensor of the entire flow field. The GPU parallel finite element solver compressed the single injection simulation time to an acceptable range. Based on the predicted three-dimensional orientation tensor distribution, the holding pressure and injection rate were adjusted. The orientation tensor prediction results matched the actual detection results well.

[0105] In stage S06, the impact strength and elongation at break of the spline were tested. The test results are shown in Table 2. Both indicators met the target threshold requirements, so there was no need to trigger incremental training and parameter readjustment.

[0106] Table 2. Results of spline mechanical property tests

[0107]

[0108] Compared to traditional methods, this invention represents a significant advancement in technical principles. Traditional processes rely on manual experience to set the coupling agent dosage and screw parameters, lacking a mechanistic description of interfacial adsorption kinetics and the evolution of multi-segment coupled materials. This results in a lack of quantitative basis for process adjustments and insufficient controllability of fiber length distribution. This invention, through a kinetic inverse algorithm, establishes the coupling agent dosage design based on physically meaningful kinetic parameters. By incorporating multi-source heterogeneous time-series signals into a graph structure through a fusion prediction model and ensuring the physical consistency of predictions with a physical constraint layer, process adjustments are transformed from experience-based trial and error to goal-driven quantitative optimization. This fundamentally improves the prediction accuracy and controllability of glass fiber length distribution, thereby endowing the material with stable high toughness.

[0109] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0110] Table 3. Variable Explanation Table

[0111]

[0112] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for preparing a high-toughness glass fiber reinforced polyamide material, characterized in that, Includes the following steps: The polyamide chips were dried, and the glass fibers were surface-treated with silane coupling agent. The coating temperature curve and dosage of the coupling agent were determined by the inverse algorithm of interfacial coupling reaction kinetics coupled with Fick diffusion and Arrhenius reaction rate equation. Surface-treated glass fibers and dried polyamide chips are fed into a twin-screw extruder with a fiber length protection screw configuration. Simultaneously, an online near-infrared pull-and-trade spectroscopy system is started to collect melt data, and the optimal screw parameters are output in real time by a multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model. Based on the glass fiber length distribution index and melt degradation index output by the online near-infrared Raman spectroscopy system, the screw speed, temperature of each section and residence time are adjusted in a multivariate coordinated manner using a model predictive control algorithm; After the extruded melt passes through the die, the Kalman filter soft measurement module is used to estimate the melt molecular weight in real time. When the estimated melt molecular weight is lower than the set threshold, vacuum exhaust treatment is added and the drying time is increased. The composite granules obtained by extrusion granulation were injection molded to prepare standard specimens, which were then used to improve the Reduced Strain Closure and machine learning correction model to predict the three-dimensional orientation tensor of glass fiber in the full flow field of injection molding. The holding pressure and injection rate were adjusted based on the prediction results. Impact strength and elongation at break of the spline are tested. If the test value is lower than the target threshold, the test data is sent back to the multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model for incremental training, and then returned to the aforementioned extrusion step to readjust the process parameters.

2. The method for preparing the high-toughness glass fiber reinforced polyamide material according to claim 1, characterized in that, The inverse algorithm for interfacial coupling reaction kinetics, which couples Fick diffusion with the Arrhenius reaction rate equation, specifically models the adsorption process of silane coupling agent on the glass fiber surface as a set of coupled partial differential equations of two-dimensional Fick diffusion equation and Arrhenius reaction rate equation. Using measured adsorption data as constraints, the diffusion coefficient, activation energy, and pre-exponential factor are estimated inversely using the quasi-Newton method.

3. The method for preparing the high-toughness glass fiber reinforced polyamide material according to claim 2, characterized in that, The convergence criterion for the inverse estimation is that the Frobenius norm of the parameter change between two adjacent iterations is lower than the convergence threshold, and the reference value for the convergence threshold is [value missing]. .

4. The method for preparing the high-toughness glass fiber reinforced polyamide material according to claim 3, characterized in that, The measured adsorption data were obtained by taking samples at multiple coating time points under multiple temperature gradients. The number and range of each temperature gradient and coating time point constituted the experimental design matrix. The mean of the parameter when the ratio of the standard deviation to the mean of the parameter estimation result was lower than the stability threshold was used as the reference value. The reference value of the stability threshold was 5%.

5. The method for preparing the high-toughness glass fiber reinforced polyamide material according to claim 4, characterized in that, The multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model represents each functional segment of the twin-screw extruder with a graph structure. The bottom layer is a temporal convolutional network module, the middle layer is a GraphSAGE variant graph convolutional module, and the top layer is a physical constraint layer. The three layers are fused by residual connections.

6. The method for preparing the high-toughness glass fiber reinforced polyamide material according to claim 5, characterized in that, The receptive field width of the temporal convolutional network module is dynamically determined based on the screw rotation speed parameter. When the screw rotation speed is higher than the rotation speed threshold, the receptive field width takes the first width value, and when it is lower than or equal to the rotation speed threshold, the receptive field width takes the second width value. The reference value of the rotation speed threshold is 150 r / min, the reference value of the first width value is 32 time steps, and the reference value of the second width value is 16 time steps.

7. The method for preparing the high-toughness glass fiber reinforced polyamide material according to claim 6, characterized in that, The GraphSAGE variant graph convolution module employs a message-passing iterative mechanism, using the hidden state vectors of each node... Norm change is used as a convergence criterion; when all nodes... Iteration stops when all changes are below the iteration convergence threshold. The reference value for the iteration convergence threshold is [value missing]. The reference value for the maximum number of iterations is 6.

8. The method for preparing the high-toughness glass fiber reinforced polyamide material according to claim 7, characterized in that, The physical constraint layer embeds the analytical expressions of the Carreau viscosity equation and the Weibull fiber fracture probability model into the network in a differentiable form. The output of the physical layer serves as the residual basis, and the output of the data-driven layer serves as the superposition of correction terms.

9. The method for preparing the high-toughness glass fiber reinforced polyamide material according to claim 8, characterized in that, The multi-scale spatiotemporal graph convolution and physical embedding fusion prediction model supports conditional generation functionality. Given the target glass fiber length distribution, it obtains parameter suggestions for the optimal screw speed and temperature curve by backpropagating to the input layer.

10. The method for preparing the high-toughness glass fiber reinforced polyamide material according to claim 9, characterized in that, The Kalman filter soft measurement module uses online near-infrared spectral signals as observations and the polyamide hydrolysis degradation kinetic equation as the equation of state to recursively estimate the molecular weight of the melt. The molecular weight threshold for initiating vacuum degassing is determined by comparing offline end-group titration data with the estimated value.

Images