Cloth antibacterial deodorization multi-objective optimization method under safety constraint

Abstract

The invention provides a cloth antibacterial deodorization multi-objective optimization method under safety constraints, and the method specifically comprises the steps: obtaining a cloth microstructure feature coding vector, functional agent physicochemical attribute data and process parameter data, constructing comprehensive input data, and inputting the comprehensive input data into a multi-task performance prediction model; the antibacterial rate, the deodorization durability and the skin irritation index are synchronously output; according to the method, the variety, proportion and process parameters of functional agents are taken as decision variables, a multi-objective optimization problem taking maximization of antibacterial rate and deodorization durability as an objective and taking skin irritation not exceeding a safety threshold as a constraint is constructed, a multi-objective evolutionary algorithm with constraint processing capacity is adopted for solving, and a Pareto optimal scheme set meeting the safety constraint is obtained. According to the method, multi-objective collaborative optimization under security constraints is realized, all candidate schemes are ensured to meet skin security requirements, and effective technical support is provided for intelligent design and manufacturing of functional textiles.

Description

Technical Field

[0001] This invention belongs to the field of textile functional finishing and intelligent optimization technology, specifically relating to a multi-objective optimization method for fabric antibacterial and deodorizing under safety constraints. Background Technology

[0002] With the popularization of health-conscious consumption, the demand for functional textiles in everyday wear and medical protection scenarios continues to grow, among which the antibacterial and deodorizing properties of fabrics have become key indicators for measuring their added value. Current textile functionalization technologies mainly focus on material compounding, physical modification, or structural composites, aiming to enhance the fabric's ability to inhibit microorganisms and adsorb and decompose odor molecules. However, these methods generally lack a systematic consideration of the intrinsic microstructure of the fabric, and do not take human skin safety as a rigid boundary condition in the design stage. This leads to problems such as unstable performance, questionable safety, or excessively high overall costs in practical applications.

[0003] While material compounding strategies based on natural extracts or nano-functional agents can achieve certain antibacterial and deodorizing effects, their formulation selection relies heavily on trial and error, making it difficult to establish a quantifiable synergistic optimization mechanism between antibacterial rate, deodorizing durability, production cost, and skin irritation. Meanwhile, functionalizing fabric surfaces using physical processes such as plasma or ion deposition, while avoiding chemical additives, is limited by equipment compatibility and substrate sensitivity, failing to achieve adaptive control based on differences in fiber topology and pore networks, and lacks effective verification of biocompatibility under long-term use. Furthermore, integrating different functional modules through multi-layer composite structures, while intuitively combining multiple properties, sacrifices fabric breathability and comfort, and causes non-linear functional decay due to interlayer interface effects, making it difficult to achieve precise functional at the fiber scale.

[0004] Therefore, there is an urgent need for an intelligent optimization method that can integrate fabric microstructure information, be based on embedded safety constraints, and support multi-objective collaborative optimization, so as to promote a new paradigm of functional textile R&D from experience-driven to model-driven. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-objective optimization method for antibacterial and deodorizing fabrics under safety constraints, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A multi-objective optimization method for fabric antibacterial and deodorizing under safety constraints includes the following steps:

[0008] Obtain microstructural feature encoding vectors that can quantitatively characterize the fiber topology and pore network distribution of the target fabric;

[0009] Obtain the physicochemical property data of the candidate antibacterial agents and deodorants, as well as the loading process parameters to be optimized;

[0010] The microstructure feature encoding vector, the physicochemical property data of the selected functional agent, and the process parameter data are combined to form comprehensive input data;

[0011] The comprehensive input data is input into the trained multi-task performance prediction model, which then outputs the antibacterial rate, deodorization duration, and skin irritation index values ​​simultaneously.

[0012] Using the selection of functional agents, the mass ratio of each functional agent, and the specific values ​​of process parameters as decision variables, a multi-objective optimization problem is constructed. The optimization objective is to maximize the antibacterial rate and the deodorization duration, and the constraint is that the skin irritation index must not exceed the preset safety threshold.

[0013] The optimization problem is solved using a multi-objective evolutionary algorithm with constraint handling capabilities. Under the premise that the skin irritation index value does not exceed a preset safety threshold, a set of non-dominant Pareto optimal candidate solutions are obtained.

[0014] Select the target scheme from the Pareto optimal candidate schemes, output the corresponding functional agent combination type, the mass ratio of each functional agent, the specific values ​​of the process parameters, and generate a visualization result of the spatial load distribution of functional agents in the fabric microstructure of the scheme.

[0015] Furthermore, the microstructure feature encoding vector is obtained through any of the following paths:

[0016] Path A: Image the target fabric using a scanning electron microscope or micro-CT device to obtain two-dimensional / three-dimensional image data with sub-micron resolution; perform image segmentation and skeleton extraction on the image data to construct a graph structure. , where the node set Characterizing fiber intersections or pore geometric centers, where n is the total number of nodes and edge sets. Characterize the fiber connections or pore channel connectivity; [The diagram structure is then described]. Input a graph neural network encoder, output a fixed-dimensional structural feature encoding vector. ;

[0017] Path B: Collect the fabric density of the target fabric using a fabric density meter, thickness gauge, air permeability tester, and near-infrared spectrometer. ,thickness Gram per square meter Breathability and near-infrared reflectance spectral matrix Where L is the number of wavelengths and M is the number of sampling points, forming a set of macroscopic physical parameters. ;Will Input structure encoding mapping model Output the predicted structural feature encoding vector The structure encoding mapping model The training samples were obtained through supervised training. Group basic fabric corresponding Mapping pairs, where Generated from path A.

[0018] Specifically, the graph structure Middle node initial feature vector Constructed from its local geometric properties, including: the curvature at the location of the node. Mean pore area within the neighborhood and fiber diameter statistical moments ,Right now ;side initial feature vector Includes connection length Equivalent radius of cross section .

[0019] Furthermore, the multi-task performance prediction model includes a shared feature encoding module and three parallel task output heads; the shared feature encoding module receives the concatenated vector. And output high-level semantic features Each task's output header is an independent fully connected subnet, and it outputs the following respectively:

[0020] Antibacterial rate prediction ,in It is the Sigmoid activation function. , For learnable bias terms;

[0021] Odor retention prediction value ,in It is a linear rectifier unit. for For learnable bias terms;

[0022] Predicted Skin Irritation Index ,in , This is a learnable bias term.

[0023] Specifically, the shared feature encoding module is implemented using a three-layer fully connected network, wherein the first layer... The layer output is:

[0024] in , , For Gaussian error linear units, It is the standard normal cumulative distribution function.

[0025] Furthermore, the multi-task performance prediction model employs a weighted multi-task loss function during the training phase:

[0026]

[0027] in , , , , .

[0028] Furthermore, the multi-objective evolutionary algorithm defines the individual constraint violation degree in constraint processing. Defined as:

[0029]

[0030] In the non-dominated sorting process, all satisfying Individuals are forcibly relegated to the worst non-dominated class, only... The subsequent stratification and congestion distance calculations are performed on the feasible solution subset.

[0031] Furthermore, during the training phase of the multi-task performance prediction model, a graph structure contrastive learning regularization term is introduced, specifically: for two different graph augmented views of the same fabric sample... and Its corresponding structural feature encoding , The similarity constraint should be satisfied, and the contrastive loss function is:

[0032]

[0033] in For cosine similarity, For temperature coefficient, This is an indicator function.

[0034] Furthermore, the visualization result of the spatial loading distribution of the functional agent is a heatmap. The heatmap H is generated by inputting the decision variables corresponding to the Pareto optimal solution into a graph attention model trained with interpretability. This model uses a graph structure. For input, output each node Functional agent loading probability And mapped to the pixel grid according to the node spatial coordinates, where Calculate using the following formula:

[0035] in For graph neural networks Layer node embedding, For nodes The set of first-order neighbors, For graph attention aggregation function, , The grid resolution of the heatmap is [value], and ≥256, the heat map is used to drive a six-axis robotic arm to control the nozzle to perform gradient spraying according to a probability distribution.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] This invention achieves digital characterization of the fabric's microstructure by acquiring microstructural feature encoding vectors that quantitatively characterize the fiber topology and pore network distribution of the target fabric, providing a physical basis for subsequent performance prediction. A complete functional processing parameter system is constructed by acquiring the physicochemical property data of candidate antibacterial agents and deodorants, as well as the load process parameters to be optimized. The microstructural feature encoding vectors, the physicochemical property data of the selected functional agents, and the process parameter data are combined to form comprehensive input data, achieving the fusion of multi-source heterogeneous data. This comprehensive input data is fed into a trained multi-task performance prediction model, which simultaneously outputs antibacterial rate, deodorization durability, and skin irritation index, achieving efficient and accurate prediction of the fabric's functional performance. A multi-objective optimization problem is constructed using the selection of functional agents, the mass ratio of each functional agent, and the specific values ​​of the process parameters as decision variables. The optimization objective is to maximize the antibacterial rate and deodorization durability, with the constraint that the skin irritation index must not exceed a preset safety threshold, establishing a scientifically sound optimization framework. A multi-objective evolutionary algorithm with constraint handling capability is employed to solve the optimization problem. Under the premise that the skin irritation index does not exceed a preset safety threshold, a set of non-dominant Pareto optimal candidate solutions is obtained, ensuring the comprehensiveness and diversity of the optimization results. From the Pareto optimal candidate solutions, a target solution is selected, and the corresponding functional agent combination type, the mass ratio of each functional agent, and the specific values ​​of process parameters are output. A visualization of the spatial load distribution of functional agents on the fabric microstructure is also generated, realizing a complete closed loop from digital optimization to practical application. This method fundamentally solves the problem of achieving comprehensive optimization of antibacterial and deodorizing performance under safety constraints in traditional fabric functionalization R&D. It realizes the "safety-by-design" design concept, ensuring that all candidate solutions meet skin safety requirements. Simultaneously, multi-objective optimization provides a scientific decision-making space for performance trade-offs, significantly improving R&D efficiency and reducing costs. Attached Figure Description

[0038] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0039] Figure 1 This is an overall flowchart of a multi-objective optimization method for fabric antibacterial and deodorizing under safety constraints provided in an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the structure of the multi-task performance prediction model in an embodiment of the present invention.

[0041] Figure 3 This is a schematic diagram of the structure of the shared feature encoding module in an embodiment of the present invention.

[0042] Figure 4 This is a block diagram of a functional agent spatial load distribution prediction model based on graph attention networks in an embodiment of the present invention.

[0043] Figure 5 This is a verification diagram of the prediction accuracy of the multi-task performance prediction model in this embodiment of the invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0045] Example 1

[0046] like Figure 1 As shown, Embodiment 1 of the present invention discloses a multi-objective optimization method for fabric antibacterial and deodorizing under safety constraints, comprising the following steps:

[0047] Step S1: Obtain the microstructural feature encoding vector that can quantitatively characterize the fiber topology and pore network distribution of the target fabric.

[0048] Specifically, this step aims to transform the complex submicron-level physical structure of the fabric into a fixed-dimensional, differentiable numerical vector, providing a unified, physically meaningful input basis for subsequent performance prediction. This step offers two parallel implementation paths to adapt to the accuracy and efficiency requirements of different application scenarios.

[0049] Specifically, step S1 includes the following two paths:

[0050] Step S11: Path A: Image the target fabric using a scanning electron microscope or micro-CT device to obtain two-dimensional / three-dimensional image data with sub-micron resolution; perform image segmentation and skeleton extraction on the image data to construct a graph structure. , where the node set Characterizing fiber intersections or pore geometric centers, where n is the total number of nodes and edge sets. Characterize the fiber connections or pore channel connectivity; [The diagram structure is then described]. Input a graph neural network encoder, output a fixed-dimensional structural feature encoding vector. .

[0051] Specifically, firstly, high-resolution images or volumetric data of the fabric surface or interior are acquired using a scanning electron microscope (SEM, operating voltage 15kV) or micro-CT (spatial resolution 0.5μm). Next, a U-Net-based semantic segmentation model is used to process the images, distinguishing between the fibrous phase and the porous phase, and a skeleton extraction algorithm (such as the Zhang-Suen algorithm) is employed to obtain the connectivity paths between the fiber skeleton centerline and the pores. Based on this, an undirected graph is constructed. Each node Represents the geometric center of a fiber branch point or a pore region, with spatial coordinates (x, y, y). i ,y i ,z i The coordinate system is derived from the image coordinate system. If the physical structures represented by two nodes are adjacent in space (e.g., the Euclidean distance is less than 5 pixels and they satisfy 8-connectivity or 26-connectivity), then an edge is established between them. .

[0052] Furthermore, the construction and encoding of the graph structure in step S11 specifically includes a detailed definition of the initial features of nodes and edges:

[0053] Step S111: Node and edge feature construction: the graph structure Middle node v i initial feature vector Constructed from its local geometric properties, including: the curvature at the location of the node. Mean pore area within the neighborhood and fiber diameter statistical moments ,Right now ;side initial feature vector Includes connection length Equivalent radius of cross section .

[0054] Specifically, for each node Calculate its local curvature (Unit: micrometers⁻¹), reflecting the degree of fiber bending; the arithmetic mean of the cross-sectional areas of all pores within a 5-micrometer radius centered on the fiber is calculated. (Unit: square micrometers), characterizing local spatial capacity; calculating the first moment (mean) of the fiber diameter along a 20-micrometer segment extending along the fiber direction. (Unit: micrometers), representing the thickness of local fibers. These three scalars are concatenated to form the initial 3D node feature vector. For each edge The Euclidean distance between the two endpoints is calculated as the connection length (unit: micrometers). The minimum cross-sectional area and wetted perimeter of the connection channel are calculated through image analysis, and the hydraulic radius is obtained as the equivalent radius of the cross section. (Unit: micrometers) Concatenate these two scalars to form a 2D edge initial feature vector. By endowing the graph structure G with these initial features that have clear physical meaning, a strong inductive bias is provided for the subsequent graph neural network to understand the microscopic geometry and topological properties of the cloth.

[0055] The constructed graph structure G is input into a pre-defined graph neural network encoder. This encoder aggregates feature information from nodes and their neighbors through multiple rounds of message passing, and finally outputs a fixed 128-dimensional structural feature encoding vector through global average pooling layers and linear projection layers. This vector encapsulates key structural information such as the connection patterns and pore distribution morphology of the fabric fiber network.

[0056] Step S12: Path B: Collect the fabric density of the target fabric using a fabric density meter, thickness gauge, air permeability tester, and near-infrared spectrometer. ,thickness Gram per square meter Breathability and near-infrared reflectance spectral matrix Where L is the number of wavelengths and M is the number of sampling points, forming a set of macroscopic physical parameters. ;Will Input structure encoding mapping model Output the predicted structural feature encoding vector The structure encoding mapping model The training samples were obtained through supervised training. Group basic fabric corresponding Mapping pairs, where Generated from path A.

[0057] Specifically, this approach is applicable to rapid inspection scenarios on production lines. First, using conventional textile inspection equipment, macroscopic parameters of the target fabric are non-destructively collected: fabric density ρ (unit: threads / 10 cm), thickness h (unit: mm), weight per square meter W (unit: g / m²), air permeability Q (unit: mm / s), and an L×M dimensional near-infrared reflectance spectral matrix S (L is the number of wavelengths, M=32 sampling points) collected in the 900-1700 nm wavelength range. These parameters are then flattened and stitched together to form a set of macroscopic parameters. Subsequently, Input a pre-trained structure encoding mapping model The model is a three-layer fully connected neural network, and its training data comes from a large number of basic cloth samples. Each sample simultaneously possesses a "truth value" encoding generated through path A. and its corresponding macroscopic parameters The model establishes a robust mapping from easily measurable macroscopic parameters to microscopic structure encodings through supervised learning (such as minimizing mean squared error loss). Therefore, for new fabrics, equivalent microstructural feature encoding vectors can be quickly obtained through conventional detection without complex imaging. This significantly improves the engineering applicability of the method.

[0058] Through step S1, whether using the high-fidelity path A or the high-efficiency path B, a unified and quantitative digital identity card of the microstructure can be obtained for any target fabric (i.e., This allows the intrinsic physical properties of the fabric to be incorporated into the subsequent computational optimization framework, solving the fundamental problem of traditional methods that design formulations without considering the material structure.

[0059] Step S2: Obtain the physicochemical property data of the candidate antibacterial agents and deodorants, as well as the loading process parameter data to be optimized.

[0060] Specifically, step S2 aims to digitize the chemical reagents and process conditions. For candidate antibacterial agents (such as nano-silver and chitosan quaternary ammonium salts) and deodorants (such as activated carbon microspheres and β-cyclodextrin), key physicochemical properties are extracted, such as molecular weight, logP (octanol-water partition coefficient), water solubility, Zeta potential, particle size, and specific surface area, forming a 64-dimensional property vector for each type of reagent. Simultaneously, process parameters to be optimized are defined, such as impregnation time t (range: 30-300 seconds), impregnation temperature T (range: 25-80 degrees Celsius), pH value, and drying conditions, forming a 6-dimensional process parameter vector. All these data are standardized to ensure numerical compatibility with the structure-encoded vector.

[0061] Step S3: Combine the microstructure feature encoding vector, the physicochemical property data of the selected functional agent, and the process parameter data to form comprehensive input data.

[0062] Specifically, step S3 concatenates the three data parts generated in steps S1 and S2 in a fixed order. The concatenation order is: a 128-dimensional microstructure feature encoding vector + a 64-dimensional antibacterial agent attribute vector + a 64-dimensional deodorant attribute vector + a 6-dimensional process parameter vector, ultimately forming a comprehensive input vector with a total dimension of 262. This splicing method preserves the independence of each modality's information, providing a foundation for the model to learn the complex interactions between them.

[0063] Step S4: Input the comprehensive input data into the trained multi-task performance prediction model, which will then output the antibacterial rate, deodorization durability, and skin irritation index values ​​simultaneously.

[0064] Specifically, the core of this step is a pre-trained multi-task neural network model, the structure of which is as follows: Figure 2 As shown. The model receives the 262-dimensional integrated input vector generated in step S3. It can predict three key performance indicators at once.

[0065] Specifically, the multi-task performance prediction model includes a shared feature encoding module and three parallel task output heads.

[0066] Step S41: Shared Feature Encoding Module: The shared feature encoding module receives the concatenated vector xin and outputs high-level semantic features. .

[0067] Furthermore, the specific implementation of the shared feature encoding module in step S41 is as follows:

[0068] like Figure 3 As shown, the shared feature encoding module is implemented using a three-layer fully connected network, and its layer l output is:

[0069] ,

[0070] in , , For Gaussian error linear units, It is the standard normal cumulative distribution function.

[0071] Specifically, the module consists of three fully connected layers with hidden layer dimensions of 512, 256, and 128, respectively. Each layer performs a linear transformation on the input followed by a non-linear mapping using the GeLU activation function. The GeLU function is smooth and non-monotonic, better handling complex patterns in the input data. Finally, the module outputs a 128-dimensional high-level shared semantic feature vector. It integrates the abstract characteristics resulting from the interaction of fabric structure, functional agent properties, and process conditions.

[0072] Step S42 (Parallel Task Output Headers): Each task output header is an independent fully connected sub-network, outputting the following respectively:

[0073] Antibacterial rate prediction ,in It is the Sigmoid activation function. , For learnable bias terms;

[0074] Odor retention prediction value ,in It is a linear rectifier unit. for For learnable bias terms;

[0075] Predicted Skin Irritation Index ,in , This is a learnable bias term.

[0076] Specifically, the three task heads operate independently. The antibacterial rate head uses the Sigmoid function to constrain the output to the (0,1) interval, then multiplies it by 100% to obtain the percentage prediction. The deodorization persistence head uses the ReLU function to ensure the output is non-negative, typically in hours. The skin irritation index head uses the Softplus function to ensure the output is always positive and to simulate the trend of irritation intensity changing with increasing dose. Through this design, a single forward propagation of the model can efficiently and synchronously obtain quantitative predictions of the three performance metrics.

[0077] In a preferred embodiment, the training process of the multi-task performance prediction model employs a carefully designed loss function and regularization strategy:

[0078] The multi-task performance prediction model uses a weighted multi-task loss function during the training phase:

[0079]

[0080] in , , , , .

[0081] Specifically, the model uses a large number of known performance labels. The sample pairs are trained end-to-end. Total loss. It is the weighted sum of the mean squared error losses of the three prediction tasks. The weight ratio of 1:1.2:0.8 is set based on the importance of each indicator in practical applications and the characteristics of data distribution: giving a slightly higher weight to deodorization persistence to cope with its larger prediction variance, while balancing the learning progress of the three to ensure that the model has a good fit to all tasks.

[0082] In another preferred embodiment, in order to improve the structural feature encoding generated by the graph neural network encoder in path A... To enhance robustness and representational ability, additional self-supervised learning constraints are introduced during the training phase:

[0083] During the training phase of the multi-task performance prediction model, a graph structure contrastive learning regularization term is also introduced, specifically: for two different graph augmentation views of the same fabric sample... and Its corresponding structural feature encoding , The similarity constraint should be satisfied, and the contrastive loss function is:

[0084] ,

[0085] in For cosine similarity, For temperature coefficient, This is an indicator function.

[0086] Specifically, during training, two types of data augmentation (such as randomly masking some nodes or edges) are randomly applied to the graph structure G of the same fabric to obtain two views G1 and G2, which are then encoded separately. and Comparative losses The goal is to maximize positive sample pairs ( , The regularization calculates the similarity between the samples and the encodings of other samples in the same batch (negative samples), while minimizing their similarity to the encodings of other samples in the same batch (negative samples). This regularization enables the encoder to learn robust features that are insensitive to local perturbations and better capture the essential structural information of the fabric, thereby indirectly improving the accuracy of subsequent multi-task predictions.

[0087] Step S4 establishes an efficient and accurate surrogate model that can predict the three properties of any combination of "fabric-functional agent-process" at millisecond speeds, replacing time-consuming and labor-intensive physical experiments and laying the foundation for subsequent large-scale optimization searches.

[0088] Step S5: Using the selection of functional agents, the mass ratio of each functional agent, and the specific values ​​of process parameters as decision variables, construct a multi-objective optimization problem. The optimization objective is to maximize the antibacterial rate and the deodorization persistence value, and the constraint is that the skin irritation index value must not exceed the preset safety threshold.

[0089] Specifically, step S5 formalizes the actual optimization problem into a standard constrained multi-objective optimization problem. The decision variables x include: selecting the types of antibacterial agents and deodorizers from a pre-set library (discrete variables), setting the mass ratio r of the two (continuous variables, e.g., antibacterial agent proportion 30%-70%), and setting specific process parameter values ​​(6 continuous variables). The optimization objectives are twofold: maximizing the predicted antibacterial rate. And maximize the prediction of deodorization durability At the same time, a hard constraint is introduced: predicting the skin irritation index. It must be less than or equal to a preset safety threshold τ (e.g., τ=1.5). This threshold is calibrated according to in vitro skin irritation testing standards (such as OEC DTG 439) to ensure that all feasible solutions meet the basic safety requirements.

[0090] Step S6: The optimization problem is solved using a multi-objective evolutionary algorithm with constraint handling capabilities. Under the premise that the skin irritation index value does not exceed the preset safety threshold, a set of non-dominant Pareto optimal candidate solutions are obtained.

[0091] Specifically, this step uses a modified NSGA-III algorithm as the solver. The core of the algorithm is how to strictly handle the constraint of skin irritation during the evolution process.

[0092] Specifically, the core mechanism of constraint handling in step S6 is as follows:

[0093] The multi-objective evolutionary algorithm defines the individual constraint violation degree in constraint processing. Defined as:

[0094]

[0095] During the non-dominated sorting process, all individuals that satisfy CV(x)>0 are forced to be assigned to the worst non-dominated rank, and subsequent stratification and crowding distance calculations are performed only in the feasible subset of solutions where CV(x)=0.

[0096] Specifically, for each candidate solution (individual) x in the population, its constraint violation degree is first calculated. .if If CV(x) = 0, it is a feasible solution; otherwise, CV(x) > 0, it is an infeasible solution. In the non-dominated ranking of each generation, all infeasible solutions with CV(x) > 0 are uniformly classified into the worst level, thus being prioritized for elimination in survival selection. Ranking and diversity maintenance (such as reference point-based selection) are only performed on feasible solutions. This "feasibility-first" principle ensures that the evolutionary search is always guided to a decision space region that satisfies safety constraints, and eventually converges to the Pareto optimal frontier (i.e., a set of candidate solutions that are non-dominant in antibacterial and deodorizing performance and are all safe) within that region.

[0097] Step S7: Select the target scheme from the Pareto optimal candidate schemes, output the corresponding functional agent combination type, the mass ratio of each functional agent, the specific values ​​of the process parameters, and generate a visualization result of the spatial load distribution of functional agents on the fabric microstructure of the scheme.

[0098] Figure 4 A block diagram of a functional agent spatial load distribution prediction model based on graph attention networks is presented. Specifically, step S6 outputs a Pareto optimal solution set, rather than a single solution. Users can select a final solution from the solution set according to their actual needs (such as a greater emphasis on antibacterial or deodorizing). The system will output the specific configuration of the solution: for example, the antibacterial agent is "nano-silver", the deodorizing agent is "activated carbon microspheres", the mass ratio is 1:1, and the impregnation process is "temperature 50℃, time 120 seconds, pH=7", etc.

[0099] Furthermore, step S7 also includes an innovative visualization and precise execution step:

[0100] The visualization result of the spatial load distribution of the functional agents is a heat map. The heatmap H is generated by inputting the decision variables corresponding to the Pareto optimal solution into a graph attention model trained with interpretability. This model uses a graph structure. Given input, output the functional agent loading probability at each node vi∈V. And mapped to the pixel grid according to the node spatial coordinates, where Calculate using the following formula:

[0101]

[0102] in For graph neural networks Layer node embedding, For nodes The set of first-order neighbors, For graph attention aggregation function, , The grid resolution of the heatmap is [value], and ≥256, the heat map is used to drive a six-axis robotic arm to control the nozzle to perform gradient spraying according to a probability distribution.

[0103] Specifically, to achieve precise load loading at the fiber level, this method introduces an interpretable graph attention model. This model takes the fabric microstructure G constructed in step S1 and the selected decision variables in step S7 as input. After computation by the graph attention network, the model is defined for each node in graph G. (Representing a micro-region) Predict the probability of a functional agent loading. The probability depends on the local structure at that location (e.g., areas with high curvature and many pores have a higher probability). Subsequently, the probability values ​​of these discrete nodes are interpolated based on their actual spatial coordinates to generate a high-resolution (e.g., 512×512) heatmap H. Brighter areas in the heatmap indicate a higher recommended concentration of functional agent. This heatmap can be directly converted into control commands for a six-axis robotic arm, driving a micro-nozzle to perform non-uniform "gradient spraying" on the fabric surface. This precisely translates the ideal distribution optimized in the digital world onto the physical fabric, maximizing functional agent utilization efficiency and ensuring performance uniformity.

[0104] In summary, through steps S1 to S7, this invention constructs a complete intelligent closed loop, from microstructure characterization to macroscopic performance prediction, then to multi-objective automatic optimization under safety constraints, and finally to precise visualization and execution. This method completely changes the outdated model of traditional textile functionalization R&D, which relies on trial and error, ignores pre-existing safety constraints, and cannot quantify and weigh multiple performance aspects, significantly improving R&D efficiency, safety, and overall product performance.

[0105] To verify the accuracy of the multi-task performance prediction model, it was evaluated on a test set containing 1000 samples. Figure 5 Scatter plots comparing the model's predicted values ​​for antibacterial rate, deodorization durability, and skin irritation index with their actual experimental measurements are presented. The results show that the root mean square errors for the three prediction tasks are 2.1%, 4.7 hours, and 0.13, respectively, with coefficients of determination all higher than 0.93. All data points are closely distributed on both sides of the diagonal, demonstrating that the model constructed using this method can simultaneously predict the three key performance indicators with high accuracy, providing a solid foundation for reliable optimization under subsequent safety constraints.

[0106] The foregoing has shown and described the basic principles, main features, and advantages of this invention. Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A multi-objective optimization method for antibacterial and deodorizing fabrics under safety constraints, characterized in that, Includes the following steps: Obtain microstructural feature encoding vectors that can quantitatively characterize the fiber topology and pore network distribution of the target fabric; Obtain the physicochemical property data of the candidate antibacterial agents and deodorants, as well as the loading process parameters to be optimized; The microstructure feature encoding vector, the physicochemical property data of the selected functional agent, and the process parameter data are combined to form comprehensive input data; The comprehensive input data is input into the trained multi-task performance prediction model, which then outputs the antibacterial rate, deodorization duration, and skin irritation index values ​​simultaneously. Using the selection of functional agents, the mass ratio of each functional agent, and the specific values ​​of process parameters as decision variables, a multi-objective optimization problem is constructed. The optimization objective is to maximize the antibacterial rate and the deodorization duration, and the constraint is that the skin irritation index must not exceed the preset safety threshold. The optimization problem is solved using a multi-objective evolutionary algorithm with constraint handling capabilities. Under the premise that the skin irritation index value does not exceed a preset safety threshold, a set of non-dominant Pareto optimal candidate solutions are obtained. Select the target scheme from the Pareto optimal candidate schemes, output the corresponding functional agent combination type, the mass ratio of each functional agent, the specific values ​​of the process parameters, and generate a visualization result of the spatial load distribution of functional agents in the fabric microstructure of the scheme.

2. The method according to claim 1, characterized in that, The microstructure feature encoding vector is obtained through any of the following paths: Path A: Image the target fabric using a scanning electron microscope or micro-CT device to obtain two-dimensional / three-dimensional image data with sub-micron resolution; perform image segmentation and skeleton extraction on the image data to construct a graph structure. , where the node set Characterizing fiber intersections or pore geometric centers, where n is the total number of nodes and edge sets. Characterize the fiber connections or pore channel connectivity; [The diagram structure is then described]. Input a graph neural network encoder, output a fixed-dimensional structural feature encoding vector. ; Path B: Collect the fabric density of the target fabric using a fabric density meter, thickness gauge, air permeability tester, and near-infrared spectrometer. ,thickness Gram per square meter Breathability and near-infrared reflectance spectral matrix Where L is the number of wavelengths and M is the number of sampling points, forming a set of macroscopic physical parameters. ;Will Input structure encoding mapping model Output the predicted structural feature encoding vector The structure encoding mapping model The training samples were obtained through supervised training. Group basic fabric corresponding Mapping pairs, where Generated from path A.

3. The method for path A according to claim 2, characterized in that, The graph structure Middle node initial feature vector Constructed from its local geometric properties, including: the curvature at the location of the node. Mean pore area within the neighborhood and fiber diameter statistical moments ,Right now ;side initial feature vector Includes connection length Equivalent radius of cross section .

4. The method according to claim 1, characterized in that, The multi-task performance prediction model includes a shared feature encoding module and three parallel task output headers; the shared feature encoding module receives the concatenated vector. And output high-level semantic features Each task's output header is an independent fully connected subnet, and it outputs the following respectively: Antibacterial rate prediction ,in It is the Sigmoid activation function. , For learnable bias terms; Odor retention prediction value ,in It is a linear rectifier unit. for For learnable bias terms; Predicted Skin Irritation Index ,in , This is a learnable bias term.

5. The method according to claim 4, characterized in that, The shared feature encoding module is implemented using a three-layer fully connected network, the third of which... The layer output is: in , , For Gaussian error linear units, It is the standard normal cumulative distribution function.

6. The method according to claim 1, characterized in that, The multi-task performance prediction model uses a weighted multi-task loss function during the training phase: in , , , , .

7. The method according to claim 1, characterized in that, The multi-objective evolutionary algorithm defines the individual constraint violation degree in constraint processing. Defined as: In the non-dominated sorting process, all satisfying Individuals are forcibly relegated to the worst non-dominated class, only... The subsequent stratification and congestion distance calculations are performed on the feasible solution subset.

8. The method according to claim 1, characterized in that, During the training phase of the multi-task performance prediction model, a graph structure contrastive learning regularization term is also introduced, specifically: for two different graph augmentation views of the same fabric sample... and Its corresponding structural feature encoding , The similarity constraint should be satisfied, and the contrastive loss function is: in For cosine similarity, For temperature coefficient, This is an indicator function.

9. The method according to claim 1, characterized in that, The visualization result of the spatial load distribution of the functional agents is a heat map. The heatmap H is generated by inputting the decision variables corresponding to the Pareto optimal solution into a graph attention model trained with interpretability. This model uses a graph structure. For input, output each node Functional agent loading probability And mapped to the pixel grid according to the node spatial coordinates, where Calculate using the following formula: in For graph neural networks Layer node embedding, For nodes The set of first-order neighbors, For graph attention aggregation function, , The grid resolution of the heatmap is [value], and ≥256, the heat map is used to drive a six-axis robotic arm to control the nozzle to perform gradient spraying according to a probability distribution.

Images