AI parameter optimization control method and system in melt extrusion process of regenerated polyester

Abstract

The invention discloses an AI parameter optimization control method and system in a melt extrusion process of regenerated polyester, and relates to the technical field of material processing. Comprising the following steps: S1, carrying out data fusion on the spectral data, the imaging data and the process parameter time sequence data to obtain fusion features; s2, constructing a hybrid model through a physical information neural network and a meta-learning hybrid model, and determining initial process parameters according to the fusion features; and S3, a reward function is constructed through the fiber strength, the unit energy consumption and the equipment wear rate, the initial process parameters are corrected according to the reward function, and the corrected process parameters are obtained. According to the method, high-dimensional feature expression is constructed through three-dimensional space-time-spectrum tensor fusion, so that the isomerism of raw material characteristics can be covered, meanwhile, space-spectrum relevance can be dynamically captured through node edge weights, and then the prediction precision of impurity distribution and melt flow is improved.

Description

Technical Field

[0001] The present invention relates to the field of material processing technology, and in particular to an AI parameter optimization control method and system for a recycled polyester melt extrusion process. Background Art

[0002] Recycled polyester refers to fiber material produced from recycled polyethylene terephthalate (PET) materials (such as used bottles and textile scraps) after physical or chemical treatment. With growing environmental awareness and the increasing demand for resource recycling, recycled polyester is increasingly used in the textile industry. While it possesses similar physical properties to virgin polyester, its production process presents challenges such as complex raw material sources, uneven molecular structure, and fluctuating impurity levels, placing higher demands on processing technology.

[0003] Melt extrusion is a key process before spinning recycled polyester. The main process includes: drying → melt plasticization → filtration → extrusion molding. Its core goal is to uniformly heat the recycled PET pellets to a molten state. Through screw conveying, mixing, and pressurization, the melt enters the subsequent spinning assembly at a stable pressure and temperature.

[0004] Chinese invention patent publication number CN118966475A discloses an energy consumption optimization system for polyester fiber production, comprising: a spinning task acquisition module for obtaining fiber production spinning tasks; a spinning control feature association module for associating control features with melt spinning equipment; a spinning control decision module for making control parameter decisions based on the spinning control feature association space; a spinning control expectation optimization module for performing multi-feature expectation coupling optimization in the spinning control decision space; a spinning control energy consumption optimization module for performing multi-dimensional energy consumption optimization in the spinning control expectation optimization space; and a spinning module for spinning. This system addresses the technical problem of existing polyester fiber production energy consumption optimization systems, which excessively pursue energy reduction while neglecting stability control during the production process, resulting in poor overall production performance. Instead, it achieves the technical effect of achieving the coordinated optimization of energy consumption and other key factors, ensuring overall performance improvement in the polyester fiber production process.

[0005] However, because waste textiles, bottle flakes, and industrial waste can all serve as sources of recycled polyester raw materials, their physicochemical properties (such as molecular weight distribution, impurity types and content, and crystallinity) exhibit high-dimensional heterogeneity. Consequently, during the melt extrusion process of recycled polyester, parameters such as temperature, pressure, and viscosity exhibit strong nonlinear coupling relationships with the raw material properties. During the pilot production phase of new raw materials, only extremely sparse pilot production data is available, making it difficult to capture the multidimensional mapping space between raw material properties and process responses. Consequently, traditional data-driven AI models cannot accurately model the impact of raw material differences on melt extrusion dynamics, resulting in error rates in predicted parameters exceeding process tolerances. Summary of the Invention

[0006] The purpose of the present invention is to provide an AI parameter optimization control method and system for the recycled polyester melt extrusion process to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an AI parameter optimization control method for a regenerated polyester melt extrusion process, comprising: S1: Acquiring multimodal data: acquiring spectral data, imaging data, and process parameter time series data, and fusing the spectral data, imaging data, and process parameter time series data to obtain fusion features; S2: Constructing a hybrid model: Constructing a hybrid model through a physical information neural network and a meta-learning hybrid model, and determining initial process parameters based on the fusion features, including: S2.1: Obtaining predicted melt flow rate: Using the physical information neural network and fusion features, a momentum conservation equation is constructed to obtain the predicted melt flow rate. The momentum conservation equation is specifically: ; in: is the density of the melt, is the predicted pressure gradient value at node m, is the dynamic viscosity of the melt, is the Laplace operator of the velocity field, is the impurity resistance term, is the sampling time; S2.2: Constructing a loss function: Obtaining data matching terms, PDE residual terms, and safety penalty terms through the melt prediction flow rate, momentum conservation equation, and real-time melt temperature, and determining a total loss function; S2.3: Determine initial process parameters: Based on the total loss function and the meta-learning hybrid model, construct a hybrid model, and use the existing process parameters of the raw material as input to the hybrid model, and output the corresponding initial process parameters; S3: Process parameter correction: A reward function is constructed based on fiber strength, unit energy consumption, and equipment wear rate, and the initial process parameters are corrected according to the reward function to obtain corrected process parameters.

[0008] Furthermore, the fusion features are obtained, including: S1.2.1: Construct a three-dimensional space-time-spectral tensor: Stack the spectral data in chronological order and form a four-dimensional tensor based on the spectral band, spatial coordinates, and time size of the imaging data. Normalize the four-dimensional tensor to obtain the normalized reflectance value, specifically: ; in: is the normalized reflectance value of the j-th pixel at wavelength λ, is the original reflectance value of the j-th pixel at wavelength λ, is the minimum pixel reflectance value at wavelength λ, is the maximum pixel reflectance value at wavelength λ; S1.2.2: Graph Neural Network Modeling: Based on the normalized reflectivity values, set the nodes in the graph neural network model to obtain the edge weights between different nodes, specifically: ; in: is the edge weight between node m and node n, is the base of natural logarithms, is the normalized reflectivity value of node m, is the normalized reflectivity value of node n, is the spatial coordinate of node m, is the spatial coordinate of node n, is the spatial scale parameter; S1.2.3: Obtain the predicted pressure gradient value: Determine the attention coefficient of the graph neural network model based on the edge weight, and update the feature vector of the node in the graph neural network model based on the attention coefficient to obtain the updated feature vector.

[0009] Furthermore, graph neural network modeling is performed based on the edge weights between different nodes, including: S1.2.2.1: Set model nodes: Divide the surface of the polyethylene raw material into a grid, and sort the nodes in the grid according to the divided grid, and determine the number of each node, specifically: ; in: is the node number, is the row coordinate of the node in the grid space, is the total number of columns in the grid space, is the column coordinate of the node in the grid space; S1.2.2.2: Obtain spatial distance item: Determine the corresponding spatial distance and spatial weight attenuation based on the node number and coordinates, specifically: ; in: is the spatial coordinate of node m, is the spatial coordinate of node n, is the row coordinate corresponding to node m, is the row coordinate corresponding to node n, is the column coordinate corresponding to node m, is the column coordinate corresponding to node n, is the spatial weight decay, is the base of natural logarithms, is the spatial scale parameter; S1.2.2.3: Obtain spectral similarity: Determine the spectral similarity between different nodes based on the normalized reflectance value of each node, specifically: ; in: is the cosine similarity, is the normalized reflectivity value of node m, is the normalized reflectivity value of node n; S1.2.2.4: Update node features: Determine the edge weights between different nodes based on the spatial weight decay and the spectral similarity between different nodes, construct an edge weight matrix, and update the node feature matrix of each layer in the graph neural network model based on the edge weight matrix and the node feature matrix. Specifically: ; in: is the node feature matrix of the l+1th layer, is a nonlinear activation function, is the degree matrix, is the edge weight matrix, is the node feature matrix of the lth layer, is the learnable weight matrix of layer l.

[0010] Furthermore, the updated feature vector is obtained, including: S1.2.3.1: Determine the attention coefficient: Based on the edge weights between different nodes, determine the attention coefficient between different nodes, specifically: ; in: is the attention coefficient between node m and node n, is the edge weight between node m and node n, is the neighborhood set of node m, is the edge weight between node m and node f, 、 、 is the node index; S1.2.3.2: Obtain updated feature vectors: Update the feature vector at each node based on the attention coefficient, specifically: ; in: is the updated normalized reflectivity value of node m, is a nonlinear activation function, is the learnable weight matrix, is the normalized reflectivity value of node n, is the neighborhood set of node m, is the node index, is the attention coefficient between node m and node n; S1.2.3.3: Determine the predicted pressure gradient value: Based on the updated eigenvector, obtain the predicted pressure gradient value corresponding to the node, specifically: ; in: is the predicted pressure gradient value at node m, is the learnable weight vector, is the updated normalized reflectivity value of node m, is the bias term.

[0011] Furthermore, the total loss function is determined, including: S2.2.1: Determine data matching items: Based on the predicted melt flow rate and the real-time melt flow rate, determine the data matching items, specifically: ; in: For data matching, is the total number of samples, The predicted flow rate of the melt for the vth sample, is the actual melt flow rate of the vth sample; S2.2.2: Determine the PDE residual term: According to the momentum conservation equation, determine the residual corresponding to each sampling point in the flow field and obtain the PDE residual term, specifically: ; in: is the PDE residual term, is the total number of nodes, is the density of the melt, is the predicted pressure gradient value at node m, is the dynamic viscosity of the melt, is the sampling time, is the Laplace operator of the velocity field at node m, is the impurity resistance term at node m; S2.2.3: Determine safety penalty items: Determine safety penalty items based on the real-time melt temperature and temperature safety threshold, specifically: ; in: For security penalties, is the real-time melt temperature, is the temperature safety threshold; S2.2.4: Determine the total loss function. Based on the data matching term, the PDE residual term, and the security penalty term, determine the total loss function, specifically: ; in: is the total loss function, For data matching, is the PDE residual term, 、 is the weight coefficient, It is a safety penalty item.

[0012] Furthermore, the corresponding initial process parameters are output, including: S2.3.1: Obtaining initial model parameters: The existing process parameters of the raw material are used as the input of the hybrid model, and the initial dynamic viscosity of the raw material at different melt temperatures and the initial learnable coefficient in the impurity resistance term are obtained as output. The specific formula for obtaining the initial dynamic viscosity is: ; in: The real-time melt temperature of the melt The dynamic viscosity under is the base viscosity of the melt at the reference temperature, is the base of natural logarithms, is the viscosity-temperature sensitivity coefficient, is the real-time melt temperature, is the reference temperature; S2.3.2: Initial model parameter update: Based on the total loss function, the initial dynamic viscosity and initial learnable coefficient are updated to obtain updated model parameters, specifically: ; in: are the updated model parameters, are the initial model parameters, is the learning rate, is the total loss function, is the total loss function Initial model parameters gradient; S2.3.3: Obtaining initial process parameters: updating the parameters of the hybrid model according to the updated model parameters, and obtaining corresponding initial process parameters according to the hybrid model after parameter update.

[0013] Furthermore, the corrected process parameters are obtained, including: S3.1: Obtain current reward value: Based on the fiber strength, unit energy consumption, and equipment wear rate during the melt extrusion process, a reward function is constructed to obtain a comprehensive reward value, specifically: ; in: is the comprehensive reward value, is the fiber strength weight coefficient, is the fiber strength index, is the unit energy consumption weight coefficient, is the energy consumption per unit product, is the equipment wear rate weight coefficient, is the equipment wear rate; S3.2: Determine the optimization strategy: Compare the comprehensive reward value with the preset reward threshold, and determine the final process parameters based on the comparison result, specifically: When the comprehensive reward value is less than the preset reward threshold, the initial process parameters are the final process parameters. Otherwise, the initial process parameters are adjusted, and steps S3.1 and S3.2 are repeated until the comprehensive reward value corresponding to the adjusted process parameters is less than the preset reward threshold.

[0014] Furthermore, the initial process parameters are adjusted by adjusting the formula to obtain the adjusted process parameters, specifically: ; in: is the melt temperature, is the system pressure, is the lower temperature threshold, is the upper pressure threshold, is the temperature-pressure compensation coefficient, is the pressure-temperature compensation coefficient, is the lower pressure threshold.

[0015] An AI parameter optimization control system for a regenerated polyester melt extrusion process uses any of the above-mentioned AI parameter optimization control methods for a regenerated polyester melt extrusion process.

[0016] Compared with the prior art, the present invention has the following beneficial effects: First, the present invention constructs a high-dimensional feature representation by fusing spectral data, imaging data, and process time series data into a three-dimensional spatiotemporal tensor, thereby covering the heterogeneity of raw material properties. Furthermore, through node-edge weights, it can dynamically capture spatial-spectral correlations, thereby improving the prediction accuracy of impurity distribution and melt flow. Second, the present invention fuses the momentum conservation equation and the PDE residual term to determine the total loss function of the physical information neural network. This is then combined with a meta-learning hybrid model to construct a hybrid model that conforms to the laws of fluid mechanics, thereby reducing dependence on sample size. Furthermore, the model parameters can be initialized based on a small number of samples, and the update strategy can be used to quickly adapt to new raw materials, thereby solving the generalization problem under sparse data. Third: The present invention dynamically balances quality, energy consumption and equipment life through a comprehensive reward value, and combines it with a process parameter adjustment formula to ensure that process parameters are within a safe threshold while avoiding equipment damage or energy consumption surges caused by excessive temperature / pressure. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the flow of the AI parameter optimization control method of the present invention; Figure 2 This is a slice diagram of the three-dimensional space-time-spectral tensor at 1 minute in the present invention; Figure 3 This is a slice diagram of the three-dimensional space-time-spectral tensor at 5 minutes in the present invention; Figure 4 It is a comparison chart of production indicators in the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Because recycled polyester raw materials, such as waste textiles, bottle flakes, and industrial waste, can all serve as sources, their physicochemical properties (such as molecular weight distribution, impurity types and content, and crystallinity) exhibit high-dimensional heterogeneity. Consequently, during the melt extrusion process of recycled polyester, parameters such as temperature, pressure, and viscosity exhibit strong nonlinear coupling relationships with the raw material properties. However, during the pilot production phase for new raw materials, only extremely sparse pilot production data is available, making it difficult to capture the multidimensional mapping space between raw material properties and process responses. Consequently, traditional data-driven AI models cannot accurately model the impact of raw material differences on melt extrusion dynamics, leading to error rates in predicted parameters exceeding process tolerances. Example 1

[0020] refer to Figure 1-Figure 4 This embodiment provides an AI parameter optimization control method for a recycled polyester melt extrusion process, the AI parameter optimization control method comprising the following steps: Step S1: Obtain fusion features. This involves aligning the spectral data, imaging data, and process parameter time series data, and then performing data fusion to obtain a 3D spatial-spectral-time series fusion feature. The details are as follows: Step S1.1: Acquire multimodal data. This involves using a near-infrared spectrometer to examine a polyethylene raw material with a known molecular weight distribution, obtaining corresponding near-infrared spectral data. A hyperspectral imager is then used to examine the surface of the polyethylene raw material, obtaining corresponding hyperspectral imaging data.

[0021] Furthermore, temperature sensors are installed at the inlet and outlet of the extruder barrel and die, a pressure sensor is installed at the die inlet, and a torque sensor is installed on the drive shaft between the drive motor and the reduction gearbox to obtain corresponding temperature, pressure, and torque data. Specifically, based on the obtained temperature, pressure, and torque data, the corresponding temperature uniformity index, melt viscosity, and torque fluctuation range are obtained.

[0022] During the specific implementation, five temperature zones were divided along the axial direction of the extruder barrel, and a thermocouple was set in each temperature zone, with the spacing between two adjacent temperature sensors set to 200nm. At the same time, a thermocouple was set at each inlet and outlet of the die to monitor the melt flow temperature gradient. Specifically, the temperature of each corresponding zone was obtained through the set thermocouples, and the corresponding temperature uniformity index was obtained based on the temperature of each zone, which was specifically: ; in: is the temperature uniformity index, is the maximum temperature data, This is the minimum temperature data.

[0023] In this embodiment, the seven temperature data obtained are: 212° C., 214° C., 215° C., 217° C., 218° C., 213° C. and 210° C., and the corresponding temperature uniformity index is 8° C.

[0024] Furthermore, a piezoresistive sensor is set at the inlet of the die head to obtain corresponding pressure data, and the average pressure is determined based on the obtained pressure data. At the same time, the corresponding melt viscosity is determined based on the average pressure, specifically: ; in: is the melt viscosity, is the radius of the die head, is the average pressure, is the volume of melt passing through the die per unit time, is the length of the die head.

[0025] In this embodiment, the pressure data fluctuates in the range of 12.2-12.8 MPa, and the corresponding average pressure is 12.5 MPa. At the same time, the radius of the die is 5 mm, the length of the die is 50 mm, and the volume of the melt passing through the die per unit time is 10 cm 3 / s, the corresponding melt viscosity is 1200Pa / cdotps.

[0026] Furthermore, a torque flange sensor is provided on the transmission shaft between the drive motor and the reduction box to obtain the corresponding instantaneous value of the torque and determine the corresponding torque fluctuation range.

[0027] Step S1.2: Feature fusion. The near-infrared spectral data, hyperspectral imaging data, temperature data, pressure data, and torque data obtained in step S1.1 are fused through the graph neural network model to obtain the predicted pressure gradient value. The details are as follows: Step S1.2.1: Construct a three-dimensional spatiotemporal-spectral tensor. This involves combining the near-infrared spectral data and hyperspectral imaging data obtained in step S1.1 with the corresponding time series to construct a spatiotemporal-spectral tensor. Specifically, the hyperspectral imaging data is stacked in chronological order and combined with the spectral bands in the near-infrared spectral data. This is done by following the order of spatial row coordinates, spatial column coordinates, spectral bands, and time to form a corresponding four-dimensional tensor.

[0028] refer to Figure 2 and Figure 3 , Figure 2 This is a slice diagram of the three-dimensional space-time-spectral tensor at 1 minute in this embodiment, Figure 2 is a slice diagram of the three-dimensional space-time-spectral tensor at 5 minutes in this embodiment, Figure 2 and Figure 3 It can be seen that within the spatial coordinate region (10:20, 25:30), all wavelengths and time slices display higher normalized reflectivity values, indicating the presence of a highly reflective impurity or special material in this region, with a reflectivity approximately 60% higher than that of other regions. Furthermore, the reflectivity distribution and intensity of this highly reflective region are essentially identical in both the 1-minute and 5-minute slice images, indicating that the reflectivity characteristics of this impurity remain stable over the 5-minute period, with no observed time-dependent decay or diffusion.

[0029] Furthermore, according to the obtained four-dimensional tensor, each band is normalized to determine the corresponding normalized reflectance value, specifically: ; in: is the normalized reflectance value of the j-th pixel at wavelength λ, is the original reflectance value of the j-th pixel at wavelength λ, is the minimum pixel reflectance value at wavelength λ, is the maximum pixel reflectance value at wavelength λ.

[0030] Step S1.2.2: Graph neural network model building. That is, according to the normalized reflectivity values obtained in step S1.2.1, the nodes in the graph neural network model are set, and based on the set nodes, the edge weights between different nodes are obtained, specifically: ; in: is the edge weight between node m and node n, is the base of natural logarithms, is the normalized reflectivity value of node m, is the normalized reflectivity value of node n, is the spatial coordinate of node m, is the spatial coordinate of node n, is the spatial scale parameter.

[0031] Step S1.2.3: Obtain the predicted pressure gradient value. That is, based on the edge weights between different nodes obtained in step S1.2.2, determine the attention coefficient of the graph neural network model, and based on the determined attention coefficient, update the feature vector at each node to obtain the updated feature vector corresponding to each node. The details are as follows: Step S1.2.3.1: Determine the attention coefficient. That is, based on the edge weights between different nodes obtained in step S1.2.2, determine the attention coefficient between different nodes, specifically: ; in: is the attention coefficient between node m and node n, is the edge weight between node m and node n, is the neighborhood set of node m, is the edge weight between node m and node f, 、 、 is the node index.

[0032] In the specific implementation process, the neighborhood of the node at coordinate (10,25) includes the node at coordinate (10,26) and the node at coordinate (30,40), where the edge weight between the node at coordinate (10,25) and the node at coordinate (10,26) is 0.96, and the edge weight between the node at coordinate (10,25) and the node at coordinate (30,40) is 0.02. The attention coefficient between the node at coordinate (10,25) and the node at coordinate (10,26) is 0.72.

[0033] Step S1.2.3.2: Obtain the updated feature vector. That is, based on the attention coefficients between different nodes determined in step S1.2.3.1, update the feature vector at each node, specifically: ; in: is the updated normalized reflectivity value of node m, is a nonlinear activation function, is the learnable weight matrix, is the normalized reflectivity value of node n, is the neighborhood set of node m, is the node index, is the attention coefficient between node m and node n.

[0034] Step S1.2.3.3: Determine the predicted pressure gradient value. That is, based on the updated feature vector at the node obtained in step S1.2.3.2, convert it into the predicted pressure gradient value corresponding to the node, specifically: ; in: is the predicted pressure gradient value at node m, is the learnable weight vector, is the updated normalized reflectivity value of node m, is the bias term.

[0035] Step S2: Construct a hybrid model. That is, based on the predicted pressure gradient value determined in step S1.2.3.3, combine the physical information neural network and the meta-learning hybrid model to construct a hybrid model and determine the corresponding initial process parameters. The details are as follows: Step S2.1: Obtaining the predicted melt flow rate. The predicted pressure gradient value determined in step S1.2.3.3 and the multimodal data obtained in step S1.1 are used as inputs to the physical information neural network. Combined with the constructed momentum conservation equation, the corresponding predicted melt flow rate is output.

[0036] Furthermore, according to the predicted pressure gradient value determined in step S1.2.3.3, the corresponding impurity resistance term is obtained, specifically: ; in: is the impurity resistance term, is the learnable coefficient, is the predicted pressure gradient value at node m.

[0037] Furthermore, based on the obtained impurity resistance term, the corresponding momentum conservation equation is constructed, specifically: ; in: is the density of the melt, is the predicted pressure gradient value at node m, is the dynamic viscosity of the melt, is the Laplace operator of the velocity field, is the impurity resistance term, is the sampling time.

[0038] Step S2.2: Construct a loss function. Specifically, based on the predicted melt velocity obtained in step S2.1, the constructed momentum conservation equation, and the real-time melt temperature, the data matching term, PDE residual term, and safety penalty term are determined. Based on the determined data matching term, PDE residual term, and safety penalty term, the corresponding total loss function is obtained. Specifically, the function is as follows: Step S2.2.1: Determine data matching items. That is, based on the predicted melt flow rate obtained in step S2.1 and the melt flow rate obtained by actual measurement, obtain data matching items, specifically: ; in: For data matching, is the total number of samples, The predicted flow rate of the melt for the vth sample, is the actual melt flow rate of the vth sample.

[0039] Step S2.2.2: Determine the PDE residual term. That is, based on the momentum conservation equation constructed in step S2.1, determine the residual corresponding to each sampling point in the flow field. At the same time, based on the residual corresponding to each sampling point, obtain the PDE residual term, specifically: ; in: is the PDE residual term, is the total number of nodes, is the density of the melt, is the predicted pressure gradient value at node m, is the dynamic viscosity of the melt, is the sampling time, is the Laplace operator of the velocity field at node m, is the impurity resistance term at node m.

[0040] Step S2.2.3: Determine the safety penalty item. That is, based on the real-time melt temperature and the set temperature safety threshold, obtain the safety penalty item, specifically: ; in: For security penalties, is the real-time melt temperature, is the temperature safety threshold.

[0041] Step S2.2.4: Determine the total loss function. That is, based on the data matching term determined in step S2.2.1, the PDE residual term determined in step S2.2.2, and the security penalty term determined in step S2.2.3, obtain the corresponding total loss function, specifically: ; in: is the total loss function, For data matching, is the PDE residual term, 、 is the weight coefficient, It is a safety penalty item.

[0042] Step S2.3: Determine the initial process parameters. This involves combining the total loss function determined in step S2.2.4 with the meta-learning hybrid model to obtain a hybrid model. The melt temperature, pressure gradient, and melt flow rate corresponding to each raw material are used as inputs to the hybrid model, and the corresponding initial process parameters are output. The details are as follows: Step S2.3.1: Obtain initial model parameters. The melt temperature, pressure gradient, and melt flow rate corresponding to each raw material are used as inputs to the meta-learning hybrid model. The output is the model parameters corresponding to each raw material, namely, the initial dynamic viscosity and initial learnable coefficient of the impurity resistance term for each raw material at different melt temperatures.

[0043] In this embodiment, the formula for obtaining the initial dynamic viscosity of the raw material at different melt temperatures is specifically: ; in: The real-time melt temperature of the melt The dynamic viscosity under is the base viscosity of the melt at the reference temperature, is the base of natural logarithms, is the viscosity-temperature sensitivity coefficient, is the real-time melt temperature, is the reference temperature.

[0044] During the specific implementation, the raw material data corresponding to 100 groups of bottle flakes are: melt temperature range [200, 220] °C, pressure gradient range [0.1, 0.3] MPa / mm, flow rate range [8, 12] m / s. At the same time, the base viscosity of the melt at the reference temperature is 500 Pa\cdotps, and the viscosity-temperature sensitivity coefficient is set to 0.02K. -1 , then at a temperature of 215°C, the dynamic viscosity of the bottle flakes at 215°C is 300Pa\cdotps, and the learnability coefficient of the bottle flakes is 0.28.

[0045] Step S2.3.2: Update the initial model parameters. That is, update the initial model parameters obtained in step S2.3.1 using the total loss function determined in step S2.2.4 to obtain the corresponding updated model parameters, specifically: ; in: are the updated model parameters, are the initial model parameters, is the learning rate, is the total loss function, is the total loss function Initial model parameters gradient.

[0046] In this specific implementation, the data matching term is 0.11, the PDE residual term is 0.15, and the safety penalty term is 0, resulting in a total loss function of 0.125. Furthermore, the learning rate in this embodiment is set to 0.01. The learnable coefficient of the bottle flake obtained in step S2.3.1 is 0.28, and the corresponding updated learnable coefficient is 0.255. Furthermore, if the dynamic viscosity of the bottle flake obtained in step S2.3.1 at 215°C is 300 Pa / cdotps, the corresponding updated dynamic viscosity is 301.5.

[0047] Step S2.3.3: Obtaining initial process parameters. That is, based on the updated model parameters obtained in step S2.3.2, the constructed hybrid model is updated, and based on the updated hybrid model, the corresponding initial process parameters are obtained, including melt temperature and extrusion pressure.

[0048] Step S3: Process parameter correction. This involves constructing a reward function based on fiber strength, unit energy consumption, and equipment wear rate. Based on this reward function, the initial process parameters obtained in step S2.3.3 are corrected to obtain the corrected process parameters. The details are as follows: Step S3.1: Obtain the current reward value. That is, during the melt extrusion process, the corresponding fiber strength, unit energy consumption, and equipment wear rate are obtained in real time. At the same time, a reward function is constructed based on the obtained fiber strength, unit energy consumption, and equipment wear rate. Specifically, it is: ; in: is the comprehensive reward value, is the fiber strength weight coefficient, is the fiber strength index, is the unit energy consumption weight coefficient, is the energy consumption per unit product, is the equipment wear rate weight coefficient, is the equipment wear rate.

[0049] Furthermore, based on the initial process parameters obtained in step S2.3.3, the polyethylene raw material is melt-extruded. During the melt-extrusion process, the fiber strength of the melt-extruded recycled polyester is tested using an online laser stretching instrument. Energy consumption during the melt-extrusion process is monitored using a smart meter, and equipment wear rate is determined through vibration spectrum analysis and screw usage time. The current reward value is determined based on the real-time fiber strength, unit energy consumption, and equipment wear rate.

[0050] Step S3.2: Determine the optimization strategy. This involves comparing the current reward value obtained in step S3.1 with the preset reward threshold, and determining the final process parameters based on the comparison results. Specifically, If the current reward value is less than the preset reward threshold, the initial process parameters obtained in step S2.3.3 become the final process parameters. Conversely, if the current reward value is not less than the preset reward threshold, the initial process parameters obtained in step S2.3.3 are adjusted to obtain the corresponding adjusted process parameters. Steps S3.1 and S3.2 are repeated until the current reward value corresponding to the adjusted process parameters is less than the preset reward threshold.

[0051] In this embodiment, the initial process parameters are adjusted by an adjustment formula, and the adjustment formula is specifically: ; in: is the melt temperature, is the system pressure, is the lower temperature threshold, is the upper pressure threshold, is the temperature-pressure compensation coefficient, is the pressure-temperature compensation coefficient, is the lower pressure threshold.

[0052] refer to Figure 4 , Figure 4 is a comparison chart of production indicators in this embodiment, Figure 4 It can be seen that: the fiber strength increased from 80% to 95%, that is, the material performance was significantly enhanced after optimization; the unit energy consumption was reduced from 100% to 75%, that is, the energy utilization efficiency in the production process was improved; the equipment wear rate was greatly reduced from 60% to 30%, that is, the durability of the equipment or material was significantly improved.

[0053] This embodiment also provides an AI parameter optimization control system for a regenerated polyester melt extrusion process, which uses the above-mentioned AI parameter optimization control method for a regenerated polyester melt extrusion process. Example 2

[0054] This embodiment provides an AI parameter optimization control method for the regenerated polyester melt extrusion process. Its specific implementation method is the same as that of Example 1, except that in step S1.2.2, the normalized reflectivity value obtained in step S1.2.1 is used to set the nodes in the graph neural network model, and based on the set nodes, the edge weights between different nodes are obtained to construct an edge weight matrix. At the same time, graph neural network modeling is performed based on the edge weight matrix and the node feature matrix. The present invention is illustrated below with reference to the specific implementation methods of this embodiment.

[0055] In this embodiment, graph neural network modeling is performed based on the edge weight matrix and the node feature matrix, as follows: Step S1.2.2.1: Set model nodes. The surface of the polyethylene material is divided into grids, with each grid corresponding to a pixel, and each pixel is considered a node. Specifically, the nodes are sorted in order according to the grid divisions, and the number of each node is determined as follows: ; in: is the node number, is the row coordinate of the node in the grid space, is the total number of columns in the grid space, is the column coordinate of the node in the grid space.

[0056] In the specific implementation process, in the constructed 50*50 grid space, the node number corresponding to the coordinate (10,25) is: 10*50+25=525. That is, the node number corresponding to the coordinate (10,25) is 525.

[0057] Furthermore, for each node, the corresponding normalized reflectivity values are arranged according to the wavelength order to form a corresponding multi-dimensional feature vector.

[0058] Step S1.2.2.2: Obtain the spatial distance item. That is, according to the node number and node coordinates determined in step S1.2.2.1, determine the corresponding spatial distance and spatial weight attenuation, specifically: ; in: is the spatial coordinate of node m, is the spatial coordinate of node n, is the row coordinate corresponding to node m, is the row coordinate corresponding to node n, is the column coordinate corresponding to node m, is the column coordinate corresponding to node n, is the spatial weight decay, is the base of natural logarithms, is the spatial scale parameter.

[0059] In the specific implementation process, in this embodiment, the scale parameter for controlling the influence of spatial proximity is 5, and the spatial distance between the node with coordinates (10, 25) and the node with coordinates (10, 26) is 1, then the corresponding spatial weight attenuation is 0.98.

[0060] Step S1.2.2.3: Obtain spectral similarity. That is, based on the normalized reflectance value obtained in step S1.2.1, determine the normalized reflectance value corresponding to each node, and based on the normalized reflectance value corresponding to each node, determine the spectral similarity between different nodes, specifically: ; in: is the cosine similarity, is the normalized reflectivity value of node m, is the normalized reflectivity value of node n.

[0061] Step S1.2.2.4: Update node features. That is, based on the spatial weight decay obtained in step S1.2.2.2 and the spectral similarity obtained in step S1.2.2.3, determine the edge weights between different nodes. Furthermore, based on the edge weights between all the determined nodes, construct an edge weight matrix. Based on the three-dimensional spatiotemporal-spectral tensor corresponding to each node, construct a node feature matrix. At the same time, based on the constructed edge weight matrix and the initial node feature matrix, update the node feature matrix of each layer in the graph neural network model, specifically: ; in: is the node feature matrix of the l+1th layer, is a nonlinear activation function, is the degree matrix, is the edge weight matrix, is the node feature matrix of the lth layer, is the learnable weight matrix of layer l.

[0062] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is limited by the accompanying embodiments and their equivalents.

Claims

1. An AI parameter optimization control method for a recycled polyester melt extrusion process, characterized in that: Includes: S1: Acquiring multimodal data: acquiring spectral data, imaging data, and process parameter time series data, and fusing the spectral data, imaging data, and process parameter time series data to obtain fusion features; S2: Constructing a hybrid model: Constructing a hybrid model through a physical information neural network and a meta-learning hybrid model, and determining initial process parameters based on the fusion features, including: S2.1: Obtaining predicted melt flow rate: Using the physical information neural network and fusion features, a momentum conservation equation is constructed to obtain the predicted melt flow rate. The momentum conservation equation is specifically: ； in: is the density of the melt, is the predicted pressure gradient value at node m, is the dynamic viscosity of the melt, is the Laplace operator of the velocity field, is the impurity resistance term, is the sampling time; S2.2: Constructing a loss function: Obtaining data matching terms, PDE residual terms, and safety penalty terms through the melt prediction flow rate, momentum conservation equation, and real-time melt temperature, and determining a total loss function; S2.3: Determine initial process parameters: Based on the total loss function and the meta-learning hybrid model, construct a hybrid model, and use the existing process parameters of the raw material as input to the hybrid model, and output the corresponding initial process parameters; S3: Process parameter correction: A reward function is constructed based on fiber strength, unit energy consumption, and equipment wear rate, and the initial process parameters are corrected according to the reward function to obtain corrected process parameters.

2. The AI parameter optimization control method for the regenerated polyester melt extrusion process according to claim 1, characterized in that: The fusion features obtained include: S1.2.1: Construct a three-dimensional space-time-spectral tensor: Stack the spectral data in chronological order and form a four-dimensional tensor based on the spectral band, spatial coordinates, and time size of the imaging data. Normalize the four-dimensional tensor to obtain the normalized reflectance value, specifically: ； in: is the normalized reflectance value of the j-th pixel at wavelength λ, is the original reflectance value of the j-th pixel at wavelength λ, is the minimum pixel reflectance value at wavelength λ, is the maximum pixel reflectance value at wavelength λ; S1.2.2: Graph Neural Network Modeling: Based on the normalized reflectivity values, set the nodes in the graph neural network model to obtain the edge weights between different nodes, specifically: ； in: is the edge weight between node m and node n, is the base of natural logarithms, is the normalized reflectivity value of node m, is the normalized reflectivity value of node n, is the spatial coordinate of node m, is the spatial coordinate of node n, is the spatial scale parameter; S1.2.3: Obtain the predicted pressure gradient value: Determine the attention coefficient of the graph neural network model based on the edge weight, and update the feature vector of the node in the graph neural network model based on the attention coefficient to obtain the updated feature vector.

3. The AI parameter optimization control method for the regenerated polyester melt extrusion process according to claim 2, characterized in that: Based on the edge weights between different nodes, graph neural network modeling is performed, including: S1.2.2.1: Set model nodes: Divide the surface of the polyethylene raw material into a grid, and sort the nodes in the grid according to the divided grid, and determine the number of each node, specifically: ； in: is the node number, is the row coordinate of the node in the grid space, is the total number of columns in the grid space, is the column coordinate of the node in the grid space; S1.2.2.2: Obtain spatial distance item: Determine the corresponding spatial distance and spatial weight attenuation based on the node number and coordinates, specifically: ； in: is the spatial coordinate of node m, is the spatial coordinate of node n, is the row coordinate corresponding to node m, is the row coordinate corresponding to node n, is the column coordinate corresponding to node m, is the column coordinate corresponding to node n, is the spatial weight decay, is the base of natural logarithms, is the spatial scale parameter; S1.2.2.3: Obtain spectral similarity: Determine the spectral similarity between different nodes based on the normalized reflectance value of each node, specifically: ； in: is the cosine similarity, is the normalized reflectivity value of node m, is the normalized reflectivity value of node n; S1.2.2.4: Update node features: Determine the edge weights between different nodes based on the spatial weight decay and the spectral similarity between different nodes, construct an edge weight matrix, and update the node feature matrix of each layer in the graph neural network model based on the edge weight matrix and the node feature matrix. Specifically: ； in: is the node feature matrix of the l+1th layer, is a nonlinear activation function, is the degree matrix, is the edge weight matrix, is the node feature matrix of the lth layer, is the learnable weight matrix of layer l.

4. The AI parameter optimization control method for the regenerated polyester melt extrusion process according to claim 2, characterized in that: Get the updated feature vector, including: S1.2.3.1: Determine the attention coefficient: Based on the edge weights between different nodes, determine the attention coefficient between different nodes, specifically: ； in: is the attention coefficient between node m and node n, is the edge weight between node m and node n, is the neighborhood set of node m, is the edge weight between node m and node f, 、 、 is the node index; S1.2.3.2: Obtain updated feature vectors: Update the feature vector at each node based on the attention coefficient, specifically: ； in: is the updated normalized reflectivity value of node m, is a nonlinear activation function, is the learnable weight matrix, is the normalized reflectivity value of node n, is the neighborhood set of node m, is the node index, is the attention coefficient between node m and node n; S1.2.3.3: Determine the predicted pressure gradient value: Based on the updated eigenvector, obtain the predicted pressure gradient value corresponding to the node, specifically: ； in: is the predicted pressure gradient value at node m, is the learnable weight vector, is the updated normalized reflectivity value of node m, is the bias term.

5. The AI parameter optimization control method for the regenerated polyester melt extrusion process according to claim 1, characterized in that: Determine the total loss function, including: S2.2.1: Determine data matching items: Based on the predicted melt flow rate and the real-time melt flow rate, determine the data matching items, specifically: ； in: For data matching, is the total number of samples, The predicted flow rate of the melt for the vth sample, is the actual melt flow rate of the vth sample; S2.2.2: Determine the PDE residual term: According to the momentum conservation equation, determine the residual corresponding to each sampling point in the flow field and obtain the PDE residual term, specifically: ； in: is the PDE residual term, is the total number of nodes, is the density of the melt, is the predicted pressure gradient value at node m, is the dynamic viscosity of the melt, is the sampling time, is the Laplace operator of the velocity field at node m, is the impurity resistance term at node m; S2.2.3: Determine safety penalty items: Determine safety penalty items based on the real-time melt temperature and temperature safety threshold, specifically: ； in: For security penalties, is the real-time melt temperature, is the temperature safety threshold; S2.2.4: Determine the total loss function. Based on the data matching term, the PDE residual term, and the security penalty term, determine the total loss function, specifically: ； in: is the total loss function, For data matching, is the PDE residual term, 、 is the weight coefficient, It is a safety penalty item.

6. The AI parameter optimization control method for the regenerated polyester melt extrusion process according to claim 1, characterized in that: Output the corresponding initial process parameters, including: S2.3.1: Obtaining initial model parameters: The existing process parameters of the raw material are used as the input of the hybrid model, and the initial dynamic viscosity of the raw material at different melt temperatures and the initial learnable coefficient in the impurity resistance term are obtained as output. The specific formula for obtaining the initial dynamic viscosity is: ； in: The real-time melt temperature of the melt The dynamic viscosity under is the base viscosity of the melt at the reference temperature, is the base of natural logarithms, is the viscosity-temperature sensitivity coefficient, is the real-time melt temperature, is the reference temperature; S2.3.2: Initial model parameter update: Based on the total loss function, the initial dynamic viscosity and initial learnable coefficient are updated to obtain updated model parameters, specifically: ； in: are the updated model parameters, are the initial model parameters, is the learning rate, is the total loss function, is the total loss function Initial model parameters gradient; S2.3.3: Obtaining initial process parameters: updating the parameters of the hybrid model according to the updated model parameters, and obtaining corresponding initial process parameters according to the hybrid model after parameter update.

7. The AI parameter optimization control method for the regenerated polyester melt extrusion process according to claim 1, characterized in that: Obtain the corrected process parameters, including: S3.1: Obtain current reward value: Based on the fiber strength, unit energy consumption, and equipment wear rate during the melt extrusion process, a reward function is constructed to obtain a comprehensive reward value, specifically: ； in: is the comprehensive reward value, is the fiber strength weight coefficient, is the fiber strength index, is the unit energy consumption weight coefficient, is the energy consumption per unit product, is the equipment wear rate weight coefficient, is the equipment wear rate; S3.2: Determine the optimization strategy: Compare the comprehensive reward value with the preset reward threshold, and determine the final process parameters based on the comparison result, specifically: When the comprehensive reward value is less than the preset reward threshold, the initial process parameters are the final process parameters. Otherwise, the initial process parameters are adjusted, and steps S3.1 and S3.2 are repeated until the comprehensive reward value corresponding to the adjusted process parameters is less than the preset reward threshold.

8. The AI parameter optimization control method for the regenerated polyester melt extrusion process according to claim 7, characterized in that: The initial process parameters are adjusted by adjusting the formula to obtain the adjusted process parameters, specifically: ； in: is the melt temperature, is the system pressure, is the lower temperature threshold, is the upper pressure threshold, is the temperature-pressure compensation coefficient, is the pressure-temperature compensation coefficient, is the lower pressure threshold.

9. An AI parameter optimization control system for a recycled polyester melt extrusion process, characterized in that: The AI parameter optimization control method for the regenerated polyester melt extrusion process described in any one of claims 1 to 8 is used.

Images