Yarn tensile property prediction and process optimization system based on multi-source data fusion

Abstract

The invention discloses a yarn tensile property prediction and process optimization system based on multi-source data fusion, relates to the technical field of crossing of high-end textile intelligent manufacturing and industrial artificial intelligence, and solves the problem that it is difficult to carry out accurate and automatic quantitative characterization on a microstructure in a yarn. The technical problems that multi-source heterogeneous data such as fiber attributes, yarn structures, production processes and equipment states are difficult to be deeply fused, and a prediction mechanism capable of simultaneously predicting the two mutually-restricted performance indexes of the tensile strength and the elongation at break of the yarn at high precision is lacked are solved. Comprising a fiber and structure quantification module, a production process monitoring module, a fusion prediction model module and a process tuning module which work cooperatively; accurate quantification of a yarn three-dimensional microstructure, deep fusion of multi-source heterogeneous data under guidance of a physical mechanism, high-precision performance prediction based on bottleneck diagnosis and decoupling correction, and multi-target robust process optimization of fusion structure consistency constraint are realized.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of high-end textile intelligent manufacturing and industrial artificial intelligence, specifically involving a yarn tensile performance prediction and process optimization system based on multi-source data fusion. Background Technology

[0002] Tensile strength and elongation at break of yarn are key mechanical indicators that determine the quality of the final product and the performance of downstream textile products. In traditional technology, these two properties are mainly obtained through offline physical testing, which results in a lag and cannot be used for real-time control and optimization of the production process.

[0003] In existing technologies, some studies attempt to predict yarn performance using partial process parameters or sensor data from the production process, combined with machine learning methods. However, these methods generally suffer from the following intractable technical problems: Most existing technologies rely on limited process parameters (twist, speed) or simple online sensor signals, neglecting the fundamental factors determining yarn performance—fiber properties (strength, fineness) and the internal microstructure of the yarn (fiber arrangement, orientation, radial distribution). This fragmentation leads to models that "know what but not why," resulting in bottlenecks in prediction accuracy and generalization ability. In the few studies that attempt to integrate multiple data sources, simple feature splicing or shallow fusion is typically used, failing to delve into the complex, nonlinear interaction mechanisms between static fiber / structure properties and dynamic disturbances in the production process. This results in insufficient information fusion, making it difficult for the model to cope with complex and ever-changing actual working conditions. Existing prediction models are mostly "black boxes," with poor interpretability of prediction results, making it difficult to directly and reliably guide the reverse optimization of production process parameters. Process optimization often relies on engineer experience or simple single-objective search, lacking systematic and intelligent optimization capabilities under multiple performance indicators and physical structural constraints.

[0004] Therefore, there is an urgent need for a yarn tensile performance prediction and process optimization system based on multi-source data fusion. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a yarn tensile property prediction and process optimization system based on multi-source data fusion, to solve the following technical problems: The difficulty in accurately and automatically quantifying the microstructure of yarn (the orientation angle and radial distribution of fiber arrangement) leads to a lack of key intrinsic information that determines the macroscopic mechanical properties of yarn, making it difficult to establish an accurate "structure-performance" correlation model. Furthermore, it is difficult to deeply integrate inherently static fiber properties and yarn structure with dynamic and varied production processes and equipment conditions from multiple sources, resulting in insufficient information utilization and weak generalization ability in prediction models. There is a lack of prediction mechanisms capable of simultaneously and accurately predicting the two mutually constraining performance indicators of yarn tensile strength and elongation at break. There is also a lack of multi-objective intelligent optimization methods that use high-precision prediction models as the core while simultaneously considering the stability of production processes and the rationality of yarn microstructure. This makes it difficult to efficiently and reliably recommend optimal or near-optimal solutions from massive combinations of process parameters that comprehensively improve product quality and production robustness.

[0006] To address the aforementioned problems, this invention provides a yarn tensile property prediction and process optimization system based on multi-source data fusion, comprising the following modules: Fiber and Structure Quantification Module: Used to acquire quantitative data on fiber properties and yarn structure; Production process monitoring module: used to acquire time-series sensor data and real-time process parameter data during the production process; The fusion prediction model module connects the fiber and structure quantification module with the production process monitoring module. The fusion prediction model module includes a first feature extraction network, a second feature extraction network, and a feature fusion unit. The inputs to the first feature extraction network are the fiber property quantification data and yarn structure quantification data, and the inputs to the second feature extraction network are the time-series sensing data and real-time process parameter data. The feature fusion unit couples the first feature vector output by the first feature extraction network with the second feature vector output by the second feature extraction network, and outputs predicted values ​​for yarn tensile strength and elongation at break. Process optimization module: Connected to the fusion prediction model module, the process optimization module uses the predicted value as a performance indicator, sets two-layer optimization constraints, and adjusts the controllable parameters in the real-time process parameter data through optimization algorithms to generate recommended process parameter values.

[0007] Preferably, the quantitative data on fiber properties includes fiber type, length, fineness, and strength, and the quantitative data on yarn structure includes yarn twist, fiber orientation angle, and fiber radial distribution.

[0008] Preferably, the orientation angle of the fiber arrangement and the radial distribution of the fibers include: For the same yarn sample, two perpendicular longitudinal side views and one cross-sectional view are acquired simultaneously, and the three views are spatially registered to construct a unified two-dimensional coordinate system. Joint semantic segmentation is performed on the registered image group to identify the pixel regions of all fibers; based on the centroid position of the fiber in the cross-sectional view and its edge continuity in the two longitudinal side views, a unique identifier is assigned to each independent fiber, and cross-view tracking is performed in the two-dimensional coordinate system. The spatial angle between the fitted axis of each fiber and the overall axial direction of the yarn is taken as the three-dimensional true orientation angle of the fiber; in the cross-sectional view, the cross-sectional area of ​​each fiber is calculated with the geometric center of the yarn as the origin; combined with the three-dimensional true orientation angle, the effective load-bearing contribution of the fiber to the yarn at the corresponding radial position is calculated, specifically: ,in, In order to effectively carry the contribution value, Let be the cross-sectional area of ​​the fiber. This represents the true orientation angle in three dimensions. The three-dimensional true orientation angles of all fibers are aggregated according to their identity and effective load-bearing contribution value and their radial position to generate the three-dimensional orientation angle distribution field and the radial effective load-bearing distribution field inside the yarn, which serve as the quantitative output of the orientation angle and radial distribution of the fiber arrangement.

[0009] Preferably, the first feature vector output by the first feature extraction network includes: The first feature extraction network organizes the fiber property quantification data and yarn structure quantification data into three physical levels, including: fiber level features, aggregate level features and interaction level features; The fiber-level features consist of a feature group composed of fiber strength, fiber length, and fiber fineness; the aggregate-level features consist of a feature group composed of fiber radial distribution, fiber orientation angle distribution, and yarn twist; the interactive-level features are derived features obtained by cross-calculating fiber-level features and aggregate-level features, including weighted statistics of fiber strength and orientation angle distribution, and the rate of change of fiber fineness distribution in the radial position. The first feature extraction network processes the feature groups of the three physical layers respectively through a fully connected encoder, and concatenates their outputs to form the first feature vector.

[0010] Preferably, the second feature vector output by the second feature extraction network includes: The second feature extraction network organizes the time-series sensing data and real-time process parameter data into two information streams, including a steady-state process stream and a dynamic disturbance stream; The steady-state process flow standardizes and embeds encoding real-time process parameter data; the dynamic disturbance flow performs time-frequency analysis on time-series sensor data and extracts the energy and dominant frequency of a specific frequency band as disturbance features. The second feature extraction network processes the dynamic disturbance flow through a gated recurrent unit network, concatenates its final hidden state with the embedding vector of the steady-state process flow, and then compresses it through a fully connected layer to form the second feature vector.

[0011] Preferably, the feature fusion unit includes: The first feature vector is input into the bottleneck hypothesis network, which outputs a multidimensional bottleneck hypothesis vector. Each dimension of the multidimensional bottleneck hypothesis vector corresponds to a potential performance constraint pattern hypothesis; the second feature vector is input into the evidence extraction network, and the output is a process evidence vector. It is used to quantify the strength of evidence supporting or refuting various bottleneck hypotheses; Calculate the matching degree matrix: ,in, The outer product operation yields the matching degree matrix. Each element Indicates the first Bottleneck assumption and the first The degree of matching of process evidence; the degree of matching of the matching matrix. Perform singular value decomposition, and use the left and right singular vectors corresponding to the maximum singular value as consensus bottleneck modes. Evidence of the consensus process ; The final fused feature vector is obtained by concatenating all the vectors. ,in, The first eigenvector, This is the second feature vector. For transpose, This is the Sigmoid function.

[0012] Preferably, the fused feature vector includes: The fused feature vector The data is input to a dual-path decoupled prediction head, which is used to generate predicted values ​​for yarn tensile strength and elongation at break. Structural baseline path: The first feature vector is input into the baseline feedforward network, and the structural baseline vector is output. , ,in, The baseline tensile strength is based on fiber and structural data. The baseline breaking elongation is used; the process correction path involves inputting the fused feature vector into a correction feedforward network, which outputs a process correction matrix. Specifically: , where each element This indicates the impact of process information on performance. In the process of making corrections, with performance Related coupling components, For performance The coupling components in the self-correction process For performance Performance Coupling components under correction For performance Performance Coupling components under correction For performance Coupling components in self-correction; According to the structural reference vector With process correction matrix The final tensile strength correction and elongation at break correction are calculated, and the structural reference and process correction are added together to obtain the final predicted value.

[0013] Preferably, the calculation of the final tensile strength correction and elongation at break correction involves adding the structural baseline and process correction to obtain the final predicted value, including: The final tensile strength correction and elongation at break correction are obtained through linear transformation, specifically: ,Right now ; ;in, This is the tensile strength correction amount. This is the correction amount for elongation at break. For transpose; Adding the structural baseline and the process correction, we obtain the final predicted value: ; ;in, This is the predicted value of the yarn tensile strength. This is the predicted value for elongation at break.

[0014] Preferably, the method includes a baseline feedforward network, a modified feedforward network, and all network parameters involved in the fused feature vector, and uses a sample dataset with true strength labels and true elongation labels to perform joint end-to-end supervised training by minimizing the composite loss function. The composite loss function is specifically as follows: ; ; ; in, For composite loss function, Based on the baseline, predict the loss. To ultimately predict the loss, This is a coupling strength regularization term. True strength label For true elongation labeling, This is the preset regularization strength coefficient.

[0015] Preferably, the process optimization module includes: Using the controllable parameters in the real-time process parameter data as decision variables, an optimization problem with three objectives is constructed. The three objectives are: first objective: maximizing the predicted value of the yarn tensile strength; second objective: maximizing the predicted value of the breaking elongation; and third objective: minimizing the process disturbance index. The process disturbance index is calculated by the component related to dynamic disturbance in the second feature vector output by the second feature extraction network. The dual-layer optimization constraints include: the first layer constraint is the allowable numerical range of the process parameters themselves; the second layer constraint is the structural consistency constraint. For any candidate set of process parameters, it needs to be input together with the current fiber property quantification data into the fiber and structure quantification module and the first feature extraction network to calculate the predicted yarn structure quantification data. The Euclidean distance between the predicted yarn structure quantification data and the preset ideal structure reference vector is less than a set threshold. The fusion prediction model module is used as a surrogate model to evaluate the first and second objectives; a Bayesian optimization framework is employed to iteratively search the decision variable space under the condition of satisfying the bi-level optimization constraints. In each iteration step, a data acquisition function is used. The acquisition function selects the next evaluation point from the candidate set. ,in, and These are the mean and standard deviation of the proxy model's predictions of the combined utility of the first and second objectives, respectively. This is the predicted value of the process disturbance index at the evaluation point. and This is the balance coefficient; After the optimization process is completed, a set of Pareto optimal solutions that are not dominated on the first, second and third objectives is output. Each solution in the optimal solution set contains a set of recommended values ​​for process parameters and their corresponding predicted values ​​for the three objectives.

[0016] The beneficial effects of this invention are: This invention uses multi-view synchronous imaging, cross-view fiber identity tracking, and three-dimensional reconstruction technology to calculate the "three-dimensional true orientation angle" and "effective load-bearing contribution" of each fiber, and aggregates them to generate the "three-dimensional orientation angle distribution field" and "radial effective load-bearing distribution field", realizing the leap from qualitative description to high-dimensional quantitative characterization of microstructure; This invention organizes fiber and structural data into three physical levels: "fiber-assembly-interaction"; classifies production process data into "steady-state process flow" and "dynamic disturbance flow"; and adopts a fusion process of "bottleneck hypothesis-evidence matching-game consensus" to find the optimal consensus mode between static structure and dynamic process through singular value decomposition, thereby realizing data-driven physical mechanism mining. This invention proposes a dual-path prediction head consisting of a "structural baseline path" and a "process correction path," and introduces a process correction matrix to show the corrective effect of modeling the production process on the coupling relationship between strength and elongation. It is trained using a specific composite loss function to ensure that the model learns both robust baseline performance and accurately captures the complex effects of process disturbances. This invention introduces minimizing the process disturbance index as a third objective and applies a "structural consistency constraint" to ensure that the recommended process parameters not only improve performance but also guarantee production stability and the physical rationality of the resulting yarn microstructure. An improved Bayesian optimization acquisition function is used to guide the search toward a region of high performance and high stability. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a yarn tensile property prediction and process optimization system based on multi-source data fusion, comprising the following modules: Fiber and Structure Quantification Module: Used to acquire quantitative data on fiber properties and yarn structure; Production process monitoring module: used to acquire time-series sensor data and real-time process parameter data during the production process; The fusion prediction model module connects the fiber and structure quantification module with the production process monitoring module. The fusion prediction model module includes a first feature extraction network, a second feature extraction network, and a feature fusion unit. The inputs to the first feature extraction network are the fiber property quantification data and yarn structure quantification data, and the inputs to the second feature extraction network are the time-series sensing data and real-time process parameter data. The feature fusion unit couples the first feature vector output by the first feature extraction network with the second feature vector output by the second feature extraction network, and outputs predicted values ​​for yarn tensile strength and elongation at break. Process optimization module: Connected to the fusion prediction model module, the process optimization module uses the predicted value as a performance indicator, sets two-layer optimization constraints, and adjusts the controllable parameters in the real-time process parameter data through optimization algorithms to generate recommended process parameter values.

[0020] Specifically, quantitative data on fiber properties such as fiber type, length, fineness, and strength are obtained through laboratory testing equipment (e.g., fiber strength tester, fineness tester). Simultaneously, yarn samples are taken, and multi-view images are acquired using a microscopic imaging system. Image processing algorithms are then used to calculate yarn twist, the three-dimensional true orientation angle of fiber arrangement, and the effective radial load distribution, which serve as quantitative data for yarn structure. Furthermore, sensors (tension sensors, vibration sensors, speed encoders) installed at key locations on spinning equipment (e.g., drawing frames, roving frames, spinning frames) collect real-time data on the timing of the production process. Simultaneously, real-time process parameter data such as draft ratio, twist coefficient, and speed are acquired from the production management system. The acquired quantified fiber property data and yarn structure data are input into a first feature extraction network. This network encodes and abstracts the input data, outputting a high-dimensional first feature vector representing the fiber properties and yarn microstructure. The acquired time-series sensor data and real-time process parameter data are input into a second feature extraction network. This network first extracts time-frequency features from the time-series sensor data, embeds and encodes the process parameters, and then passes them through a gated loop. The ring unit network processes temporal features and ultimately fuses and outputs a high-dimensional second feature vector representing the dynamic state of the production process. The first and second feature vectors are input into a feature fusion unit, and the two heterogeneous features are deeply coupled through a specific fusion mechanism to generate a unified fused feature vector. Finally, this fused feature vector is input into a prediction head to directly output the predicted values ​​of yarn tensile strength and elongation at break. The process parameters that need to be adjusted are used as decision variables, and the output predicted values ​​are used as the performance targets to be maximized. At the same time, a two-layer optimization constraint is set: the first layer is the equipment allowable range of process parameters; the second layer is the structural consistency constraint, which requires that the yarn structure data predicted based on the candidate process parameters should be kept within a certain deviation range from the ideal structure reference value. The constructed fused prediction model is used as a surrogate model to evaluate the performance target. A Bayesian optimization intelligent optimization algorithm is used to iteratively search within the decision space that satisfies the above constraints. The optimization process aims to find the process parameter combination that can make the prediction performance optimal and meet the structural constraints. After optimization, a set of Pareto optimal process parameter recommended values ​​and their corresponding prediction performance are output for production decision-making.

[0021] In one embodiment of the present invention, the fiber property quantification data includes fiber type, length, fineness and strength, and the yarn structure quantification data includes yarn twist, fiber orientation angle and fiber radial distribution.

[0022] In one embodiment of the present invention, the orientation angle of the fiber arrangement and the radial distribution of the fibers include: For the same yarn sample, two perpendicular longitudinal side views and one cross-sectional view are acquired simultaneously, and the three views are spatially registered to construct a unified two-dimensional coordinate system. Joint semantic segmentation is performed on the registered image group to identify the pixel regions of all fibers; based on the centroid position of the fiber in the cross-sectional view and its edge continuity in the two longitudinal side views, a unique identifier is assigned to each independent fiber, and cross-view tracking is performed in the two-dimensional coordinate system. The spatial angle between the fitted axis of each fiber and the overall axial direction of the yarn is taken as the three-dimensional true orientation angle of the fiber; in the cross-sectional view, the cross-sectional area of ​​each fiber is calculated with the geometric center of the yarn as the origin; combined with the three-dimensional true orientation angle, the effective load-bearing contribution of the fiber to the yarn at the corresponding radial position is calculated, specifically: ,in, In order to effectively carry the contribution value, Let be the cross-sectional area of ​​the fiber. This represents the true orientation angle in three dimensions. The three-dimensional true orientation angles of all fibers are aggregated according to their identity and effective load-bearing contribution value and their radial position to generate the three-dimensional orientation angle distribution field and the radial effective load-bearing distribution field inside the yarn, which serve as the quantitative output of the orientation angle and radial distribution of the fiber arrangement.

[0023] Specifically, a representative yarn sample is taken and fixed on a specially designed sample stage. A simultaneous imaging system containing three high-resolution industrial cameras is used to instantaneously acquire data on the same sample: Cameras A and B acquire two longitudinal side views of the yarn along mutually perpendicular directions to observe the projection and orientation of the fibers along the yarn's length; Camera C acquires a cross-sectional view of the yarn along its axial direction (usually requiring cryo-embedding and precise slicing of the sample) to observe the distribution of fibers within the cross-section; the three cameras are then jointly calibrated using a calibration plate. The feature point matching algorithm performs spatial registration on the three acquired views, establishing a unified two-dimensional coordinate system covering the three-dimensional space of the yarn (usually using the cross-sectional view plane as the reference plane to establish a planar coordinate system, with the longitudinal view providing depth cues). The registered three-view image group is input into a pre-trained semantic segmentation neural network (e.g., U-Net), which is trained to identify "single fiber" and "background." After segmentation, the precise pixel regions of all fibers in each view are obtained. In the cross-sectional view, the geometric centroid coordinates of the connected regions of each fiber are calculated. In the two longitudinal side views, based on the continuity of fiber edges, an image tracking algorithm (e.g., based on optical flow or contour matching) is used to track the projection trajectory of each fiber along the yarn axis. An identity matching algorithm associates the centroid of each fiber in the cross-sectional view with the fiber projections with continuous trajectories in the two longitudinal views. Successfully matched fibers are assigned a globally unique identifier and cross-view tracking is completed under the unified two-dimensional coordinate system, thereby reconstructing the spatial position and orientation of each independent fiber in the yarn at the data level. For each tracked fiber, its position in the two longitudinal side views is used to... The projected profile in the image is used to fit the main spatial axis of the fiber in three-dimensional space using the principle of triangulation. The angle between this spatial axis and the overall axial direction of the yarn (Z-axis) is calculated. This angle is the true three-dimensional orientation angle of the fiber, and the angle range is usually between 0 and 90 degrees. In the cross-sectional view, with the geometric center of the yarn cross-section as the origin, the pixel area of ​​the fiber on the cross-section is calculated and converted into the physical cross-sectional area through calibration. The load-bearing capacity of the fiber is related to its projection on the yarn axial direction. Therefore, the effective load-bearing contribution of fiber k to the yarn at its radial position is defined as: , The closer the value is to 1, the higher the fiber strength utilization rate; conversely, the lower the value is. By statistically analyzing the three-dimensional true orientation angles of all fibers according to their fiber ID and radial position (distance from the yarn core), a two-dimensional histogram or probability density distribution field can be generated to describe the statistical characteristics of the fiber orientation angles at different radial positions. The effective load-bearing contribution values ​​of all fibers are binned and accumulated according to their radial positions (i.e., the distance from the centroid to the origin), and finally a function or vector describing the distribution of the total effective load-bearing capacity in each radial loop on the yarn cross section is generated, i.e., the radial effective load-bearing distribution field.

[0024] In one embodiment of the present invention, the first feature vector output by the first feature extraction network includes: The first feature extraction network organizes the fiber property quantification data and yarn structure quantification data into three physical levels, including: fiber level features, aggregate level features and interaction level features; The fiber-level features consist of a feature group composed of fiber strength, fiber length, and fiber fineness; the aggregate-level features consist of a feature group composed of fiber radial distribution, fiber orientation angle distribution, and yarn twist; the interactive-level features are derived features obtained by cross-calculating fiber-level features and aggregate-level features, including weighted statistics of fiber strength and orientation angle distribution, and the rate of change of fiber fineness distribution in the radial position. The first feature extraction network processes the feature groups of the three physical layers respectively through a fully connected encoder, and concatenates their outputs to form the first feature vector.

[0025] Specifically, the raw data obtained from the fiber and structure quantification module will be organized into three levels of feature groups according to their physical meaning: fiber-level feature groups. : Directly includes the inherent properties of the fiber, specifically including: fiber strength : Tensile breaking strength of single fiber, fiber length The average or main length of the fiber, and the fiber fineness. The linear density of the fiber; after normalizing these three scalars, they are directly combined into a three-dimensional vector: ; Collection-level feature groups : Describes the overall structure of the yarn formed by fiber aggregation, specifically including: fiber radial distribution The radial effective load distribution field is discretized, for example, divided into M equidistant radial rings. The total effective load contribution value within each ring is calculated, forming an M-dimensional vector; fiber orientation angle distribution. Statistical quantification of the two-dimensional orientation angle distribution field, for example, calculating the mean and standard deviation of the orientation angles, as well as the proportion of fibers in different angle intervals (0-30°, 30-60°, 60-90°), forms a K-dimensional vector; yarn twist. The metric twist or twist coefficient of the yarn is taken as a scalar; these characteristics are standardized and then concatenated into an (M+K+1) dimensional vector. Interactive hierarchical feature groups Through cross-calculation, the fiber properties are characterized under specific structures, specifically calculating the following derived features: fiber strength-orientation weighting factor. ,in, Average fiber strength, The weighting coefficients are those associated with the i-th fiber orientation angle interval, and the weighting coefficients are based on... The vector calculation reflects the effective utilization rate of fiber strength under different orientation distributions; the radial gradient of fiber fineness. Used to characterize the variation trend of fiber fineness in the yarn radial direction. ,in, and These are the estimated average fineness values ​​of the fibers within the outer and inner ring layers, respectively. Let be the yarn radius; this gradient reflects the uniformity of fiber distribution. Normalizing these two scalars, they are combined into a two-dimensional vector: The feature sets from the three levels are input into three independent fully connected encoder (FC Encoder) subnetworks: fiber-level encoder: input Output a low-dimensional dense vector ;Collection-level encoder: Input Output a low-dimensional dense vector Interactive hierarchical encoder: Input Output a low-dimensional dense vector Each encoder typically consists of 1-3 fully connected layers and an activation function (ReLU), which extracts a high-order abstract representation of the data within each layer. The output vectors of the three encoders are concatenated to form the final first feature vector. This vector integrates comprehensive and hierarchical physical information from microscopic fiber properties to macroscopic yarn structure and then to property-structure interaction, providing rich and structured input for subsequent feature fusion.

[0026] In one embodiment of the present invention, the second feature vector output by the second feature extraction network includes: The second feature extraction network organizes the time-series sensing data and real-time process parameter data into two information streams, including a steady-state process stream and a dynamic disturbance stream; The steady-state process flow standardizes and embeds encoding real-time process parameter data; the dynamic disturbance flow performs time-frequency analysis on time-series sensor data and extracts the energy and dominant frequency of a specific frequency band as disturbance features. The second feature extraction network processes the dynamic disturbance flow through a gated recurrent unit network, concatenates its final hidden state with the embedding vector of the steady-state process flow, and then compresses it through a fully connected layer to form the second feature vector.

[0027] Specifically, the acquired real-time data stream is divided into two information streams based on its representation of the production state: Steady-state process stream: This stream corresponds to settable and relatively stable process parameters during production. Typical parameters include, but are not limited to: main drafting zone drafting ratio, rear drafting zone drafting ratio, twist or twist coefficient, and front roller surface speed. After acquiring the real-time values ​​of these parameters, they are first standardized to eliminate the influence of dimensions. Then, each standardized parameter value is mapped to a low-dimensional dense vector through an embedding layer, with each process parameter corresponding to an independent embedding vector. Finally, the embedding vectors of all process parameters are summed or averaged to obtain a comprehensive steady-state process feature vector, which represents the current "preset, desired" process state. Dynamic disturbance stream: This stream corresponds to data from the equipment... The time-series sensing data from sensors contains dynamic disturbance information such as random fluctuations, mechanical vibrations, and tension transients that cannot be completely eliminated during the production process. Typical data include: yarn tension signals between the front and middle rollers (time-series), vibration acceleration signals of the spindle or ring (time-series), and main motor current fluctuation signals (time-series). For each time-series signal (e.g., tension signal), sampling is performed within a fixed time window (e.g., 10 seconds). The signals within the window are subjected to Short-Time Fourier Transform (STFT) or Wavelet Transform to extract their time-frequency characteristics. Specifically, specific frequency bands related to mechanical states are considered (e.g., the low-frequency vibration band of 5-50Hz, the high-frequency noise band of 100-500Hz), and the signal energy and dominant frequency within each band of interest are calculated. Frequency), as the perturbation feature of this time window, is used to concatenate multiple energy and main frequency features extracted from multiple signals (tension, vibration) to form a dynamic perturbation feature sequence. Assuming there are 2 signals, each focusing on 2 frequency bands, the feature dimension of each time step is (2 signals * 2 frequency bands * 2 features = 8 dimensions), and the sequence length is the number of frames within the window. The dynamic perturbation feature sequence is input into a gated recurrent unit (GRU) network. The GRU network can effectively capture long-term dependencies in time series data. After processing by the GRU network, the hidden state of the last time step is taken as a general representation of the dynamic perturbation mode within the entire time window. The vector representing the steady-state process is concatenated with the final state of the GRU representing the dynamic perturbation to obtain an intermediate fusion vector. Subsequently, this intermediate fusion vector is fed into a compressed fully connected layer, which is usually composed of 1 to 2 fully connected layers and uses an activation function such as ReLU.Its function is to perform nonlinear transformation and dimensionality reduction on the spliced ​​high-dimensional features, filter out redundant information, and extract the most core production process state representation that is related to the final performance prediction. The output of the compressed fully connected layer is the final second feature vector, which simultaneously encodes the "set steady-state process" and the "actual dynamic disturbances", comprehensively and compactly describing the real state of the current production process.

[0028] In one embodiment of the present invention, the feature fusion unit includes: The first feature vector is input into the bottleneck hypothesis network, which outputs a multidimensional bottleneck hypothesis vector. Each dimension of the multidimensional bottleneck hypothesis vector corresponds to a potential performance constraint pattern hypothesis; the second feature vector is input into the evidence extraction network, and the output is a process evidence vector. It is used to quantify the strength of evidence supporting or refuting various bottleneck hypotheses; Calculate the matching degree matrix: ,in, The outer product operation yields the matching degree matrix. Each element Indicates the first Bottleneck assumption and the first The degree of matching of process evidence; the degree of matching of the matching matrix. Perform singular value decomposition, and use the left and right singular vectors corresponding to the maximum singular value as consensus bottleneck modes. Evidence of the consensus process ; The final fused feature vector is obtained by concatenating all the vectors. ,in, The first eigenvector, This is the second feature vector. For transpose, This is the Sigmoid function.

[0029] Specifically, the first feature vector is input into a bottleneck hypothesis network, which is a multilayer perceptron (MLP) with a fixed output layer dimension of N. After training, the network learns to infer from the first feature vector the N potential performance constraint modes that the current yarn may have under ideal production conditions, and outputs a vector. , where each scalar Representing the The strength or likelihood of such bottleneck assumptions, for example, This may correspond to "insufficient utilization of fiber strength". Corresponding to issues such as "uneven twist distribution", vector This reflects the performance limitations predicted by static factors. The second feature vector is input into an evidence extraction network, which is also an MLP with a fixed output layer dimension of B. After training, the network learns to extract B types of process evidence from the second feature vector that can confirm or refute the bottleneck hypothesis, and outputs a vector... , where each scalar Representing the The strength of such evidence, for example, This may correspond to "intense tension fluctuations" (supporting the hypothesis that "fiber arrangement is disturbed"). Corresponding to "abnormal vibration of the drawing mechanism" (supporting the hypothesis of "uneven strands"), etc., vector This reflects the supporting or opposing information provided by the dynamic process; Calculate the outer product of the bottleneck hypothesis vector and the process evidence vector to obtain an N*B dimension matching degree matrix. The matrix elements... This intuitively represents the degree of coexistence between the "i-th bottleneck hypothesis" and the "j-th process evidence." A larger value indicates a higher degree of matching between the hypothesis and the evidence, and a greater likelihood that the hypothesis is supported by the evidence in the current production process. The matching matrix is ​​then subjected to singular value decomposition, sorted in descending order, and the left and right singular vectors corresponding to the largest singular values ​​are selected. The left singular vector is then normalized and used as... , Each component Represents in Within the evidentiary context, the relative importance weights of various bottleneck hypotheses reflect the "filtering" and "focusing" of dynamic process information on static bottleneck hypotheses, deriving the most likely bottleneck combination patterns; the right singular vector is normalized as... , Each component Represents in Under the assumption of [specific assumption], the relative support weights of various types of procedural evidence reflect the "interpretation" and "classification" of dynamic evidence by the static bottleneck assumption, deriving the most relevant evidence combination pattern; by concatenating all the aforementioned key vectors and adding a scalar representing consensus strength, a fused feature vector is finally formed; among them, For a scalar, calculate and The inner product of the two terms is then mapped to the (0, 1) interval using the Sigmoid function. This value represents the strength or confidence level of the consensus. and The greater the consistency (the smaller the angle), the larger the inner product, indicating a higher degree of matching between static bottlenecks and dynamic evidence, and a stronger consensus; ultimately, the feature vectors are fused. It is a high-dimensional vector that contains not only the original static and dynamic information, but also the most likely bottleneck pattern, the most relevant combination of evidence, and the confidence level of this match, obtained through game matching.

[0030] In one embodiment of the present invention, the fused feature vector includes: The fused feature vector The data is input to a dual-path decoupled prediction head, which is used to generate predicted values ​​for yarn tensile strength and elongation at break. Structural baseline path: The first feature vector is input into the baseline feedforward network, and the structural baseline vector is output. , ,in, The baseline tensile strength is based on fiber and structural data. The baseline breaking elongation is used; the process correction path involves inputting the fused feature vector into a correction feedforward network, which outputs a process correction matrix. Specifically: , where each element This indicates the impact of process information on performance. In the process of making corrections, with performance Related coupling components, For performance The coupling components in the self-correction process For performance Performance Coupling components under correction For performance Performance Coupling components under correction For performance Coupling components in self-correction; According to the structural reference vector With process correction matrix The final tensile strength correction and elongation at break correction are calculated, and the structural reference and process correction are added together to obtain the final predicted value.

[0031] Specifically, the first feature vector is input into the baseline feedforward network, which is a multilayer perceptron (MLP). Its output layer is fixed with two neurons and does not use an activation function (or uses linear activation). The network outputs a two-dimensional vector. ;in, It represents the baseline tensile strength, which is the tensile strength predicted by the model based solely on fiber properties and yarn microstructure under the assumption that the production process is completely ideal and undisturbed. Its physical unit is consistent with the actual strength. This represents the baseline breaking elongation, which, similarly, is the breaking elongation (percentage) predicted solely based on the yarn's microstructure. The fused feature vector is input into a modified feedforward network, which is also an MLP, but its output layer is fixed at four neurons and does not use an activation function. The network output is reshaped into a 2x2 process correction matrix. The structural reference vector is considered as a 2*1 column vector. The correction amount is obtained by matrix multiplication of the process correction matrix and the reference vector, i.e. ; Therefore, the final correction amount and Instead of independent calculations, the model calculates a linear combination of baseline strength and baseline elongation under a coupling relationship defined by the correction matrix. This allows the model to express the complex physical phenomenon that "when the process changes, not only are the strength and elongation baselines corrected individually, but they also influence each other." The structural baseline is added to the calculated correction amount to obtain the system's final performance prediction for the current batch of yarn. Through this dual-path design, the model's prediction results are explicitly decoupled into two parts: an interpretable baseline part. This reflects the "theoretical potential" of material and structural design; analyzable corrections. It quantifies the "realization deviation" of a specific production process from its theoretical potential, and can further understand whether the process affects performance directly or in a coupled manner by analyzing the correction matrix.

[0032] In one embodiment of the present invention, the calculation of the final tensile strength correction and elongation at break correction, by adding the structural baseline and the process correction to obtain the final predicted value, includes: The final tensile strength correction and elongation at break correction are obtained through linear transformation, specifically: ,Right now ; ;in, This is the tensile strength correction amount. This is the correction amount for elongation at break. For transpose; Adding the structural baseline and the process correction, we obtain the final predicted value: ; ;in, This is the predicted value of the yarn tensile strength. This is the predicted value for elongation at break.

[0033] In one embodiment of the present invention, the baseline feedforward network, the modified feedforward network, and all network parameters involved in the fused feature vector are jointly trained end-to-end using a sample dataset with true strength labels and true elongation labels by minimizing the composite loss function. The composite loss function is specifically as follows: ; ; ; in, For composite loss function, Based on the baseline, predict the loss. To ultimately predict the loss, This is a coupling strength regularization term. True strength label For true elongation labeling, This is the preset regularization strength coefficient.

[0034] Specifically, a historical production dataset containing several samples is collected. For each sample, the following information needs to be collected and associated synchronously: Input data: fiber property quantification data, yarn structure quantification data, time-series sensor data and real-time process parameter data during the production process of this batch; Real label: standard physical tests (single yarn strength test) are performed on the yarn samples of the final output of this batch to obtain the true tensile strength and elongation at break; The total dataset is randomly divided into training set, validation set and test set according to the proportion; All weights and bias parameters in the first feature extraction network, the second feature extraction network, the feature fusion unit (including the bottleneck hypothesis network and the evidence extraction network), and the dual-path decoupled prediction head (including the baseline feedforward network and the modified feedforward network) are randomly initialized using methods such as Xavier or He; Sample input data of a batch is taken from the training set, and according to the corresponding forward propagation path, it passes through each network module in sequence to finally calculate the predicted value of each sample in this batch: baseline prediction and final prediction; For each sample in this batch, its composite loss is calculated, where, and These are the off-diagonal elements of the correction matrix output by this sample in the process correction path. It is a preset hyperparameter (e.g., initially set to 0.01) used to control the regularization terms. The strength of the sample loss is calculated; the average loss of all samples in the batch is calculated to obtain the total batch loss; using automatic differentiation, the gradient of the total loss with respect to all trainable parameters of the model is calculated, and optimization algorithms (such as the Adam optimizer) are used to update these parameters according to the gradient; all batches in the training set are traversed to complete one training cycle, and multiple training cycles are repeated until the model's performance on the validation set (such as the mean square error of the final prediction) no longer improves significantly or begins to decline; the model performance under different hyperparameter combinations (such as learning rate, batch size, regularization strength coefficient) is evaluated using the validation set, and the hyperparameter combination with the minimum comprehensive loss on the validation set is selected; using the finally determined hyperparameters and the trained model weights, forward propagation is performed on the test set to calculate the final tensile strength prediction value and elongation at break prediction value; by comparing the predicted values ​​with the true labels, indicators such as root mean square error and mean absolute percentage error are calculated to quantitatively evaluate the model's prediction accuracy and generalization ability; The design of this composite loss function implements multiple supervision and constraints: Ensuring that the baseline feedforward network must learn a sufficiently accurate mapping of "theoretical performance" from static features lays the physical foundation for prediction; Ensuring that the final output of the entire model closely approximates the true value is the direct guarantee of overall accuracy; As a coupling strength regularization term, its core function is to prevent the correction network from over-relying on off-diagonal elements to fit noise in the training data. The incentive model can be directly modified In cases where explanations are available, direct corrections should be used first; larger coupling corrections should only be used when the data clearly show that the process change has indeed caused a systematic change in the strength-stretch relationship.

[0035] In one embodiment of the present invention, the process optimization module includes: Using the controllable parameters in the real-time process parameter data as decision variables, an optimization problem with three objectives is constructed. The three objectives are: first objective: maximizing the predicted value of the yarn tensile strength; second objective: maximizing the predicted value of the breaking elongation; and third objective: minimizing the process disturbance index. The process disturbance index is calculated by the component related to dynamic disturbance in the second feature vector output by the second feature extraction network. The dual-layer optimization constraints include: the first layer constraint is the allowable numerical range of the process parameters themselves; the second layer constraint is the structural consistency constraint. For any candidate set of process parameters, it needs to be input together with the current fiber property quantification data into the fiber and structure quantification module and the first feature extraction network to calculate the predicted yarn structure quantification data. The Euclidean distance between the predicted yarn structure quantification data and the preset ideal structure reference vector is less than a set threshold. The fusion prediction model module is used as a surrogate model to evaluate the first and second objectives; a Bayesian optimization framework is employed to iteratively search the decision variable space under the condition of satisfying the bi-level optimization constraints. In each iteration step, a data acquisition function is used. The acquisition function selects the next evaluation point from the candidate set. ,in, and These are the mean and standard deviation of the proxy model's predictions of the combined utility of the first and second objectives, respectively. This is the predicted value of the process disturbance index at the evaluation point. and This is the balance coefficient; After the optimization process is completed, a set of Pareto optimal solutions that are not dominated on the first, second and third objectives is output. Each solution in the optimal solution set contains a set of recommended values ​​for process parameters and their corresponding predicted values ​​for the three objectives.

[0036] Specifically, identify the process parameters to be optimized, namely the controllable parameters in the real-time process parameter data, such as: front zone draft ratio, rear zone draft ratio, twist coefficient, and front roller speed; The process disturbance index is calculated as follows: the candidate process parameters and the time-series sensing data (take the recent average or typical operating conditions) in the current production are input into the second feature extraction network. The sub-vector part corresponding to the "dynamic disturbance flow" in the output second feature vector is extracted, and the L2 norm of the sub-vector is calculated as the disturbance index. The smaller the value, the more stable the production system is expected under the set of process parameters. Among them, structural consistency constraints: defining constraint functions. For candidate parameters The fiber properties quantification data, along with the fixed fiber property quantification data for the current batch, are input into the system. First, the fiber and structure quantification module predicts the possible yarn structure quantification data (orientation angular distribution, radial distribution) under this process. Then, this predicted structure data is encoded into a structure feature vector through the first feature extraction network. The Euclidean distance between the structure feature vector and a preset ideal structure reference vector (which can be obtained statistically from the structure data of high-quality historical products) is calculated, and this distance is required to be less than a set threshold. This is to ensure that the optimized process can produce yarn with a reasonable microstructure; Among them, Bayesian optimization based on the improved acquisition function: treats the fusion prediction model module as a whole, considering it as an evaluation... and The proxy model; for Its computation relies on a portion of the output of the second feature extraction network, and its computational cost is relatively low; among which, , and These represent combinations of process parameters. When inputting, the first, second, and third objectives are functions; within the decision variable space, a small number (e.g., 10-20) of initial points are selected using methods such as Latin hypercube sampling, and the surrogate model and perturbation index calculation module are invoked to obtain the corresponding values ​​for each point. Values ​​constitute the initial observation dataset. For the target and Based on the current observation set Construct a Gaussian process regression model for each new point. Given the predicted mean and prediction uncertainty (standard deviation) of the target value; to balance performance exploration, uncertainty exploration, and stability preference, the following acquisition function is designed to select the next evaluation point, wherein the acquisition function is: , yes and The weighted sum of the predicted means represents the expected improvement in overall performance. ; yes and The aggregation of forecast uncertainty represents overall uncertainty; ; This is the predicted value of the process disturbance index at the evaluation point. and For balance coefficient, The degree of exploration of uncertain regions is controlled, and is usually optimized between 0.01 and 5; The degree of preference for low-disturbance (high-stability) regions is controlled, typically tuned between 0.1 and 3; this is done while satisfying two layers of constraints and Within the feasible region, find the acquisition function. The point with the largest value ;right Perform a realistic evaluation (call the proxy model to calculate) , The disturbance calculation module is called to obtain ; New observations Add to dataset Update the Gaussian process model until the preset number of iterations (100 times) or the convergence criterion is reached. After the optimization iterations are completed, perform a non-dominated sorting based on all feasible solutions (satisfying the constraints) in the final observed dataset to find all Pareto optimal solutions, i.e., those solutions for all three objectives for which no other solution is worse than it in all objectives and is strictly better than it in at least one objective. Then, assign these Pareto optimal solutions and their corresponding objective values. The solution set is compiled into a recommended set of outputs. Each solution is a set of optimized process parameter recommendations, along with predicted performance strength, elongation, and process disturbance index. Engineers can make the final selection from this set of solutions based on their actual production priorities (whether they prioritize strength or production stability).

[0037] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A yarn tensile property prediction and process optimization system based on multi-source data fusion, characterized in that, Includes the following modules: Fiber and Structure Quantification Module: Used to acquire quantitative data on fiber properties and yarn structure; Production process monitoring module: used to acquire time-series sensor data and real-time process parameter data during the production process; The fusion prediction model module connects the fiber and structure quantification module with the production process monitoring module. The fusion prediction model module includes a first feature extraction network, a second feature extraction network, and a feature fusion unit. The inputs to the first feature extraction network are the fiber property quantification data and yarn structure quantification data, and the inputs to the second feature extraction network are the time-series sensing data and real-time process parameter data. The feature fusion unit couples the first feature vector output by the first feature extraction network with the second feature vector output by the second feature extraction network, and outputs predicted values ​​for yarn tensile strength and elongation at break. Process optimization module: Connected to the fusion prediction model module, the process optimization module uses the predicted value as a performance indicator, sets two-layer optimization constraints, and adjusts the controllable parameters in the real-time process parameter data through optimization algorithms to generate recommended process parameter values.

2. The yarn tensile property prediction and process optimization system based on multi-source data fusion according to claim 1, characterized in that, The quantitative data on fiber properties includes fiber type, length, fineness, and strength, while the quantitative data on yarn structure includes yarn twist, fiber orientation angle, and fiber radial distribution.

3. The yarn tensile property prediction and process optimization system based on multi-source data fusion according to claim 2, characterized in that, The orientation angle and radial distribution of the fibers include: For the same yarn sample, two perpendicular longitudinal side views and one cross-sectional view are acquired simultaneously, and the three views are spatially registered to construct a unified two-dimensional coordinate system. Joint semantic segmentation is performed on the registered image group to identify the pixel regions of all fibers; based on the centroid position of the fiber in the cross-sectional view and its edge continuity in the two longitudinal side views, a unique identifier is assigned to each independent fiber, and cross-view tracking is performed in the two-dimensional coordinate system. The spatial angle between the fitted axis of each fiber and the overall axial direction of the yarn is taken as the three-dimensional true orientation angle of the fiber; in the cross-sectional view, the cross-sectional area of ​​each fiber is calculated with the geometric center of the yarn as the origin; combined with the three-dimensional true orientation angle, the effective load-bearing contribution of the fiber to the yarn at the corresponding radial position is calculated, specifically: ,in, In order to effectively carry the contribution value, Let be the cross-sectional area of ​​the fiber. This represents the true orientation angle in three dimensions. The three-dimensional true orientation angles of all fibers are aggregated according to their identity and effective load-bearing contribution value and their radial position to generate the three-dimensional orientation angle distribution field and the radial effective load-bearing distribution field inside the yarn, which serve as the quantitative output of the orientation angle and radial distribution of the fiber arrangement.

4. The yarn tensile property prediction and process optimization system based on multi-source data fusion according to claim 1, characterized in that, The first feature vector output by the first feature extraction network includes: The first feature extraction network organizes the fiber property quantification data and yarn structure quantification data into three physical levels, including: fiber level features, aggregate level features and interaction level features; The fiber-level features consist of a feature group composed of fiber strength, fiber length, and fiber fineness; the aggregate-level features consist of a feature group composed of fiber radial distribution, fiber orientation angle distribution, and yarn twist; the interactive-level features are derived features obtained by cross-calculating fiber-level features and aggregate-level features, including weighted statistics of fiber strength and orientation angle distribution, and the rate of change of fiber fineness distribution in the radial position. The first feature extraction network processes the feature groups of the three physical layers respectively through a fully connected encoder, and concatenates their outputs to form the first feature vector.

5. The yarn tensile property prediction and process optimization system based on multi-source data fusion according to claim 1, characterized in that, The second feature vector output by the second feature extraction network includes: The second feature extraction network organizes the time-series sensing data and real-time process parameter data into two information streams, including a steady-state process stream and a dynamic disturbance stream; The steady-state process flow standardizes and embeds encoding real-time process parameter data; the dynamic disturbance flow performs time-frequency analysis on time-series sensor data and extracts the energy and dominant frequency of a specific frequency band as disturbance features. The second feature extraction network processes the dynamic disturbance flow through a gated recurrent unit network, concatenates its final hidden state with the embedding vector of the steady-state process flow, and then compresses it through a fully connected layer to form the second feature vector.

6. The yarn tensile property prediction and process optimization system based on multi-source data fusion according to claim 1, characterized in that, The feature fusion unit includes: The first feature vector is input into the bottleneck hypothesis network, which outputs a multidimensional bottleneck hypothesis vector. Each dimension of the multidimensional bottleneck hypothesis vector corresponds to a potential performance constraint pattern hypothesis; the second feature vector is input into the evidence extraction network, and the output is a process evidence vector. It is used to quantify the strength of evidence supporting or refuting various bottleneck hypotheses; Calculate the matching degree matrix: ,in, The outer product operation yields the matching degree matrix. Each element Indicates the first Bottleneck assumption and the first The degree of matching of process evidence; the degree of matching of the matching matrix. Perform singular value decomposition, and use the left and right singular vectors corresponding to the maximum singular value as consensus bottleneck modes. Evidence of the consensus process ; The final fused feature vector is obtained by concatenating all the vectors. ,in, The first eigenvector, This is the second feature vector. For transpose, This is the Sigmoid function.

7. The yarn tensile property prediction and process optimization system based on multi-source data fusion according to claim 1, characterized in that, The fused feature vector includes: The fused feature vector The data is input to a dual-path decoupled prediction head, which is used to generate predicted values ​​for yarn tensile strength and elongation at break. Structural baseline path: The first feature vector is input into the baseline feedforward network, and the structural baseline vector is output. , ,in, The baseline tensile strength is based on fiber and structural data. The baseline breaking elongation is used; the process correction path involves inputting the fused feature vector into a correction feedforward network, which outputs a process correction matrix. Specifically: , where each element This indicates the impact of process information on performance. In the process of making corrections, with performance Related coupling components, For performance The coupling components in the self-correction process For performance Performance Coupling components under correction For performance Performance Coupling components under correction This refers to the coupling component in the performance self-correction process; According to the structural reference vector With process correction matrix The final tensile strength correction and elongation at break correction are calculated, and the structural reference and process correction are added together to obtain the final predicted value.

8. The yarn tensile property prediction and process optimization system based on multi-source data fusion according to claim 7, characterized in that, The calculation of the final tensile strength correction and elongation at break correction involves adding the structural baseline and process correction to obtain the final predicted value, including: The final tensile strength correction and elongation at break correction are obtained through linear transformation, specifically: ,Right now ; ;in, This is the tensile strength correction amount. This is the correction amount for elongation at break. For transpose; Adding the structural baseline and the process correction, we obtain the final predicted value: ; ;in, This is the predicted value of the yarn tensile strength. This is the predicted value for elongation at break.

9. The yarn tensile property prediction and process optimization system based on multi-source data fusion according to claim 7, characterized in that, The baseline feedforward network, the modified feedforward network, and all network parameters involved in the fused feature vector are jointly trained end-to-end using a sample dataset with true strength labels and true elongation labels by minimizing the composite loss function. The composite loss function is specifically as follows: ； ； ； in, For composite loss function, Based on the baseline, predict the loss. To ultimately predict the loss, This is a coupling strength regularization term. True strength label For true elongation labeling, This is the preset regularization strength coefficient.

10. The yarn tensile property prediction and process optimization system based on multi-source data fusion according to claim 1, characterized in that, The process optimization module includes: Using the controllable parameters in the real-time process parameter data as decision variables, an optimization problem with three objectives is constructed. The three objectives are: first objective: maximizing the predicted value of the yarn tensile strength; second objective: maximizing the predicted value of the breaking elongation; and third objective: minimizing the process disturbance index. The process disturbance index is calculated by the component related to dynamic disturbance in the second feature vector output by the second feature extraction network. The dual-layer optimization constraints include: the first layer constraint is the allowable numerical range of the process parameters themselves; the second layer constraint is the structural consistency constraint. For any candidate set of process parameters, it needs to be input together with the current fiber property quantification data into the fiber and structure quantification module and the first feature extraction network to calculate the predicted yarn structure quantification data. The Euclidean distance between the predicted yarn structure quantification data and the preset ideal structure reference vector is less than a set threshold. The fusion prediction model module is used as a surrogate model to evaluate the first and second objectives; a Bayesian optimization framework is employed to iteratively search the decision variable space under the condition of satisfying the bi-level optimization constraints. In each iteration step, a data acquisition function is used. The acquisition function selects the next evaluation point from the candidate set. ,in, and These are the mean and standard deviation of the proxy model's predictions of the combined utility of the first and second objectives, respectively. This is the predicted value of the process disturbance index at the evaluation point. and This is the balance coefficient; After the optimization process is completed, a set of Pareto optimal solutions that are not dominated on the first, second and third objectives is output. Each solution in the optimal solution set contains a set of recommended values ​​for process parameters and their corresponding predicted values ​​for the three objectives.

Images