Polyester process parameter configuration and optimization method based on residual neural network and spatial-temporal feature remodeling

Abstract

The invention discloses a polyester process parameter configuration and optimization method based on a residual neural network and spatial-temporal feature remodeling. In order to overcome the defects that an existing configuration method is long in period and high in cost, and a conventional algorithm is prone to being influenced by large delay of continuous production, the method comprises the steps that whole-process historical data is obtained, and time sequence backtracking offset is conducted based on material retention time to construct a time-space aligned data set; the one-dimensional parameters are remodeled into a two-dimensional process feature matrix according to process topology and physical attributes, a residual neural network is trained, and an agent model is obtained; and finally, performing reverse limited optimization by adopting a Bayesian algorithm containing a physical penalty term, and outputting an optimal process parameter combination. According to the method, the chemical mechanism and deep learning are fused, safe and rapid optimization of high-dimensional parameters can be realized without relying on complex mechanism modeling, the research and development period is remarkably shortened, raw material and energy consumption is reduced, and the product quality stability is improved.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary application technology of chemical process control and artificial intelligence, and in particular to a method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping. Background Technology

[0002] Polyester possesses excellent properties such as high strength, low elongation, heat resistance, and fatigue resistance, making it widely used in demanding industrial fields such as conveyor belts, rubber frames, and automotive seat belts. Polyester production is a highly continuous, multi-stage, and complex chemical process. Taking a typical polyester production line as an example, the material must sequentially undergo core processes such as esterification, polycondensation, thickening, and spinning, involving the coordinated control of at least 56 process parameters, including temperature, pressure, residence time, and rotational speed. The configuration of these process parameters directly determines the core performance indicators of the final finished yarn, such as viscosity, breaking strength, and shrinkage.

[0003] Traditional process parameter development and configuration mainly rely on repeated trial productions and testing feedback based on human experience. This approach has significant drawbacks: on the one hand, the R&D cycle is extremely long, and multiple trial productions lead to high costs in raw materials, energy, and time; on the other hand, due to the complex and strong coupling relationships between process parameters and the highly nonlinear relationship between parameters and finished product performance, traditional mechanism modeling is difficult, resulting in insufficient parameter control precision and easy fluctuations in product quality.

[0004] In recent years, with the development of the Industrial Internet, using artificial intelligence algorithms such as deep learning (e.g., BP neural networks, convolutional neural networks) to directly mine parameter mapping patterns from historical production data has become a research hotspot in the industry. However, existing conventional artificial intelligence algorithms, when applied to complex continuous flow chemical production lines like polyester, have revealed the following insurmountable technical bottlenecks: First, the spatiotemporal misalignment of data leads to the model extracting "spurious correlation" features. Polyester production is a continuous flow process with significant lag. The material flows from the esterification reactor at the beginning to the finished product winding at the end, often requiring several hours of physical residence. Existing data acquisition methods typically package sensor data from the entire production line at the same moment into a single sample. This results in a complete physical causal disconnect between the upstream process parameters input into the model (such as the esterification temperature several hours ago) and the performance of the finished product at the current moment, making it difficult for the model to learn the true evolution of process lags.

[0005] Secondly, one-dimensional data input loses the spatial topology and physical coupling information of the production line. Existing parameter configuration methods typically flatten dozens of process parameters directly into one-dimensional vectors and input them into a neural network. This approach severs the lateral coupling relationships between multiple physical quantities such as thermal, dynamic, and flow fields within the same process, and also obscures the longitudinal gradient evolution trend of a single physical quantity (such as temperature) along each process of the production line. As a result, ordinary convolutional networks or residual networks can only perform simple numerical weighting and cannot leverage their advantages in spatial feature extraction.

[0006] Third, the lack of physical mechanism constraints leads optimization algorithms into the "curse of high dimensionality" and "dangerous optimization." When seeking optimal process parameters in reverse, traditional pure data-driven optimization algorithms (such as standard Bayesian optimization and particle swarm optimization) only aim to minimize mathematical errors. When faced with a high-dimensional solution space of more than 50 dimensions, not only is the algorithm extremely slow to converge, but it is also prone to generating "dangerous parameters" that, while scoring high mathematically, violate common sense in physical and chemical engineering practice (e.g., recommending a downstream pipeline temperature lower than the upstream spinning box temperature, leading to melt condensation and equipment blockage).

[0007] Therefore, there is an urgent need to propose a method for configuring and optimizing polyester process parameters that can deeply integrate chemical physics mechanisms and deep learning algorithms, overcome the impact of large delays in continuous production, and achieve rapid and safe optimization of high-dimensional parameters within the physically feasible domain. Summary of the Invention

[0008] This invention addresses the shortcomings of existing technologies, such as long process parameter configuration cycles, high costs, low control precision, and the tendency of conventional deep learning algorithms to extract "pseudo-correlation" features, lose spatial topological information, and output dangerous parameters that violate physical common sense when applied to continuous chemical production due to large hysteresis characteristics. It provides a method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping.

[0009] The present invention addresses the aforementioned technical problems primarily through the following technical solution: a method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping, comprising the following steps: S1: Obtain historical production data of the entire polyester process, including high-frequency collected one-dimensional process parameters and low-frequency collected finished product performance indicators; based on the average material residence time of each process in the production line, perform time-series backtracking offset on the one-dimensional process parameters to obtain time-series aligned one-dimensional process parameters, and combine the time-series aligned one-dimensional process parameters with the corresponding finished product performance indicators to construct a time-space aligned sample dataset. S2: Reshape the time-aligned one-dimensional process parameters in the sample dataset into a two-dimensional process feature matrix; S3: Using a two-dimensional process feature matrix as input features and the corresponding finished product performance indicators as supervision labels, a pre-constructed residual neural network is trained, and features are extracted using two-dimensional convolution to obtain a surrogate model that can positively predict finished product performance indicators. S4: Receive the target preset finished product performance index, and based on the proxy model, use a Bayesian optimization algorithm including a physical penalty term to perform reverse iterative optimization to generate candidate process parameter combinations; S5: Select the optimal process parameter combination from the candidate process parameter combinations generated in each iteration and output it, and then send the optimal process parameter combination for execution.

[0010] This solution provides a complete parameter optimization closed loop that deeply integrates artificial intelligence and continuous chemical engineering mechanisms. By using time-series backtracking offset, it eliminates data misalignment caused by large-hysteresis production lines; by using feature reshaping, it endows the residual network with the ability to understand the physical topology of the production line; and by introducing a Bayesian algorithm with a physical penalty term, it transforms a purely mathematical unbounded optimization problem into a safety engineering optimization problem constrained by chemical engineering mechanisms, fundamentally solving the problems of high trial-and-error costs in traditional processes and the incompatibility of conventional AI algorithms.

[0011] The residual neural network in step S3 contains several residual modules with skip connections. These skip connections have specific chemical physics significance in this invention: polyester production exhibits a strong causal relationship where "the preceding process determines the subsequent process." Skip connections can directly transmit the underlying physical characteristics of preceding processes (such as esterification and polycondensation) to the deep network, bypassing intermediate layers. This prevents the dilution of early key chemical reaction characteristics by the nonlinear transformations of subsequent deep networks (such as spinning), thus ensuring the complete transmission of chemical causal characteristics within the surrogate model.

[0012] Preferably, the one-dimensional process parameters obtained in step S1 include the temperature, pressure, time, and rotation speed parameters of each process in the esterification section, polycondensation section, thickening section, and spinning section of the polyester production line; the finished product performance indicators include at least viscosity, breaking strength, elongation at break, and dry heat shrinkage rate.

[0013] This solution comprehensively covers the entire physicochemical transformation process of polyester from raw materials to finished products, ensuring the completeness of the process parameter space learned by the surrogate model.

[0014] Preferably, in step S1, before performing time-series backtracking offset on the one-dimensional process parameters, the data is first cleaned and preprocessed. The specific process includes: Identify erroneous values ​​in the data, use the Isolation Forest algorithm to identify outliers in the data, and use the direct deletion method to handle erroneous and outliers; Use the K-nearest neighbor algorithm to build a regression equation from the nearest data points, and replace and fill in the missing values; The cleaned data was processed using the following dimensionless normalization formula: X nom =(xx min ) / (x max -x min ), where X nom Here are the normalized data, and x is the original data. max and x min These are the maximum and minimum values ​​of the parameter, respectively.

[0015] The reason for performing cleaning before time-series backtracking is that subsequent time-series alignment relies on the aggregation of data within the time window. If uncleaned erroneous values ​​or missing values ​​due to sensor malfunctions are mixed into the window, the time window aggregation results will be severely distorted or even cause computational crashes. Furthermore, since polyester production involves multiple physical quantities such as temperature, pressure, and rotational speed, their numerical differences can reach several orders of magnitude (e.g., pressure of tens of megapascals versus temperature of hundreds of degrees Celsius). Using minimum-maximum normalization to eliminate dimensional differences is a necessary condition to ensure the smooth descent of the subsequent residual network gradient.

[0016] Preferably, before training the residual neural network, the sample dataset is divided into a training set and a test set in an 8:2 ratio, and then the training set is further divided into a training set for training and a validation set for validation in an 8:2 ratio.

[0017] By employing a double-nested 8:2 partitioning method, the validation set and test set are scientifically isolated, avoiding information leakage during model training and ensuring that the surrogate model has good generalization ability and prediction accuracy when faced with unseen production data.

[0018] Preferably, the residual neural network pre-built in step S3 is constructed based on the PyTorch deep learning framework, using a DataSet to manage data reading and loading, and a DataLoader to receive the returned data; during model training and testing, the Huber Loss loss function and the coefficient of determination R are used. 2 The mean absolute percentage error (MAPE) is used to evaluate the model's learning and predictive abilities.

[0019] Since sensor data in industrial settings inevitably contain a certain proportion of spike noise, choosing Huber Loss, which is more robust to outliers, instead of the traditional mean squared error (MSE) can effectively prevent the surrogate model from being skewed by a few outliers that have not yet been cleaned up during training, thereby improving the model's ability to resist interference in complex industrial environments.

[0020] Preferably, in step S1, the specific implementation process of performing time-series backtracking offset on the one-dimensional process parameters based on the average material residence time of each process in the production line is as follows: The time T for obtaining finished product performance data end As a trigger anchor point; Based on the equipment volume V of process i i Real-time discharge pump flow rate Q i Given the material density ρ, the average residence time Δt of process i is dynamically calculated. i =V i / (Q i ·ρ); Calculate the effective cutoff center time of process parameters for process i. , where N is the total number of sequential processes in the production line, and k is the process index from the current process i to the final product; Extraction to effectively capture the center time T i The one-dimensional process parameters of process i, based on time, are used as the time reference. end The time-aligned one-dimensional process parameters corresponding to the obtained finished product performance indicators are used to complete the time-tracking offset.

[0021] In actual continuous chemical production, fluid flow within pipelines is laminar, exhibiting stagnation and backmixing effects characterized by "slower flow velocity at the pipe walls and higher flow velocity at the center." Therefore, the finished product produced at any given moment does not necessarily originate entirely from the precise T... i A fleeting moment. To ensure the industrial feasibility of this solution, in extracting T... i When determining the process parameters at time T, the specific implementation method is as follows: A weighted average method based on Gaussian distribution (normal distribution) is used to extract the parameters at time T. i A sliding time window (T) centered on a window of size 2δ. i -δ,T i Aggregation is performed using high-frequency process parameters within +δ). The closer to the center time T... i The greater the weight of a data point, the more exponentially the weight of data points closer to the window edge decreases. This design perfectly replicates the backmixing phenomenon in fluid dynamics through mathematical distribution, thereby eliminating the spatiotemporal misalignment error caused by large hysteresis systems.

[0022] Preferably, in step S2, the specific rules for reshaping the time-aligned one-dimensional process parameters into a two-dimensional process feature matrix are as follows: The material is reshaped according to the preset process space topology sequence and physical attribute field classification, and the vertical axis of the two-dimensional process feature matrix represents the spatial topology sequence of material flow, and the horizontal axis represents the isomorphic physical attribute field. For structural gaps in the matrix caused by device sensing topology, zero-padding is used to fill them in. When the data is subsequently input into the two-dimensional convolutional layer of the residual neural network, a feature mask corresponding to the gap is introduced to prevent the gap from being updated during backpropagation of the network.

[0023] Traditional residual networks primarily process images (two-dimensional space in nature). This invention innovatively constructs an orthogonal two-dimensional mapping of "process node - physical attribute" using 56 parameters. Vertical convolution extracts the gradient change of a single physical quantity along the production line, while horizontal convolution extracts the multi-field coupling features of heat, force, and flow within the same process.

[0024] In actual production lines, sensor configurations are not perfect grids (e.g., conveyor pipes only display temperature data, not pressure data). If zeros are directly filled into the convolutional layer, the network may mistakenly identify these as edges of "process accidents where physical quantities suddenly drop to zero." Introducing a feature mask forces the network not to update gradients for these structurally empty points during backpropagation, completely shielding feature extraction interference caused by asymmetric equipment topology.

[0025] Preferably, in step S4, the specific process of using a Bayesian optimization algorithm with a physical penalty term for reverse iterative optimization is as follows: In each iteration, the Bayesian optimization algorithm calculates a score by combining the acquisition function with the performance index predicted by the surrogate model and the physical penalty term. Based on the score, it explores and outputs candidate process parameter combinations in the solution space of the process parameters. The algorithm iterates continuously until the preset maximum number of iterations is reached, or the error between the predicted performance index corresponding to the candidate process parameter combination and the target preset finished product performance index is less than the preset tolerance threshold.

[0026] Preferably, in step S4, the specific formula for calculating the score by combining the performance index predicted by the surrogate model and the physical penalty term in the acquisition function is: Score(X)=EI(X)-λ·P total (X); Where X represents the candidate combination of process parameters, EI(X) is the expected increment under unconstrained conditions, λ is the penalty coefficient, and P total (X) is the total mechanism penalty term obtained by summing up the various mechanism constraints; the mechanism constraints include the thermal gradient constraint that the temperature of the spinning box is greater than or equal to the temperature of the conveying pipeline.

[0027] The candidate process parameter combination X in the above formula is generated within the process parameter solution space. This solution space is a high-dimensional hypercube, whose upper and lower boundaries are jointly determined by the rigid mechanical limits of the production equipment (such as the maximum design temperature of the heating roller) and the historical statistical range of safe process fluctuations. Simultaneously, to overcome the "cold start" problem in high-dimensional optimization, during the initial iteration of Bayesian optimization, the K-Nearest Neighbor (KNN) algorithm is first used to retrieve the historical batches whose actual finished product performance is closest to the target preset performance from the historical database. The corresponding actual process parameters are then directly used as the initial seed solution for Bayesian optimization.

[0028] Because standard Bayesian algorithms suffer from variance explosion (the curse of dimensionality) when dealing with high-dimensional parameters such as 56-dimensional ones due to covariance matrix calculations, this invention introduces a dimensionality reduction strategy based on model sensitivity before optimization. Specifically, a pre-trained surrogate model is used to calculate the global sensitivity of each process parameter to the target performance using a gradient attribution algorithm (such as SHAP value). Only the top H (e.g., 15, the specific value can be determined according to requirements) core process parameters with the highest sensitivity are selected as free search variables, while the remaining secondary process parameters are directly fixed to the average of historical high-quality batches. This greatly reduces the optimization space and enables the algorithm to converge quickly within tens of seconds.

[0029] Preferably, in step S5, the optimal process parameter combination is selected for output. Specifically, the combination of candidate process parameters generated in each iteration with the smallest error between the predicted performance index and the target preset finished product performance index is output as the optimal process parameter combination. The selected optimal combination of process parameters is used to achieve optimal economic benefits, optimal technological innovation, optimal green chemical processes, or optimal time in actual production; Furthermore, the specific steps of issuing and executing the optimal process parameter combination include: outputting the optimal process parameter combination in text format, log format, or a format readable by the process control software.

[0030] The output of optimal process parameters is stored in a universal format, which can be directly and seamlessly integrated with the underlying control system (such as DCS system) of various existing chemical fiber production lines without the need for expensive large-scale hardware modifications to existing factory equipment, and has great industrial promotion value.

[0031] The substantial effects of this invention are: (1) Completely solve the spatiotemporal misalignment of large hysteresis systems and extract true causal features. This invention introduces a dynamic residence time model and a Gaussian weighted sliding time window algorithm to accurately realize the spatiotemporal alignment of high-frequency process data and low-frequency quality inspection data across different work sections. This eliminates the “pseudo-correlation” learning caused by large fluid hysteresis and laminar back mixing in traditional methods, and lays the foundation for high prediction accuracy of the surrogate model from the data source.

[0032] (2) Integrating chemical engineering spatial topology significantly improves the model's fitting and generalization capabilities. This invention innovatively reshapes one-dimensional process parameters into orthogonal two-dimensional matrices with physical meaning, and supplements this with a masking mechanism, enabling the residual neural network to perfectly adapt to the multi-field coupling characteristics of the longitudinal and transverse directions of chemical production lines. Verification shows that, compared to ordinary convolutional networks that flatten the input data, the residual neural network surrogate model established in this invention achieves a higher coefficient of determination R0 on the test set. 2 The value reached 0.98173, indicating a significant improvement in model fit and generalization ability.

[0033] (3) Integrating physical mechanism constraints and dimensionality reduction strategies to ensure safe and efficient optimization. By introducing physical penalty terms such as temperature anti-condensation into the Bayesian optimized acquisition function, and combining them with sensitivity dimensionality reduction and initial seed solution strategies, this invention successfully resolves the high-dimensional curse of high-dimensional parameter optimization, completely eliminating dangerous parameters that violate physical common sense in the algorithm output. Verification shows that the time taken for a single optimization is reduced to 78.8 seconds, achieving a fast and safe optimization closed loop.

[0034] (4) Significantly shortens the R&D cycle and reduces industrial production costs. This invention breaks away from the traditional trial-and-error method, reducing the parameter configuration cycle from the traditional 72 hours to less than 2 hours, shortening the R&D cycle by more than 97%. At the same time, it greatly reduces raw material loss and energy consumption during the trial production process, reducing production costs by more than 30%.

[0035] (5) Significantly improves parameter configuration accuracy and product quality stability. Because this invention can accurately capture nonlinear strong coupling relationships and perform reverse optimization, it has been verified that the error between the product performance indicators corresponding to the configured parameters and the preset values ​​can be stably controlled within 0.1% (compared to the average error of about 5% in traditional methods, the accuracy has been significantly improved). The high-precision parameter distribution effectively reduces quality fluctuations in the production process and enhances the core competitiveness of the product in the market. Attached Figure Description

[0036] Figure 1 This is a flowchart of a method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping, according to the present invention. Detailed Implementation

[0037] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0038] Example: This example describes a method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping. Figure 1 As shown, the specific steps are as follows: S1: Acquire historical production data throughout the entire process and perform spatiotemporal alignment; Acquire historical production data for the entire polyester industrial yarn process. This historical data includes one-dimensional process parameters acquired at high frequencies (e.g., collected every minute by a DCS system) and finished product performance indicators acquired at low frequencies (e.g., tested every 4 hours by a quality control laboratory). Specifically, the one-dimensional process parameters include 56 parameters such as temperature, pressure, time, and rotation speed for processes such as esterification, polycondensation, thickening, and spinning, as detailed in Tables 1 to 4. The finished product performance indicators include 10 indicators such as finished yarn viscosity (dl / g), fineness (dtex), breaking strength (cN / dtex), breaking elongation (%), dry heat shrinkage (%), oil content (%), network density (cells / m), AA rate (%), waste yarn rate (%), and number of yarn breaks.

[0039] Before performing time-series backtracking offset, the data is first cleaned and preprocessed: Identify formatted errors in the data, use the Isolation Forest algorithm to identify outliers in the data distribution, and use direct deletion to uniformly remove errors and outliers.

[0040] The K-Nearest Neighbors (KNN) algorithm is used to establish a regression equation from the K nearest data points to replace and fill in missing values ​​caused by sensor malfunctions or other reasons.

[0041] Considering the significant dimensional differences in multiphysics, a dimensionless normalization formula is used to process the cleaned data: X nom =(xx min ) / (x max -x min ), where X nom Here are the normalized data, and x is the original data. max and x min These are the maximum and minimum values ​​of the parameter, respectively.

[0042] Subsequently, a time-backoff operation is performed to eliminate the spatiotemporal misalignment error of the large-hysteresis system: The time T at which the performance index data of a certain batch of finished products is obtained end As the trigger anchor point. Based on the equipment volume V of process i. i Real-time discharge pump flow rate Q i Given the material density ρ, the average residence time Δt of process i is dynamically calculated. i =V i / (Q i ·ρ).

[0043] Next, the effective cutoff center time of process parameters for process i is calculated. , where N is the total number of sequential processes in the production line, and k is the process index from the current process i to the final product.

[0044] Considering the laminar flow of fluid within the pipe and the backmixing effect characterized by slow velocity at the pipe wall and high velocity at the center, this embodiment employs a weighted average method based on Gaussian distribution for data extraction: the extraction effectively captures the center time T. i A time window centered at a size of 2δ (T) i -δ,T i Polymerization is performed using high-frequency process parameters within the range of +δ). The value of the half-width δ is determined by the fluid residence time distribution characteristics of the reactor or pipeline in that process. In practical engineering applications, the preferred value of δ is 5%–15% of the average residence time of that process (for example, if the average residence time of a process is 120 minutes, then δ is 6–18 minutes) to fully cover the data time axis dispersion range caused by laminar flow and agitation backmixing. Polymerization characteristic value X i =∑(w t ·x t ), where x t Here are the parameter acquisition values ​​at time t, and the weight w t Obey T i The distribution follows a Gaussian distribution with a mean of 1 / 2. The one-dimensional process parameters, aggregated from all processes, are used as the parameters relative to time T. end The time-aligned one-dimensional process parameters corresponding to the obtained finished product performance indicators are used to complete the time-tracking offset, and the two are combined to construct a spatiotemporally aligned sample dataset.

[0045] S2: Two-dimensional feature reshaping of one-dimensional process parameters after timing alignment; Based on the preset process space topology sequence and physical attribute field classification, the time-aligned one-dimensional process parameters in the sample dataset are reshaped into a two-dimensional process feature matrix.

[0046] The specific mapping rules are as follows: the vertical axis of the two-dimensional process feature matrix represents the spatial topological sequence of material flow (such as dividing it into 14 process node rows according to the downstream direction, such as esterification 1, esterification 2, polycondensation 1, etc., and winding); the horizontal axis represents the isomorphic physical property field (such as dividing it into 4 physical property columns, such as thermal field / temperature, dynamic field / pressure, motion field / speed, and state field / current).

[0047] For structural gaps in the matrix caused by equipment sensing topology (e.g., a section of conveying pipeline has only temperature sensors but no speed sensors), zero-padding with a constant of 0 is used. Simultaneously, a Boolean feature mask matrix (Mask) with the same size as the two-dimensional process feature matrix (14×4) is generated, with gaps marked as 0 and locations with actual sensor data marked as 1. This feature mask is introduced into the subsequent two-dimensional convolutional layers of the residual neural network to prevent gaps from being updated with gradients during backpropagation, thus avoiding misinterpretation by the network as a step signal indicating a sudden drop in physical quantity.

[0048] S3: Construction and training of residual neural network surrogate model; The sample dataset is divided into a training set and a test set in an 8:2 ratio. The training set is then divided into a training set for training and a validation set for validation in an 8:2 ratio.

[0049] A residual neural network with multiple residual modules was built based on the PyTorch deep learning framework. A two-dimensional process feature matrix was used as input features, and the corresponding finished product performance indicators were used as supervision labels for training. The two-dimensional convolutional kernels in the network slide across the matrix to extract the lateral coupling features of multiple physics fields within the same process and the longitudinal gradient evolution features of a single physics field along the production line. Simultaneously, skip connections in the network directly transmit the low-level physical features of preceding processes (such as the esterification stage) to the deep network, ensuring the transmission of chemical causal features where preceding processes determine subsequent processes.

[0050] During model training, Huber Loss, which is more robust to outlier noise, is used as the loss function. After training, the determination coefficient R of the residual neural network surrogate model constructed in this embodiment is verified using a test set. 2 It achieves a score of 0.98173, which is significantly better than traditional BP neural networks (0.80346) and convolutional neural networks (0.82349), demonstrating excellent positive performance prediction accuracy.

[0051] S4: Inverse Bayesian optimization based on physical constraints and sensitivity reduction; The system receives the target pre-defined performance indicators of the finished product and, based on a trained surrogate model, performs iterative optimization using a Bayesian optimization algorithm that includes a physical penalty term. To address the curse of dimensionality problem arising from optimizing 56-dimensional process parameters and improve industrial safety, the following sub-steps are specifically executed: 1. Sensitivity Reduction and Locking: Using a surrogate model, the global sensitivity weights of 56 process parameters to the target preset finished product performance indicators are calculated through the SHAP gradient attribution algorithm. The top 15 core process parameters with the highest sensitivity are selected as the free search variables for the Bayesian optimization algorithm (constructing a low-dimensional process parameter solution space); for the remaining minor process parameters, the average values ​​of the parameters are extracted from historical high-quality production batches and fixed, and they are not included in the joint optimization calculation.

[0052] 2. Definition of physical process parameter solution space: The boundary of the above low-dimensional process parameter solution space is jointly formed by the mechanical rigidity limit of the equipment (such as the maximum design temperature of the heating roller and the maximum speed of the metering pump) and the upper and lower limits of historical safe production fluctuations.

[0053] 3. Cold start initialization: Using the K nearest neighbor (KNN) algorithm, retrieve the historical production batches whose actual finished product performance is closest to the target preset finished product performance index in Euclidean distance from the historical database, and use the corresponding actual process parameters as the initial candidate process parameter combination for the Bayesian algorithm.

[0054] 4. Constrained Iterative Optimization: In each iteration, the Bayesian optimization algorithm calculates the score using the acquisition function combined with the performance metrics predicted by the surrogate model and the physical penalty term. The specific scoring formula is: Score(X) = EI(X) - λ·P total (X).

[0055] Where X represents the generated candidate process parameter combination, EI(X) is the expected increment under unconstrained conditions, λ is the penalty coefficient that dynamically and adaptively increases with the number of iterations, and P total (X) represents the total mechanistic penalty term obtained by summing up the various mechanistic constraints. For example, the mechanistic constraints include the spinning box temperature (T). box The temperature must be greater than or equal to the temperature of the pipeline (T). pipe The anti-condensation thermal gradient constraint is defined as follows: Let the constraint function be g1(X) = T. pipe -T box When g1(X)>0 (i.e., temperature inversion occurs, violating the mechanism), a penalty value P1(X)=g1(X) is generated, which greatly reduces the final score of this dangerous parameter combination in the acquisition function.

[0056] Guided by the score, the Bayesian optimization algorithm continuously generates and outputs candidate process parameter combinations for the next round within the solution space of process parameters. It stops iterative optimization when the preset maximum number of iterations (e.g., 500 times) is reached, or when the absolute error between the predicted performance index corresponding to the candidate process parameter combination and the target preset finished product performance index is less than the preset tolerance threshold (e.g., fracture strength error < 0.1 cN / dtex).

[0057] S5: Selection and distribution of optimal parameters; When the iterative optimization stops, the optimal process parameter combination is output as the one with the smallest error between the predicted performance index and the target preset finished product performance index among the candidate process parameter combinations generated in each iteration. The selected optimal process parameter combination can be used to achieve macro-level goals such as optimal economic benefits, optimal technological innovation, optimal green chemical engineering, or optimal time in actual production.

[0058] Finally, the optimal combination of process parameters is output in text, log, or CSV format readable by the process control software and sent to the DCS control system of the polyester industrial yarn production line for execution.

[0059] This embodiment extracts a set of data from the test set to test the system in order to verify the reliability of the technical solution of this embodiment. After the system is configured and optimized, the process parameter configuration optimization results can be obtained as shown in Tables 1 to 4 below, and the configuration values ​​and errors are shown in the tables.

[0060] Table 1: Process Parameter Configuration for Esterification Section Parameter name True value Configuration value error(%) Raw material color value L L L+0.1294 0.13084 Raw material color value a A A+0.00064 0.09171 Raw material color value b B B+0.00012 0.00725 EG transmittance (220nm ≥ 75%) Tr1 Tr1+0.02707 0.02943 EG transmittance (275nm ≥ 92%) Tr2 Tr2-0.01287 0.013 EG transmittance (350nm ≥ 99%) Tr3 Tr3-0.04913 0.04913 Moor ratio MR MR-0.00149 0.12922 Esterification temperature (°C) ET1 ET1+0.22705 0.08766 Esterification pressure (kPa) EP1 EP1+0.04258 0.0743 Esterification temperature (°C) ET2 ET2-0.00344 0.00128 Esterification pressure (kPa) EP2 EP2-0.00143 0.0172 Table 2: Process Parameter Configuration for Polycondensation Section Parameter name True value Configuration value error(%) Polycondensation temperature (°C) CT1 CT1+0.26254 0.09682 Polycondensation pressure (kPa) CP1 CP1+0.00058 0.00641 Polycondensation temperature (°C) CT2 CT2+0.21931 0.07962 Polycondensation pressure (Pa) CP2 CP2-1.47707 0.12744 Final polycondensation temperature (°C) CT3 CT3+0.12849 0.046 Final polycondensation pressure (Pa) CP3 CP3+0.1381 0.13028 Table 3: Process Parameter Configuration for the Thickening Section Parameter name True value Configuration value error(%) Delivery pressure (MPa) DP DP-0.0016 0.05137 Temperature (°C) VT VT-0.15305 0.05475 Vacuum pressure (Pa) VP VP-0.04975 0.07425 Low viscosity zone conveying temperature (°C) LDP LDP-0.22603 0.0816 High viscosity zone conveying temperature (°C) HDT HDT-0.28499 0.10196 High-viscosity melt delivery pressure (MPa) HDP HDP-0.00789 0.04385 Table 4: Spinning Section Process Parameter Configuration Parameter name True value Configuration value error(%) Screw temperature in Zone 1 (°C) ST1 ST1+0.35734 0.11453 Screw temperature in Zone 2 (°C) ST2 ST2+0.39333 0.12408 Three-zone screw temperature (°C) ST3 ST3+0.09249 0.02984 Screw temperature in zone four (°C) ST4 ST4+0.05088 0.01696 Screw temperature in zone 5 (°C) ST5 ST5-0.1641 0.05659 Current (A) I I+0.24163 0.13276 Head pressure (bar) HP HP-0.15628 0.12502 Measuring head temperature (°C) HT HT-0.09836 0.03274 Screw speed (rpm) SR SR-0.01298 0.07865 Biphenyl furnace temperature (°C) BFT BFT-0.38771 0.12924 Chamber temperature (°C) BT BT-0.24981 0.08226 Pump inlet pressure (bar) BPT BPT-0.00249 0.03406 Pump post-pump pressure (bar) APT APT-0.15433 0.09526 Metering pump specifications (cc) PS PS+0.01918 0.07671 Metering pump speed (r / min) PR PR-0.01818 0.13252 Post-heating (°C) PHT PHT+0.28439 0.08364 Temperature (°C) ST ST-0.01466 0.0698 humidity(%) H H-4e-05 4.41797E-5 Wind speed (m / min) WS WS-0.00126 0.13238 Wind pressure (kPa) WP WP-0.50336 0.07191 Oil pump speed / specifications (rpm, cc) PV PV-0.03124 0.1157 Hot roller speed 1+2 (m / min) HRV1 HRV1-0.25381 0.05958 Hot roller temperature 3+4 (°C) HRT3 HRT3-0.01246 0.01661 Hot roller speed 3+4 (m / min) HRV3 HRV3-0.04566 0.01047 Hot roller temperature 5+6 (°C) HRT5 HRT5+0.11827 0.12449 Hot roller speed 5+6 (m / min) HRV5 HRV5+0.27455 0.06021 The temperature of the hot roller is 7+8 (°C). HRT7 HRT7+0.00291 0.00253 Hot roller speed 7+8 (m / min) HRV7 HRV7+1.88264 0.10882 Hot roller temperature 9+10 (°C) HRT9 HRT9-0.02472 0.01129 Hot roller speed 9+10 (m / min) HRV9 HRV9-1.72128 0.06471 Stretch ratio DR DR-0.00229 0.03663 Winding speed (m / min) WS WS+1.99855 0.07994 Winding tension (cN) WT WT-0.73103 0.08808 Verification has shown that using the method of this embodiment for process parameter configuration, the single optimization time is only 78.8 seconds, and the parameter configuration cycle is shortened from the traditional 72 hours to less than 2 hours. Moreover, the error between the product performance indicators corresponding to the configured parameters and the preset values ​​in actual production is stably controlled within 0.1%, achieving unexpected technical results.

[0061] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0062] Although this paper makes extensive use of terms such as industrial parameters, finished product performance indicators, and surrogate models, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. A method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping, characterized in that, Includes the following steps: S1: Obtain historical production data for the entire polyester production process, including one-dimensional process parameters and finished product performance indicators; Based on the average material dwell time of each process in the production line, the one-dimensional process parameters are time-series backtracked and offset to obtain time-series aligned one-dimensional process parameters. The time-series aligned one-dimensional process parameters are then combined with the corresponding finished product performance indicators to construct a time-space aligned sample dataset. S2: Reshape the time-aligned one-dimensional process parameters in the sample dataset into a two-dimensional process feature matrix; S3: Using a two-dimensional process feature matrix as input features and the corresponding finished product performance index as supervision labels, a pre-constructed residual neural network is trained to obtain a surrogate model that can positively predict the finished product performance index. S4: Receive the target preset finished product performance index, and based on the proxy model, use a Bayesian optimization algorithm including a physical penalty term to perform reverse iterative optimization to generate candidate process parameter combinations; S5: Select the optimal process parameter combination from the candidate process parameter combinations generated in each iteration and output it, and then send the optimal process parameter combination for execution.

2. The method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping according to claim 1, characterized in that, The one-dimensional process parameters obtained in step S1 include the temperature, pressure, time, and rotation speed parameters of each process in the esterification section, polycondensation section, thickening section, and spinning section of the polyester production line; the finished product performance indicators include at least viscosity, breaking strength, elongation at break, and dry heat shrinkage rate.

3. The method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping according to claim 1, characterized in that, In step S1, before performing time-series backtracking offset on the one-dimensional process parameters, the data is first cleaned and preprocessed. The specific process includes: Identify erroneous values ​​in the data, use the Isolation Forest algorithm to identify outliers in the data, and use the direct deletion method to handle erroneous and outliers; Use the K-nearest neighbor algorithm to build a regression equation from the nearest data points, and replace and fill in the missing values; The cleaned data was processed using the following dimensionless normalization formula: X nom =(xx min ) / (x max -x min ), where X nom The data is normalized, and x is the original data. max and x min These are the maximum and minimum values ​​of the parameter, respectively.

4. A method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping according to claim 1 or 3, characterized in that, Before training the residual neural network, the sample dataset is divided into a training set and a test set in an 8:2 ratio. The training set is then divided into a training set for training and a validation set for validation in an 8:2 ratio.

5. The method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping according to claim 1, characterized in that, The residual neural network pre-built in step S3 is based on the PyTorch deep learning framework, using DataSet to manage data reading and loading, and DataLoader to receive the returned data. During model training and testing, the Huber Loss loss function and the coefficient of determination R are used. 2 The mean absolute percentage error (MAPE) is used to evaluate the model's learning and predictive abilities.

6. The method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping according to claim 1, characterized in that, In step S1, the specific implementation process of performing time-series backtracking offset on the one-dimensional process parameters based on the average material residence time of each process in the production line is as follows: The time T for obtaining finished product performance data end As a trigger anchor point; Based on the equipment volume V of process i i Real-time discharge pump flow rate Q i Given the material density ρ, the average residence time Δt of process i is dynamically calculated. i =V i / (Q i ·ρ); Calculate the effective cutoff center time of process parameters for process i. , where N is the total number of sequential processes in the production line, and k is the process index from the current process i to the final product; Extraction to effectively capture the center time T i The one-dimensional process parameters of process i, based on time, are used as the time reference. end The time-aligned one-dimensional process parameters corresponding to the obtained finished product performance indicators are used to complete the time-tracking offset.

7. The method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping according to claim 1, characterized in that, In step S2, the specific rules for reshaping the time-aligned one-dimensional process parameters into a two-dimensional process feature matrix are as follows: The material is reshaped according to the preset process space topology sequence and physical attribute field classification, and the vertical axis of the two-dimensional process feature matrix represents the spatial topology sequence of material flow, and the horizontal axis represents the isomorphic physical attribute field. For structural gaps in the matrix caused by device sensing topology, zero-padding is used to fill them in. When the data is subsequently input into the two-dimensional convolutional layer of the residual neural network, a feature mask corresponding to the gap is introduced to prevent the gap from being updated during backpropagation of the network.

8. The method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping according to claim 1, characterized in that, In step S4, the specific process of using a Bayesian optimization algorithm with a physical penalty term for back-iterative optimization is as follows: In each iteration, the Bayesian optimization algorithm calculates a score by combining the acquisition function with the performance index predicted by the surrogate model and the physical penalty term, and explores and outputs candidate process parameter combinations in the solution space of the process parameters based on the score. The process is iterated continuously until the preset maximum number of iterations is reached, or the error between the predicted performance index corresponding to the candidate process parameter combination and the target preset finished product performance index is less than the preset tolerance threshold.

9. The method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping according to claim 8, characterized in that, In step S4, the specific formula for calculating the score by combining the performance index predicted by the surrogate model and the physical penalty term in the acquisition function is: Score(X)=EI(X)-λ·P total (X); Where X represents the candidate combination of process parameters, EI(X) is the expected increment under unconstrained conditions, λ is the penalty coefficient, and P total (X) is the total mechanism penalty term obtained by summing up the various mechanism constraints; the mechanism constraints include the thermal gradient constraint that the temperature of the spinning box is greater than or equal to the temperature of the conveying pipeline.

10. The method for configuring and optimizing polyester process parameters based on residual neural networks and spatiotemporal feature reshaping according to claim 1, characterized in that, In step S5, the optimal process parameter combination is selected for output. Specifically, the combination of candidate process parameters generated in each iteration with the smallest error between the predicted performance index and the target preset finished product performance index is output as the optimal process parameter combination. The selected optimal combination of process parameters is used to achieve optimal economic benefits, optimal technological innovation, optimal green chemical processes, or optimal time in actual production; Furthermore, the specific steps of issuing and executing the optimal process parameter combination include: outputting the optimal process parameter combination in text format, log format, or a format readable by the process control software.

Images