A method for synergistic optimization of anti-fouling stability and safety of a trash rack anti-fouling material

Abstract

This invention belongs to the field of marine engineering antifouling material technology and artificial intelligence optimization technology. It discloses a collaborative optimization method for the antifouling stability and safety of antifouling materials used in debris-blocking nets, comprising the following steps: Step 1, collecting multi-dimensional data and preparing samples of the antifouling materials for the debris-blocking nets; Step 2, constructing and training an AI prediction model: selecting key feature parameters from the multi-dimensional data, using these key feature parameters as core feature vectors to construct a dual-objective prediction model, which outputs predicted values ​​for the material's antifouling stability index and environmental safety index; Step 3, collaborative optimization under multiple constraints: setting multiple constraints for the material; using the hybrid meta-heuristic algorithm ALO-KHO to solve the multi-objective optimization problem; selecting the optimal solution from the Pareto optimal solution to obtain the parameter combination of the antifouling materials for the debris-blocking nets with the highest score. The prediction model of this invention has high accuracy, strong generalization ability, and significant multi-objective collaborative optimization effect.

Description

Technical Field

[0001] This invention relates to the fields of marine engineering antifouling materials technology and artificial intelligence optimization technology, and in particular to a method for synergistic optimization of the antifouling stability and safety of antifouling materials for debris barriers. Background Technology

[0002] The trap screen at the intake of the nuclear power plant's cold source (such as...) Figure 1 (b) is a key piece of equipment ensuring the safe operation of nuclear power plants. However, due to its long-term immersion in the marine environment, it is susceptible to contamination by marine organisms such as jellyfish, shrimp, and algae, leading to decreased water intake efficiency, accelerated equipment corrosion, and in severe cases, shutdown accidents. Therefore, the antifouling material of the contaminant screen (such as...) Figure 1 The antifouling stability (such as antifouling cycle and coating abrasion resistance) of (a) is a core technical requirement.

[0003] Meanwhile, the ecological sensitivity of the marine environment places stringent requirements on the environmental safety of antifouling materials. The release and residue of nano-metal ions from these materials must meet environmental standards to avoid negative impacts on marine biodiversity. Traditional antifouling material optimization often relies on trial and error or single-objective optimization, which has the following shortcomings: 1. The performance data of antifouling materials (antifouling cycle, metal ion release rate, etc.) and environmental parameters (ocean current velocity, water quality composition, etc.) show significant cross-domain differences. The distribution gap between the source domain (known material data) and the target domain (new formulation material data) leads to low prediction accuracy of conventional machine learning models. Furthermore, the sample preparation cost of new formulation materials is high and the number is small, resulting in insufficient generalization ability of the model in scenarios with few samples. 2. The optimization process needs to simultaneously meet multiple constraints such as anti-fouling stability, environmental safety, mechanical performance and economic cost. Traditional optimization algorithms (such as single heuristic algorithms) are prone to getting trapped in local optima, making it difficult to balance global exploration and local development, and unable to efficiently find Pareto optimal solutions under multiple constraints. 3. Existing technologies have not achieved a collaborative closed loop of "prediction-optimization". The prediction model has not fully considered the collaborative problem of domain adaptation and few-sample classification. The optimization algorithm has not been deeply integrated with the high-precision prediction model, resulting in insufficient practicality and reliability of the optimization scheme.

[0004] Therefore, there is an urgent need to develop a collaborative optimization method for antifouling materials that can adapt to the characteristics of small sample sizes and cross-domain data, and can efficiently balance multiple constraints, so as to solve the problem of collaborative optimization between antifouling stability and environmental safety. Summary of the Invention

[0005] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a method for synergistic optimization of the anti-fouling stability and safety of anti-fouling materials for pollution-blocking nets.

[0006] The technical solution adopted to achieve the purpose of this invention is: A method for synergistically optimizing the antifouling stability and safety of antifouling materials for pollution barriers includes the following steps: Step 1, Multi-dimensional data acquisition and sample preparation: The multi-dimensional data includes a biological-environmental parameter correlation database, a material parameter database, and a material performance-environmental safety database. The samples are used for data verification. Step 2, AI Prediction Model Construction and Training: Key feature parameters are selected from the multi-dimensional data collected in Step 1. These key features include antifouling stability features, environmental safety features, material structure features, and environmental operating condition features. Using these key feature parameters as the core feature vector, a bi-objective prediction model based on a domain adversarial prototype network is constructed. After model training and optimization, the predicted value of the material's antifouling stability index is output. Predicted values ​​of environmental safety indicators ; Step 3, Collaborative optimization under multiple constraints: Step s31: Set multiple constraints for the material; Step s32: Use the hybrid meta-heuristic algorithm ALO-KHO to perform multi-objective optimization. The solution process is as follows: First, generate N initial solutions, each corresponding to a combination of optimization variables. Then, calculate the updated solutions by sequentially updating using ALO and KHO. Substitute the updated solutions into the bi-objective prediction model trained in step 2 to obtain the final solution. and Combined with mechanical performance scores and economic cost score Calculate the fitness of each updated solution, retain the N solutions with the highest fitness as the next generation population, and iterate in this way until convergence is achieved to obtain the Pareto optimal solution; Step s33: Select the optimal solution from the Pareto optimal solution to obtain the parameter combination of the antifouling material of the debris barrier with the highest score.

[0007] In the above technical solution, in step 1, the multi-dimensional data includes the source domain dataset. With the target domain few-sample dataset and to Expand to obtain the target domain sample set .

[0008] In the above technical solution, step 2, the method for constructing the bi-objective prediction model includes the following steps: Step s21, Constructing the base layer: The base layer uses ResNet18 as the backbone network to transform the core feature vectors into 512-dimensional initial feature vectors. ; Step s22: Construct a domain-adaptive embedding module, which consists of an autoencoder and an attention submodule connected in series. The autoencoder is based on... As input, the encoder and decoder output a 256-dimensional embedding vector. Attention submodule and As input, the output is a 256-dimensional domain confusion feature. ,in, This refers to the fully connected layer operation within the attention submodule, where sigmoid is the activation function. Step s23: Construct a class discrimination layer with the episode training mechanism as its core: from the source domain dataset Source domains are constructed by selecting different material categories. Divided into support sets With query set From the target domain sample set Selecting different material categories to construct the target domain Similarly, it is divided into support sets. With query set For each and For each sample in the dataset, calculate its prototype vector. ;right and For each sample in the dataset, calculate its domain confusion features. With all category prototype vectors The Euclidean distance is converted into class probabilities using the softmax function, and the prediction result is output. This includes the antifouling stability index and environmental safety index corresponding to the sample.

[0009] Step s24, construct the domain alignment enhancement layer: receive the domain confusion features from step s22. The prediction results of step s23 Furthermore, the parameters of the category discrimination layer and the domain adaptive embedding module are optimized inversely through the loss function to form a closed loop.

[0010] In the above technical solution, in step s23, , To support the centralized A sample set of similar materials To support the number of samples in each class, For the first Domain confusion features of each sample.

[0011] In the above technical solution, the prediction formula for the category probability in step s23 is: ; in, For query samples Belongs to the The probability of a class This is the function for calculating Euclidean distance.

[0012] In the above technical solution, in step s24, the total loss of the loss function The calculation formula is: ; in, For the source domain prototype loss, For the target domain prototype loss, Loss due to domain confusion To differentiate the loss between domains, , , and These are adaptive weights, , =1, 2, 3, or 4, For uncertain parameters; The calculation formula is: ; ; in, For the first One sample, For domain discriminator, This is a dot product feature fusion function based on dimensionality reduction of random matrices. For the Hadamard product operator, A randomly generated fixed matrix with dimension . , A randomly generated fixed matrix with dimension . , The feature dimensions after dimensionality reduction. Number of categories; The calculation formula is: , For domain discrimination; The calculation formula is: ; The calculation formula is: ; This represents the total number of samples in the query set within the source domain episode; The total number of samples in the query set within the target domain episode; For the sample Corresponding category The true label unique hot encoding.

[0013] In the above technical solution, step 2, the model training and optimization steps are as follows: Will The dataset is divided into training, validation, and test sets. The SGD optimizer is used to optimize the parameters of the bi-objective prediction model, and the mean squared error (MSE) on the validation set is used as the convergence metric.

[0014] In the above technical solution, the constraints in step s31 include technical constraints, performance constraints, safety constraints, and economic constraints. In step s32, the combination of optimization variables includes the type of nano-metal composition, nanoparticle size, composition ratio, deposition temperature, reaction time, and coating thickness. In the above technical solution, the ALO update process in step s32 is as follows: ; in, For the first The generation The solution of the first... The value of the variable, the first The solution is the first A combination of material parameters, Index for iteration count, For the index of the solution, For the index of the variable, For the first The lower bound of the values ​​of each optimization variable. For the first The upper limit of the values ​​of each optimization variable. For the first The optimal solution (antlion) in the population of generations There are several variable values, where rand is a uniformly random number in the interval 0-1; The KHO update process is as follows: ; in, For the first The generation The solution of the first... The speed of each variable For inertial weights, For the first The generation The solution of the first... The speed of each variable As a learning factor, For the first The solution with the highest fitness in the population is the first One variable value, For the first The generation The solution of the first... One variable value, For the first All solutions of the generation population The average of the variables.

[0015] In the above technical solution, the formula for calculating fitness in step s32 is: ; in, For fitness, To score for antifouling stability, To score environmental safety For mechanical performance scores, Score the economic cost. To constrain the penalty coefficient for violations, For the first A function representing the degree of violation of a constraint. For the current solution, The total number of constraints. To constrain the effective penalties for violations, when hour, Output 0 when hour, Output .

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention proposes a dual-objective prediction model based on Domain Adaptive Few-Shot Learning (DA-FSL), with a customized domain adversarial prototype network at its core, to solve the collaborative problem of "domain adaptation" and "few-shot classification". This model narrows the distribution gap between the source and target domains, accurately predicting antifouling stability indicators (antifouling cycle, coating wear rate) and environmental safety indicators (metal ion residual concentration, biotoxicity inhibition rate) in few-shot scenarios for new formulation materials, providing reliable data support for subsequent optimization. 2. This invention proposes a hybrid meta-heuristic algorithm (ALO-KHO), which integrates the global exploration capability of antlion optimization (ALO) with the local exploitation capability of krill swarm optimization (KHO) to solve the problem of collaborative optimization of material parameters under multiple constraints. Through a dual-process update strategy, it efficiently searches for the optimal combination of nano-metal composition, particle size, composition ratio, and preparation process parameters under multiple constraints such as technology, performance, safety, and economy. 3. This invention constructs a complete technical closed loop of "data acquisition - sample preparation - model prediction - multi-constraint optimization - scheme screening", which realizes the synergistic improvement of the antifouling stability and environmental safety of antifouling materials, while taking into account mechanical properties and economy, and provides the optimal material formula and process parameters that can be directly used for industrial production; 4. The prediction model of this invention has high accuracy and strong generalization ability: The customized domain adversarial prototype network proposed in this invention solves the cross-domain difference problem between the source domain and the target domain through a closed-loop design of "feature extraction - domain adaptation - category discrimination - loss balancing", and is also adapted to the low-sample scenario of new formulation materials. Compared with traditional machine learning models, the prediction error of this model on the validation set is less than 8%, and the error difference between the test set and the validation set is less than 2%. The prediction accuracy of key indicators such as metal ion residual concentration and anti-fouling cycle is significantly improved, providing a reliable decision basis for optimization. 5. The optimization algorithm is efficient and robust: The hybrid metaheuristic algorithm (ALO-KHO) avoids the pitfalls of single algorithms being prone to getting trapped in local optima by combining the global exploration of ALO with the local development of KHO. It converges within 3000 iterations, which is more than 30% faster than traditional optimization algorithms (such as single PSO and GA algorithms). It can also stably output a set of Pareto optimal solutions that satisfy all constraints, significantly improving optimization efficiency and solution reliability. 6. Significant Multi-Objective Synergistic Optimization Results: This invention achieves synergistic optimization of antifouling stability and environmental safety. The optimized antifouling material has an antifouling cycle of ≥180 days, a coating wear rate of ≤5%, a metal ion residual concentration of ≤0.5mg / L, and a biotoxicity inhibition rate of ≤10%, fully meeting marine environmental protection standards and performance requirements. Furthermore, the optimized scheme also considers mechanical properties (tensile strength ≥1.0MPa, fatigue resistance ≥10). 4 In terms of both cost-effectiveness (preparation cost increase ≤15%), its overall performance is superior to existing technologies; 7. High practicality and wide application scenarios: The present invention is based on the database constructed from real sea test and indoor simulation data. The optimized material formula and process parameters can be directly used for industrial production. Moreover, the model and algorithm have good scalability and can be adapted to the antifouling material optimization needs of different marine environments (such as ports, ships, etc.), with broad application prospects. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the coating applied manually to long-lasting mesh, where (a) is the antifouling material and (b) is the mesh after the antifouling material has been applied.

[0018] Figure 2 This is a flowchart of a method for synergistically optimizing the antifouling stability and safety of antifouling materials for pollution-blocking nets.

[0019] Figure 3 This is a graph comparing algorithm performance. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] like Figure 2 As shown, a method for synergistically optimizing the antifouling stability and safety of antifouling materials for pollution barriers includes the following steps: Step 1: Multi-dimensional data collection and sample preparation.

[0022] Step 1.1, collect basic data from multiple dimensions.

[0023] Typical marine organisms (jellyfish, shrimp, algae, etc.) from the intake of a nuclear power plant's cold source were selected. Through indoor simulation experiments and on-site monitoring, data on the characteristics of organism mortality under different environmental conditions were collected, including water quality factors (Fe). 2+ Cu 2+ Zn 2+ A database linking marine organisms and environmental parameters was established using data on metal ion concentration, pH value, dissolved oxygen, hydrological elements (ocean current velocity, wave height, wave period), and changes in biological morphology and physiological function.

[0024] Collect core parameter data of nano-antifouling materials, including the composition, micro-nano structure size (10-100nm), surface energy, and metal ion release rate of Cu-based, Zn-based, Ti-based and other nano-metals and their compounds, as well as the mechanical properties (strength, toughness, fatigue resistance) and dispersion characteristics of the coating matrix (functional hydrogels, epoxy resins, etc.), and construct a material parameter database.

[0025] Collect data from marine netting tests and field exposure tests, including the antifouling cycle, coating wear rate, and substrate adhesion strength of antifouling materials under different working conditions, as well as data on the residual concentration of metal ions in the seawater environment and the impact on marine biodiversity, and establish a material performance-environmental safety database.

[0026] The above three types of databases together constitute the source domain dataset for model training. With the target domain few-sample dataset ,in Includes sufficient labeled samples (various parameters fully match anti-fouling stability and safety labels). For the target domain, there are a few sample labeled data for different material combinations (each material combination contains 5-10 labeled samples).

[0027] Step 1.2, Gradient sample preparation.

[0028] Based on a material parameter database, multiple gradient samples were designed. Through different preparation processes such as vapor deposition and hydrothermal synthesis, the nanoparticle size (10-100nm), material composition ratio (nano-metal content 5%-20%), and micro-nano structure morphology (porous, layered) were controlled to prepare antifouling net material samples with different combinations of antifouling components and matrix.

[0029] Each sample group was configured with three parallel samples to ensure sample consistency for subsequent performance testing and data validation. Data augmentation techniques (including parameter perturbation and environmental condition simulation extension) were used to... The sample set of the target domain is expanded to obtain an enhanced sample set. This is to meet the training requirements of domain-adaptive few-shot learning.

[0030] Step 2: AI prediction model construction and training.

[0031] Step 2.1, Feature Engineering.

[0032] Key characteristic parameters were selected from the collected multi-dimensional data, including antifouling stability characteristics (metal ion release stability, coating wear resistance, antifouling cycle), environmental safety characteristics (metal ion residue, pH influence range, biotoxicity threshold), material structure characteristics (nanoparticle dispersion, component ratio, microstructure), and environmental conditions characteristics (flow rate, wave height, biomass density).

[0033] Normalization was used to eliminate differences in feature dimensions, and Pearson correlation analysis was used to remove redundant features (correlation coefficient |r|). (One feature is retained for a feature size of 0.8). Combined with principal component analysis for dimensionality reduction, core feature variables with a cumulative variance contribution rate of ≥90% are retained, thereby reducing the computational complexity of the model.

[0034] Step 2.2: Construct a bi-objective prediction model.

[0035] A dual-objective prediction model is constructed based on domain-adaptive few-shot learning, with the Domain Adversarial Prototype Network (DAPN) at its core. The core design logic of the dual-objective prediction model is to solve the synergistic problem of "domain adaptation" and "few-shot classification"—to narrow the distribution gap between the source domain (known material data) and the target domain (new formulation material data), while ensuring the distinguishability of different material categories and avoiding category confusion caused by domain alignment. The modules do not work independently, but form a closed loop through a series of logic: "feature extraction → domain adaptation → category discrimination → loss balancing".

[0036] Step 2.2.1, construct the base layer.

[0037] As the feature input foundation of the dual-objective prediction model, it is responsible for transforming the preprocessed M-dimensional core feature vector into high-dimensional semantic features. ResNet18 is used as the backbone network, containing four residual blocks (each residual block contains two convolutional layers + one residual connection), and the activation function is uniformly ReLU (to alleviate the gradient vanishing problem). The input is... The one-dimensional feature vector (corresponding to the fusion features of material parameters and environmental parameters) outputs as a 512-dimensional initial feature vector. The output of this foundational layer serves as the common input for all subsequent sub-modules, providing fundamental feature support for domain adaptation and category discrimination.

[0038] in: M is the dimension of the core feature vector, which is determined by the number of core features retained after screening in the feature engineering stage (such as the total number of material structure features, environmental condition features, etc.). The dimensional form of the input feature vector is a one-dimensional vector containing M core features, corresponding to the feature set after fusing the parameters of the anti-fouling material of the debris barrier with environmental parameters; ResNet18 is the backbone network model, an 18-layer deep residual network used to extract high-dimensional semantic features. It contains 4 residual blocks and is the core network structure for feature extraction. The residual block is the basic structural unit in ResNet18. Each residual block contains two convolutional layers and one residual connection, which is used to alleviate the gradient vanishing problem in deep network training and improve feature extraction capabilities. Convolutional layers are feature transformation layers in residual blocks. They perform dimensional transformation and information integration on input features through convolution operations, providing adaptive feature dimensions for residual connections. Residual connections are cross-layer connection structures in residual blocks that directly pass input features to the output and superimpose them with the output features of the convolutional layer, preserving the original feature information and assisting in gradient backpropagation. ReLU is an activation function, short for Rectified Linear Unit, used to alleviate the gradient vanishing problem in deep network training and enhance the nonlinear expressive power of the model. The original feature vector input to the backbone network, i.e. Dimensional material-environment fusion feature vector (such as the normalized combination of parameters like nanoparticle scale and ocean current velocity); is the feature mapping function of the backbone network, which represents the operation process of extracting features from the input vector x through ResNet18 and outputting high-dimensional initial features; The output of the backbone network is a 512-dimensional initial feature vector containing high-dimensional semantic information related to materials and environment, which is the basic feature for subsequent domain adaptation and category discrimination.

[0039] Step 2.2.2: Construct the domain adaptive embedding module.

[0040] Following step 2.2.1, the initial feature vector of the backbone network is 512. The autoencoder is the core module for achieving "domain alignment," consisting of an autoencoder and an attention submodule connected in series. These two modules work together to perform "feature compression and domain information filtering." First, the autoencoder learns and extracts a 256-dimensional embedding vector with domain invariance through a structure of "encoder (2 FC layers: 512→384→256) + decoder (2 FC layers: 256→384→512)." This ensures that the features have consistent semantic representation in both the source and target domains; subsequently, the attention submodule passes through one fully connected (FC) layer (inputting 512-dimensional initial features). The output (a 256-dimensional attention score) is combined with a sigmoid activation function to generate attention weights in the 0-1 range, which are then applied to the autoencoder output. Perform element-wise weighted summation, the formula is as follows: Its core function is to filter out domain-specific information (such as feature noise caused by differences in experimental environments between the source and target domains) from the initial features, retain common information that is effective in predicting the antifouling stability and safety of materials, and finally output 256-dimensional domain confusion features. This helps eliminate interference from domain disparities in subsequent small-sample classification.

[0041] in: 512 is the dimension of the initial feature vector output by the backbone network, i.e. The feature dimensions are determined by the network architecture design of ResNet18; The autoencoder is the core substructure of the domain adaptive embedding module, consisting of an encoder and a decoder. It is used to learn domain-invariant features and achieve semantic alignment between the source domain and the target domain features. The encoder is a feature compression unit of an autoencoder, containing two fully connected layers (FC layers). The dimensionality transformation path is 512→384→256, transforming the 512-dimensional initial features output by the backbone network. Compressed into 256-dimensional embedded features; The decoder is a feature reconstruction unit of the autoencoder, containing two fully connected layers (FC layers). The dimensionality transformation path is 256→384→512, which reconstructs the 256-dimensional embedded features output by the encoder into 512-dimensional features, which are used to constrain the encoder to learn domain-invariant information. The FC layer is a fully connected layer used to realize the dimensionality transformation and information integration of feature vectors. It maps the input features to the target dimension through matrix multiplication and is the core operation layer of the autoencoder and attention submodule. Let be the feature mapping function of the autoencoder, representing the initial features of the input obtained through the collaborative operation of the "encoder + decoder". The process of extracting domain-invariant features; The output of the autoencoder is a 256-dimensional domain-invariant embedding vector, which ensures that the features of the source domain (known material data) and the target domain (new formulation material data) have consistent semantic representation. 256 is the dimension of the autoencoder's output embedding vector, i.e. The feature dimension is the optimal dimension design that balances feature compression and information preservation; The attention submodule is a substructure of the domain adaptive embedding module. It consists of one fully connected (FC) layer and a sigmoid activation function, which is used to generate attention weights and filter domain-specific information. This is a fully connected layer operation in the attention submodule, with the input being the 512-dimensional initial features output by the backbone network. The output is a 256-dimensional intermediate feature vector, which provides dimensionality adaptation for attention weight generation; The sigmoid function is used to map the intermediate feature vectors output by the fully connected (FC) layer to the 0-1 interval, generating attention weights and achieving domain-invariant features. Element-wise weighted; The output of the attention submodule is a 256-dimensional attention weight vector, where each element takes a value from 1 to 2. Between, used to measure The importance of each feature dimension to the prediction task; The formula for calculating domain confusion features is derived by using the 256-dimensional domain-invariant features output by the autoencoder. Element-wise multiplication with the attention weight vector filters out domain-specific noise and retains effective common information; Let be the overall feature mapping function of the domain adaptive embedding module, representing the initial features mapped through the collaborative operation of the "autoencoder + attention submodule". The process of performing domain adaptation and noise filtering, and outputting domain obfuscated features; The final output of the domain adaptive embedding module is a 256-dimensional domain confusion feature vector, which has eliminated the distribution differences between the source and target domains and retained only the common features that are effective in predicting antifouling stability and safety.

[0042] Step 2.2.3: Construct the category discrimination layer.

[0043] Based on domain confusion features To predict the antifouling or safety indicators of material categories, the core approach is to simulate few-sample scenarios through an "episode training mechanism" to learn category prototypes and discrimination rules. The category discrimination layer and the domain adaptive embedding module have an "input-output" relationship, and their loss function backpropagates to optimize the parameters of the domain adaptive embedding module. The specific process is as follows: Step 2.2.3.1, episode generation.

[0044] From source domain dataset Twenty material categories were randomly selected (five samples from each category) to construct the source domain. Divided into support sets (One sample per class, 20 samples in total) and the query set (4 samples per class, 80 samples in total); from the target domain sample set Five material categories were randomly selected (five samples from each category) to construct the target domain. Similarly, it is divided into support sets. With query set .

[0045] in: The domain confusion feature vector has a dimension of 256 and is output by the domain adaptive embedding module. It has eliminated the distribution difference between the source domain and the target domain and only retains the common features related to the antifouling stability and safety of the material. It is the input feature of the category discrimination layer. The source domain dataset contains a sufficient number of labeled samples. The samples are data related to antifouling materials for pollution barriers with known formulations. Each sample is associated with complete material parameters, environmental parameters, and corresponding antifouling / safety index labels. The target domain sample set consists of a small sample dataset of the target domain. Data augmentation techniques, such as parameter perturbation and environmental condition simulation expansion, are used to supplement the sample size of the target domain and meet the training requirements of episode. An episode is a subset of training samples and is the core data unit that simulates few-shot scenarios. It includes a support set and a query set, and is divided into a source domain episode and a target domain episode, which are used by the model to learn few-shot classification rules. For the source domain episode, from A training subset of 20 material categories was randomly selected to learn the feature differences and discrimination rules of different material categories in the source domain; For the target domain episode, from A training subset of 5 material categories was randomly selected to adapt to the few-shot classification scenario of the target domain and improve the model's ability to predict new formulation materials. Support set ( () is a component of an episode, containing a small subset of labeled samples used to calculate prototype vectors for each category. The source domain support set consists of 20 categories, with one sample per category, for a total of 20 samples. The target domain support set consists of 5 classes, with 1 sample per class, for a total of 5 samples. query set ( The support set is a component of the episode, containing a subset of labeled samples used to validate the classification rules learned by the model based on the support set. The source domain query set consists of 20 categories, 4 samples per category, for a total of 80 samples. The target domain query set consists of 5 categories, 4 samples per category, for a total of 20 samples.

[0046] Step 2.2.3.2, Prototype Calculation.

[0047] For each episode's support set, calculate the prototype vector for each category using the following formula: The prototype vector is essentially the feature center of this type of material.

[0048] in: For the first The prototype vector of a material class is the feature center of that material class. It is calculated from the mean of the domain confusion features of that class of samples in the support set and is used for class matching of subsequent query samples. To support the centralized The sample set of materials of that class is the basic data for calculating the prototype vector of that class, and contains all samples of that class in the support set (here). (Contains only 1 sample) To support the number of samples in each class, here That is, each category contains only 1 sample in the support set, which is consistent with the few-sample training scenario; For support sets Domain confusion feature vectors of all samples Perform a summation operation to calculate the mean of the category prototype vectors later; The formula for calculating the prototype vector of a category is derived from... middle The mean of the domain confusion features of each sample is used to obtain the feature center of this type of material. This reflects the core characteristics of this type of material;

[0049] Step 2.2.3.3, classification prediction.

[0050] For each sample in the query set, calculate its domain confusion features. With all category prototype vectors The Euclidean distance is converted into class probabilities using the softmax function, as shown in the formula: Finally, the antifouling stability index (antifouling cycle, coating wear rate) and environmental safety index (metal ion residual concentration, biotoxicity inhibition rate) corresponding to the sample are output.

[0051] in: The Euclidean distance function is used to measure the domain confusion characteristics of query samples. With the Class prototype vector The similarity between them is such that the smaller the distance, the more similar their features are. Softmax is an activation function used to transform distance values ​​from multiple categories into a probability distribution. For query samples Belongs to the The probability of a class is output by the softmax function; the higher the probability value, the better the sample. The higher the probability of belonging to this category; For the classification prediction formula, first calculate the Euclidean distance between the query sample features and the prototypes of each category and take the negative square, then convert it into category probability through softmax to realize material category discrimination in scenarios with few samples; Antifouling stability index: One of the core indicators output by the model, including antifouling cycle (requirement). (days), coating wear rate (requirements) (), used to evaluate the antifouling durability of materials; Environmental safety indicators: The second core indicator output by the model includes the residual concentration of metal ions (requirements). ), biotoxicity inhibition rate (requirement) (This is used to evaluate the safety impact of materials on the marine environment.)

[0052] Step 2.2.4: Construct the domain alignment enhancement layer.

[0053] It forms a "two-way constraint" relationship with the domain adaptive embedding module and the few-shot prototype learning module, which are core sub-modules of the category discrimination layer. On the one hand, it receives domain confusion features from the embedding module. Prediction results from the prototype learning module On the other hand, the parameters of the first two modules are optimized inversely through the loss function to avoid excessive domain alignment leading to a decrease in class discrimination. This module contains two sub-losses, which respectively achieve a balance between "domain confusion" and "class discrimination".

[0054] The domain confusion feature vector, with a dimension of 256, is output by the domain adaptive embedding module and is one of the core input features of the domain alignment enhancement layer. It has initially eliminated the distribution difference between the source domain and the target domain. The prediction results of the prototype learning module are in the number of categories C, which is the probability distribution vector of the query sample belonging to each category, reflecting the model's judgment result on the sample category;

[0055] The bidirectional constraint is the interaction between the domain alignment enhancement layer and the domain adaptive embedding module and the few-shot prototype learning module. It receives the outputs of the first two modules as inputs and optimizes the parameters of the first two modules through backpropagation of the gradient of the loss function, so as to achieve a balance between "domain alignment" and "class differentiation".

[0056] Step 2.2.4.1, Construct the domain confusion loss .

[0057] To avoid feature dimension explosion ( It is 256-dimensional. (For dimension C), a dot product strategy using random matrix dimensionality reduction is employed, the formula is as follows: The loss formula is: ( For the domain discriminator, the input is the dimensionality-reduced joint features, and the output is the probability that a sample comes from the source domain. Gradient reversal is used to transfer features that obfuscate domain information to the embedding module, making the domain discriminator unable to distinguish the source of the sample.

[0058] in: The core function of the domain confusion loss is to drive the domain adaptive embedding module to learn domain-invariant features, so that the domain discriminator cannot distinguish whether the sample comes from the source domain or the target domain, thus narrowing the domain distribution gap. C represents the number of categories; Dimensional explosion refers to directly obfuscating features of a 256-dimensional domain. Prediction results with C dimension After concatenation, the feature dimension is too high ( The problem of increased computational complexity caused by dimensionality reduction needs to be addressed through dimensionality reduction strategies. A dot product feature fusion function based on random matrix dimensionality reduction is used to... and The features are merged into low-dimensional joint features to avoid dimensionality explosion. A randomly generated fixed matrix with dimension . Used for domain confusion features Dimensionality reduction mapping is performed, and the mapping is generated only once during the initial training phase and then remains fixed. A randomly generated fixed matrix with dimension . ( (Number of categories), used to evaluate the prediction results. Dimensionality reduction mapping is performed, and the mapping is generated only once during the initial training phase and then remains fixed. The feature dimension after dimensionality reduction is set to 64 here, which is the optimal dimension design to balance computational complexity and feature information preservation. The element-wise product (Hadamard product) operator is used to multiply elements in a dimensionality-reduced product. Feature vectors and dimensionality reduction Feature vectors are multiplied element-wise to achieve feature fusion.

[0059] Step 2.2.4.2, Construct the domain-discriminating loss. .

[0060] To address the class overlap issue that may result from domain confusion, the initial features output by the backbone network are directly processed. Apply a classification loss to enhance the class differentiation within the source and target domains. Construct a domain-discriminating discriminator. One FC layer: Input initial features Output the category prediction results, and the loss formula is: This loss is different from domain confusion loss. This creates a reverse constraint, ensuring that the embedded modules do not disrupt the differences between categories while aligning the domain distribution.

[0061] in: Domain-discriminative loss, its core function is to enhance the class distinction between the source and target domains, avoiding feature overlap between different material categories due to domain confusion. Forming a reverse constraint; The domain-discriminator consists of two fully connected layers (FC layers), with the dimension transformation path being... ( (Number of categories), the input is the 512-dimensional initial features output by the backbone network. The output is the predicted probability distribution of the sample's class. The initial feature vector output by the backbone network (same as the meaning in step 2.2.1) has a dimension of 512 and contains the original high-dimensional semantic information of the material and environment. It has not undergone domain confusion processing and is used to preserve the class distinguishing features. The formula for calculating the domain discrimination loss is as follows: For all samples in the source and target domains, calculate the negative logarithm sum of the predicted probabilities of the domain discrimination discriminator for each category. Minimizing this loss enhances the feature differences between categories and ensures the distinguishability of different material categories.

[0062] Step 2.2.4.3, Construct the source domain prototype loss. and target domain prototype loss .

[0063] Source domain prototype loss Based on source domain Support set and query set calculate: ; Target domain prototype loss Based on the target domain Support set and query set calculate: ; in, This represents the total number of samples in the query set within the source domain episode; The total number of samples in the query set within the target domain episode; For a single sample in the query set; This represents the total number of material categories selected in the episode. For the sample Corresponding category The true label unique hot encoding, if the sample Category ,but ,otherwise .

[0064] and This involves calculating the loss using cross-entropy and averaging the true label and the predicted class probability of each sample in the corresponding episode query set. This average reflects the model's discrimination error based on the prototype vector in scenarios with few samples in the source and target domains.

[0065] Step 2.2.5: Construct the loss balancing layer.

[0066] Because the model has four core losses (source domain prototype loss) Target domain prototype loss Domain confusion loss Domain differentiation loss Furthermore, there are conflicts in the optimization directions for different losses (e.g.) Orientation domain alignment, (Preferred Class Differentiation) This module is responsible for automatically calculating the weights of each loss, avoiding optimization imbalances caused by manual parameter tuning. The module is based on a Gaussian likelihood maximization strategy, assuming that each loss corresponds to a task-dependent uncertainty parameter. ( (corresponding to 4 losses respectively), the weights are defined as follows: The total loss formula is: During training, the model simultaneously optimizes network parameters and uncertainty parameters. This minimizes the total loss and achieves a dynamic balance between the losses of different tasks.

[0067] in: For the source domain prototype loss, based on the source domain The calculation is the classification loss of the few-shot prototype learning module on the source domain data, which reflects the model's accuracy in distinguishing the material categories in the source domain. The smaller the loss value, the better the classification effect in the source domain. For the target domain prototype loss, based on the target domain The calculation is the classification loss of the few-shot prototype learning module on the target domain data, which reflects the model's ability to adapt to new formulation materials in the target domain. The smaller the loss value, the better the prediction effect in the target domain. Domain confusion loss is used to promote domain distribution alignment, making the feature distributions of the source domain and the target domain more consistent. The smaller the loss value, the smaller the domain difference. Domain discrimination loss is used to enhance the distinguishability of material categories and avoid category overlap caused by domain alignment. The smaller the loss value, the more significant the feature differences between different material categories. Loss balancing layer: The loss fusion module of the model, which automatically calculates the weights of the four core losses based on the Gaussian likelihood maximization strategy, resolves the conflict between different loss optimization directions, and achieves dynamic balance; Gaussian likelihood maximization strategy: The core strategy for loss weight calculation. It assumes that each loss follows an independent Gaussian distribution. By maximizing the Gaussian likelihood of the model's prediction results, it automatically learns the task-dependent uncertainty parameters of each loss and thus determines the optimal weight. There are 4 uncertain parameters that depend on the task. correspond correspond correspond correspond This is used to quantify the task difficulty of each loss; the greater the uncertainty, the more difficult the task is to optimize. The adaptive weight for the j-th loss is defined as follows: With uncertainty parameters Dynamically adjusted, the greater the task difficulty ( (The larger the weight) The larger the value, the more optimization resources the model can allocate to difficult optimization tasks; The total loss calculation formula is derived using adaptive weights. The four core losses are weighted and summed, and the network parameters are optimized simultaneously during training. This minimizes the total loss and achieves collaborative optimization of domain alignment and category discrimination.

[0068] Step 2.2.6, Module Collaboration Workflow.

[0069] The entire forward propagation and backward optimization process of the domain adversarial prototype network forms a closed loop: The forward propagation steps are: Feature vector → Backbone network (output) → Branch 1: Domain Discriminator (Calculation) Branch 2: Domain Adaptive Embedding Module (Output) → Few-shot prototype learning module (computation) Output prediction results → Domain Confusion Discriminator (Input) and ,calculate → Adaptive weighted module (fuses 4 losses and outputs the total loss) ).

[0070] The steps of back optimization are: total loss Backpropagation via gradient descent simultaneously updates the parameters of the backbone network, the domain adaptive embedding module, the few-shot prototype learning module, the two discriminators, and the uncertainty parameter. This ensures that all modules are optimized collaboratively, ultimately resulting in a prediction model that simultaneously adapts to domain disparities and few-sample classification.

[0071] The model's final output consists of predicted values ​​for two key indicators: antifouling stability indicators (antifouling cycle, coating wear rate) and environmental safety indicators (metal ion residual concentration, biotoxicity inhibition rate), and must meet preset thresholds (antifouling cycle ≥ 180 days, coating wear rate ≤ 5%, metal ion residual concentration ≤ 0.5 mg / L, biotoxicity inhibition rate ≤ 10%).

[0072] Forward propagation: The forward computation process of the model, where feature vectors sequentially pass through the backbone network and branch calculations (domain discrimination loss). ), domain adaptive embedding module, few-shot prototype learning module (computation) and prediction results Domain confusion discriminator (computation) Finally, the total loss is output through the loss balancing layer. This completes one forward calculation.

[0073] Backward optimization: the parameter update process of the model, and the total loss. Backpropagation using the gradient descent algorithm simultaneously updates the network parameters of the backbone network, domain adaptive embedding module, few-shot prototype learning module, domain discriminator, and domain distinguishing discriminator, as well as the uncertainty parameters. This enables all modules to work together for optimization, gradually improving model performance.

[0074] Step 2.3, Model Training and Optimization.

[0075] Step 2.3.1, data partitioning.

[0076] Source domain dataset The dataset is divided into training, validation, and test sets in a 7:1.5:1.5 ratio; the target domain has a small sample set. (including enhanced) All samples are used to supplement the target domain samples in episode training, and no separate test set is set to avoid validation bias in scenarios with few samples.

[0077] in: The source domain dataset is divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5, which are used for model training, hyperparameter tuning, and final performance verification, respectively. For a small sample dataset of the target domain, including augmented data All samples are used to supplement the target domain samples in episode training, and no separate test set is set to avoid bias caused by insufficient validation data in scenarios with few samples.

[0078] Step 2.3.2, training parameters.

[0079] The stochastic gradient descent (SGD) optimizer was used, with a momentum of 0.9 (to accelerate convergence) and a weight decay of 0.01 (to suppress overfitting); the initial learning rate was... During training, adjustments are made using polynomial decay, the formula being: The batch size is set to 32, and the number of iterations is 100,000.

[0080] in: Stochastic gradient descent (SGD) is a model optimizer that calculates gradients and updates parameters by randomly selecting batches of samples, thereby minimizing the total loss. It is suitable for large-scale data training scenarios. Momentum is a parameter of the SGD optimizer, set to 0.9, used to accumulate historical gradient directions, accelerate model convergence, and avoid getting stuck in local optima during training. Weight decay is the regularization parameter of the SGD optimizer, set to 0.01, applied to the network weights. Regularization suppresses model overfitting and improves generalization ability; The initial learning rate is set to 0.001, which is the parameter update step size in the early stage of model training and determines the initial convergence speed. The learning rate is dynamically adjusted through polynomial decay during the training process, so that the learning rate gradually decreases as the training progresses. In the early stage, the optimal solution is quickly explored, and in the later stage, it is refined and optimized. This is the adjustment parameter for polynomial decay, set to 10, used to control the decay rate of the learning rate; The exponential parameter for polynomial decay is set to 0.75, which determines the shape of the learning rate decay curve. For training progress, the value range is: During initial training At the end of training , used to dynamically adjust the learning rate; The formula for calculating the dynamic learning rate changes as training progresses. Increase, learning rate Gradually decrease, and enhance the ability to explore and develop equilibrium models; The batch size is the sample batch size for each training iteration. It is set to 32, which means that 32 samples are used to calculate the gradient in each iteration, balancing computational efficiency and gradient estimation accuracy. The number of iterations is the total number of training rounds of the model, set to 100,000 to ensure that the model has sufficient training time to converge to a stable state.

[0081] Step 2.3.3, convergence determination.

[0082] Using the mean squared error (MSE) on the validation set as the convergence metric, the formula is as follows: ( This represents the true indicator value. These are the model's predicted values. (Number of samples in the validation set). When the MSE is lower than [value] for 1000 consecutive iterations... And the difference between the test set MSE and the validation set MSE is less than When the model fails to perform a full test, training is stopped to avoid overfitting or underfitting. Five-fold cross-validation is then used to further validate the model's generalization ability and ensure its predictive stability across different data partitions.

[0083] in: Mean mean squared error (MSE) is a convergence metric for a model, used to measure the deviation between the model's predicted values ​​and the actual values ​​on the validation set. The smaller the MSE, the higher the prediction accuracy. : Verify the true index values ​​of the i-th sample in the set, namely the actual antifouling stability index (antifouling cycle, coating wear rate) and environmental safety index (metal ion residual concentration, biotoxicity inhibition rate) of the material. : The model prediction value of the i-th sample in the validation set, that is, the prediction result of the corresponding index output by the model; : Number of validation set samples, i.e., the total number of samples participating in the MSE calculation; The MSE calculation formula quantifies the model's prediction error by calculating the mean of the squared differences between the predicted and true values ​​of all validation set samples. The convergence condition is the criterion for stopping model training. It requires two conditions to be met: first, the MSE (Mean Equation) must be lower than a certain value for 1000 consecutive iterations. Secondly, the difference between the test set MSE and the validation set MSE is less than 1. This ensures that the model converges without overfitting or underfitting. 5-fold cross-validation: A method for verifying the generalization ability of a model. The source domain training set is divided into 5 non-overlapping subsets. The model is trained with 4 subsets and validated with 1 subset in turn. This process is repeated 5 times and the average performance is taken. This avoids performance bias caused by a single data partition and ensures the prediction stability of the model under different data partitions. The antifouling stability index is one of the core indicators in the final output of the model, including the antifouling cycle (preset threshold). (days), coating wear rate (preset threshold) ), used to quantify the antifouling durability of materials; Environmental safety indicators are the second core indicator in the final output of the model, including the residual concentration of metal ions (preset threshold). Biotoxicity inhibition rate (preset threshold) (This is used to quantify the safety impact of materials on the marine environment.)

[0084] Step 3: Collaborative optimization under multiple constraints.

[0085] This invention proposes a hybrid metaheuristic algorithm (ALO-KHO), which integrates the global exploration capability of Ant-lion Optimizer (ALO) with the local exploitation capability of Krill herd Optimization (KHO) to solve the problem of collaborative optimization of material parameters under multiple constraints. This algorithm has an "input-output" relationship with the previously mentioned domain adversarial prototype network—each combination of material parameters generated by the optimization algorithm needs to be input into the domain adversarial prototype network to obtain predicted indices, and then the fitness is calculated based on the indices and constraints to guide the algorithm's iterative optimization. The specific process is as follows: Step 3.1, setting constraints.

[0086] (1) Technical constraints: The material must meet the mechanical performance requirements under wave and current environment (tensile strength ≥ 1.0 MPa, fatigue resistance ≥ 10). 4 (Secondary) Requirements for the dispersion stability of nanoparticles (agglomeration rate ≤10%).

[0087] (2) Performance constraints: antifouling period ≥ 180 days, coating adhesion strength to substrate ≥ 2.0 MPa.

[0088] (3) Safety constraints: It complies with marine environmental protection standards, the amount of metal ion released does not exceed the limit specified in GB / T25849-2010, and the toxicity impact on non-target organisms is ≤10%.

[0089] (4) Economic constraints: The material preparation cost increases by ≤15% compared with existing technologies.

[0090] Step 3.2, Solve using a multi-objective optimization algorithm.

[0091] Step 3.2.1, ALO-KHO algorithm parameter definition.

[0092] The core objective of the ALO-KHO algorithm optimization is to find a set of material parameter combinations (nanometallic composition type, nanoparticle size, composition ratio, deposition temperature, reaction time, coating thickness) that satisfies all constraints and achieves the optimal comprehensive score (antifouling stability + environmental safety + mechanical properties + economic cost). The value ranges of each optimization variable are set based on the material parameter database, specifically: The nano-metal composition type is a discrete variable (Cu-based, Zn-based, Ti-based). The nanoparticle size is a continuous variable (10-100 nm). The composition of nano-metals is a continuous variable (5%-20%). The deposition temperature is a continuous variable (100-500℃). The reaction time is a continuous variable (1-10h). The coating thickness is a continuous variable (50-200μm).

[0093] Step 3.2.2, Algorithm Iteration Process.

[0094] Step 3.2.2.1, Population initialization.

[0095] Forty initial solutions are generated (population size N=40). Each solution corresponds to a combination of optimization variables. Discrete variables are selected through random sampling, while continuous variables are randomly generated from a uniform distribution within their range of values. Each initial solution must first pass a "constraint check": if any constraint condition is violated (e.g., nanoparticle size = 150nm)... If the value is 100nm, then the variable value is regenerated to ensure that all solutions in the initial population are feasible.

[0096] Step 3.2.2.2, Dual-process update (global exploration + local development).

[0097] The ALO and KHO processes are responsible for "global search of new regions" and "local optimization of high-quality solutions," respectively, to prevent the algorithm from getting trapped in local optima. Both processes share the same population, and the update order is "ALO update → KHO update → fitness calculation," iterating sequentially as follows.

[0098] Step 3.2.2.2.1, ALO update process (global exploration).

[0099] Simulating the predatory behavior of antlions, the position of each solution (antlion) in the population is updated. The core principle is to expand the search range through "boundary contraction" and "random walk," as shown in the formula: ; The ALO update process is the global exploration phase of the hybrid metaheuristic algorithm (ALO-KHO), which simulates the behavior of antlions hunting prey. It expands the search range through "boundary contraction" and "random walk" strategies to avoid the algorithm getting stuck in local optima. The core objective is to find potential high-quality material parameter combinations.

[0100] For the first The generation The first solution (combination of material parameters) The value of the variable, that is, the value after ALO update. The first group of parameters New values ​​for parameters (such as nanoparticle size, composition ratio, etc.).

[0101] Index for iteration count, Indicates the current iteration generation. This indicates the next iteration number and is used to track the timing of parameter updates.

[0102] The index of the solution has a range of values. (Population size), corresponding to the [number]th [unit] in the population. A combination of material parameters.

[0103] For the index of the variable, the corresponding sequence number of the optimization variable (nanometallic composition type, nanoparticle size, etc.), each It corresponds to a specific material or process parameter.

[0104] For the first The lower limit of the value of each optimization variable is set based on the material parameter database and the feasibility of actual processes (e.g., the lower limit for the nanoparticle size is 10 nm).

[0105] For the first The upper limit of the values ​​for each optimization variable is set based on the material parameter database and the feasibility of the actual process (e.g., the upper limit for the nanoparticle size is...). .

[0106] For the first The optimal solution (antlion) in the population of generations The value of the variable, that is, the first value of the material parameter combination with the highest comprehensive score in the current iteration. These parameter values ​​serve as a guiding benchmark for global exploration.

[0107] rand is a uniformly distributed random number in the range of 0-1, which is randomly generated each time it is calculated. It is used to introduce randomness, realize "random walk", and expand the search coverage.

[0108] The purpose of this process is to guide the population to search near the global optimum, avoiding the oversight of potential high-quality parameter combinations.

[0109] Step 3.2.2.2.2, KHO update process (partial development).

[0110] Following the ALO-updated population, the aggregation and foraging behavior of krill swarms are simulated. The position of the solution is updated based on three factors: food concentration (fitness), population density, and water flow. The formula is as follows: ; in: The KHO update process is a partial development phase of the hybrid metaheuristic algorithm (ALO-KHO). It takes over the population after the ALO update, simulates the aggregation and foraging behavior of krill groups, and performs fine-grained optimization on potential high-quality solutions. The core objective is to improve the accuracy of parameter combinations. For the first The generation The solution of the first... The rate of change of a variable is used to control the update step size of that variable and reflects the rate of change of the variable in the iteration. The inertial weight is initially set to 0.9 and decreases linearly to 0.4 with each iteration. It is used to balance the algorithm’s global exploration and local development capabilities. In the early stage, a larger inertial weight is retained to promote global search, and in the later stage, the weight is reduced to strengthen local optimization. For the first The generation The solution of the first... The velocity of a variable, i.e. the update step size of that variable in the current iteration, provides the basis for the next velocity update; The learning factor is set to 2, which is used to adjust the intensity of "learning towards the optimal solution" and "learning towards the population mean", guiding the variable to be updated in a better direction; For the first The solution with the highest fitness (food source) in the population. The values ​​of each variable, and Consistency serves as the optimal parameter benchmark in the population, guiding other solutions to converge towards it. For the first The generation The solution of the first... The variable value is the current value of the variable after the ALO update and before the KHO update; For the first All solutions of the generation population The average value of each variable reflects the central tendency of the population on that variable, and is used to avoid local attraction to a single optimal solution and maintain population diversity. rand rand The KHO velocity update formula consists of three parts: an inertia term (retaining historical velocity), a cognitive term (learning from the optimal solution), and a social term (learning from the population mean), dynamically adjusting the variable update step size. The KHO position update formula is used to add the calculated new velocity to the current variable value to obtain the updated variable value of KHO, thus achieving local fine-grained optimization. The purpose of this process is to perform local fine-tuning optimization on the potential high-quality solutions found by ALO, thereby improving the accuracy of the solutions.

[0111] Step 3.2.2.3, Fitness Calculation (in collaboration with the domain adversarial prototype network).

[0112] Each updated solution (combination of material parameters) needs to be input into the trained domain adversarial prototype network prediction model to obtain the predicted value of the antifouling stability index. Predicted values ​​of environmental safety indicators Simultaneously, by combining mechanical property data and economic cost data from the material parameter database, a fitness function is constructed. The fitness function must simultaneously consider both "objective optimization" and "constraint satisfaction," and the formula is: ; Fitness calculation is the core step in the hybrid meta-heuristic algorithm (ALO-KHO) for evaluating the quality of solutions (combinations of material parameters). It works in conjunction with the trained domain adversarial prototype network prediction model, taking into account both objective optimization and constraint satisfaction, and outputs quantified fitness values ​​to provide a basis for population updates.

[0113] in: The solution is the combination of material parameters, including the specific values ​​of optimization variables such as the type of nano-metal composition, nanoparticle size, composition ratio, deposition temperature, reaction time, and coating thickness. It is the input object for fitness calculation. Domain adversarial prototype network prediction model: A trained bi-objective prediction model that takes a combination of material parameters as input and outputs a predicted value of the antifouling stability index. Predicted values ​​of environmental safety indicators The accuracy of its prediction directly affects the accuracy of fitness calculation; The antifouling stability score is predicted by the domain adversarial prototype network and is calculated by weighting indicators such as antifouling cycle (≥180 days) and coating wear rate (≤5%). The higher the score, the better the material's antifouling durability. The environmental safety score is predicted by the domain adversarial prototype network and is calculated by weighting indicators such as residual metal ion concentration (≤0.5mg / L) and biotoxicity inhibition rate (≤10%). The higher the score, the safer the material is to the marine environment. The mechanical property score is derived from a material parameter database, based on tensile strength (≥1.0 MPa) and fatigue resistance (≥10). 4Mechanical properties such as particle size distribution and nanoparticle aggregation rate (≤10%) are calculated, and a higher score indicates better material mechanical properties. The economic cost score is derived from a materials parameter database and is calculated based on the material preparation cost (with an increase of ≤15% compared to existing technologies). A higher score indicates better material economy. The fitness function is a mathematical expression that quantifies the merits of a problem, while also incorporating the objective optimization score and the penalty for constraint violation. The weights 0.35, 0.35, 0.2, and 0.1 are assigned by industry experts and correspond to the antifouling stability (…). ), environmental safety ), mechanical properties ( ), economic costs ( This reflects the importance of each indicator in the optimization process; The penalty coefficient for constraint violation is set to 100 to amplify the negative impact of constraint violation and force the algorithm to search for feasible solutions that satisfy the constraints. For the first A function representing the degree of violation of a constraint. Given the current solution (material parameter combination), the output is the quantified value of constraint violation (e.g., tensile strength If the constraints are satisfied, then If the constraints are violated... ; To constrain the effective penalties for violations, when Output 0 when (constraints are satisfied), no penalty; when Output when (constraints are violated) This will result in a corresponding penalty; This represents the total number of constraints, including all preset constraints such as technical constraints, performance constraints, safety constraints, and economic constraints. The total degree of violation of all constraints is represented by , and the sum of the degree of violation of individual constraints is represented by . The higher the total degree of violation, the greater the penalty.

[0114] The prediction accuracy of the domain adversarial prototype network directly determines the accuracy of fitness calculation, which is a key support for the optimization algorithm to find the optimal solution.

[0115] Step 3.2.2.4, Population update and convergence determination.

[0116] Calculate the fitness of all updated solutions, and retain the 40 solutions with the highest fitness as the next generation population (elite retention strategy); repeat the process of "ALO update → KHO update → fitness calculation → population update" until the number of iterations reaches 3000. After convergence, output all solutions in the population that satisfy the constraints, forming the Pareto optimal solution set (i.e., the set of solutions that cannot be optimized by optimizing one metric without harming other metrics).

[0117] Population update is a step in the algorithm iteration process to select high-quality solutions. An elite retention strategy is adopted, the fitness values ​​of all updated solutions are calculated, and the 40 solutions with the highest fitness are retained as the next generation of population to ensure that the population evolves in a better direction.

[0118] The elite retention strategy is the core strategy for population updates. It directly retains the top 40 solutions in terms of fitness in the current population, avoiding the loss of high-quality solutions during iteration and improving the convergence efficiency of the algorithm.

[0119] Convergence criterion is the criterion for stopping the algorithm's iteration. When the number of iterations reaches 3000, the algorithm is considered to have converged, and the cyclic process of "ALO update → KHO update → fitness calculation → population update" is stopped.

[0120] The Pareto optimal solution set is the set of optimal solutions output by the algorithm after convergence. Each solution in the set is a Pareto optimal solution - that is, it is impossible to optimize one index (such as S) without harming other indices (such as A, H, O). All solutions constitute a non-dominated solution set.

[0121] Step 3.3, Optimal Solution Selection The evaluation indicators include antifouling stability score, environmental safety score, mechanical performance score, and economic cost score. Industry experts assign weights to each indicator (antifouling stability 0.35, environmental safety 0.35, mechanical performance 0.2, economic cost 0.1).

[0122] Calculate the comprehensive score of each candidate solution. The scheme with the highest comprehensive score is selected as the final optimized scheme, and the corresponding material composition ratio, microstructure parameters and preparation process parameters are output.

[0123] The optimal solution selection is the step of choosing the final solution from the Pareto optimal solution set. Based on the comprehensive score ranking, the solution with the best comprehensive performance is selected as the final optimization solution.

[0124] The overall score is the evaluation index for the optimal solution, and its calculation formula is as follows: Consistent with the objective optimization part of the fitness function, it does not include constraint penalties (because Pareto optimal solutions all satisfy constraints).

[0125] The final optimized solution is the combination of material parameters with the highest comprehensive score. The output includes the composition ratio of nano-metals, nanoparticle size, microstructure morphology (porous / layered), preparation process parameters (deposition temperature, reaction time) and coating thickness, which can be directly used for the preparation of antifouling materials for pollution barriers.

[0126] like Figure 3 As shown, the error rate evolution trends and performance differences of the DAPN-ALO-KHO algorithm compared with five other algorithms under different training iterations: Convolutional Neural Network Optimization Based on Genetic Algorithm (CNN+GA), Convolutional Neural Network Optimization Based on Particle Swarm Optimization (CNN+PSO), Convolutional Neural Network Optimization Based on Reinforcement Learning (CNN+RL), DAPN-ALO, and DAPN-KHO.

[0127] From the perspective of overall evolution, the error rate of all algorithms shows a monotonically decreasing trend with the increase of training times. The rate of decrease is more significant in the early stage of training (100-400 times) and gradually flattens out in the later stage (700-1000 times). This pattern is consistent with the convergence characteristics of machine learning model training, that is, the model learns data features quickly in the early stage and gradually approaches the optimal solution in the later stage.

[0128] In terms of quantitative performance comparison, the DAPN-ALO-KHO algorithm consistently maintains the lowest error rate. When the number of training iterations reaches 1000, its error rate drops to 4%, significantly outperforming three hybrid algorithms based on traditional CNNs: CNN+GA (11%), CNN+PSO (9%), and CNN+RL (12%). It also outperforms the DAPN-ALO (5%) and DAPN-KHO (4%) algorithms that only employ a single heuristic optimization. This advantage stems from the dual-core design of the DAPN-ALO-KHO algorithm: the Domain Adversarial Prototype Network (DAPN) effectively alleviates the problem of insufficient generalization ability caused by cross-domain data distribution differences and sample scarcity through the synergy of domain adaptive embedding and few-shot prototype learning. Meanwhile, the hybrid strategy of Antlion Optimization (ALO) and Krill Swarm Optimization (KHO) achieves a dynamic balance between global exploration and local development, improving the efficiency and accuracy of parameter optimization.

[0129] In contrast, CNN+GA, CNN+PSO, and CNN+RL algorithms are limited by the domain adaptation limitations of traditional feature extraction networks and the search limitations of single optimization algorithms, resulting in consistently high error rates. While DAPN-ALO and DAPN-KHO leverage the DAPN model to ensure the effectiveness of feature learning, the lack of completeness in searching multi-constrained parameter spaces due to the single heuristic algorithm leads to slightly inferior performance compared to DAPN-ALO-KHO. In summary, DAPN-ALO-KHO, through deep collaborative innovation between the prediction model and the optimization algorithm, demonstrates superior stability and convergence performance in the optimization scenario of antifouling materials with few samples, cross-domain conditions, and multiple constraints, providing an effective technical reference for similar complex engineering optimization problems.

[0130] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for synergistic optimization of the antifouling stability and safety of antifouling materials for pollution-blocking nets, characterized in that, Includes the following steps: Step 1, Multi-dimensional data acquisition and sample preparation: The multi-dimensional data includes a biological-environmental parameter correlation database, a material parameter database, and a material performance-environmental safety database. The samples are used for data verification. Step 2, AI Prediction Model Construction and Training: Key feature parameters are selected from the multi-dimensional data collected in Step 1. These key features include antifouling stability features, environmental safety features, material structure features, and environmental operating condition features. Using these key feature parameters as the core feature vector, a bi-objective prediction model based on a domain adversarial prototype network is constructed. After model training and optimization, the predicted value of the material's antifouling stability index is output. Predicted values ​​of environmental safety indicators ; In step 2, the method for constructing the dual-objective prediction model includes the following steps: Step s21, Constructing the base layer: The base layer uses ResNet18 as the backbone network to transform the core feature vectors into 512-dimensional initial feature vectors. ; Step s22: Construct a domain-adaptive embedding module, which consists of an autoencoder and an attention submodule connected in series. The autoencoder is based on... As input, the encoder and decoder output a 256-dimensional embedding vector. Attention submodule and As input, the output is a 256-dimensional domain confusion feature. ,in, This refers to the fully connected layer operation within the attention submodule, where sigmoid is the activation function. Step s23: Construct a class discrimination layer with the episode training mechanism as its core: from the source domain dataset Source domain episodes are constructed by selecting different material categories. Divided into support sets With query set From the target domain sample set Constructing the target domain episode by selecting different material categories Similarly, it is divided into support sets. With query set For each and For each sample in the dataset, calculate its prototype vector. ;right and For each sample in the dataset, calculate its domain confusion features. With all category prototype vectors The Euclidean distance is converted into class probabilities using the softmax function, and the prediction result is output. This includes the antifouling stability index and environmental safety index corresponding to the sample; Step s24, construct the domain alignment enhancement layer: receive the domain confusion features from step s22. The prediction results of step s23 And by back-optimizing the parameters of the category discrimination layer and the domain adaptive embedding module through the loss function, a closed loop is formed; Step 3, Collaborative optimization under multiple constraints: Step s31: Set multiple constraints for the material; Step s32: Use the hybrid meta-heuristic algorithm ALO-KHO to perform multi-objective optimization. The solution process is as follows: First, generate N initial solutions, each corresponding to a combination of optimization variables. Then, calculate the updated solutions by sequentially updating using ALO and KHO. Substitute the updated solutions into the bi-objective prediction model trained in step 2 to obtain the final solution. and Combined with mechanical performance scores and economic cost score Calculate the fitness of each updated solution, retain the N solutions with the highest fitness as the next generation population, and iterate in this way until convergence is achieved to obtain the Pareto optimal solution; Step s33: Select the optimal solution from the Pareto optimal solution to obtain the parameter combination of the antifouling material of the debris barrier with the highest score.

2. The method for synergistic optimization of the antifouling stability and safety of the antifouling material of the debris barrier net as described in claim 1, characterized in that, In step 1, the multi-dimensional data includes the source domain dataset. With the target domain few-sample dataset and to Expand to obtain the target domain sample set .

3. The method for synergistic optimization of the antifouling stability and safety of the antifouling material of the debris barrier net as described in claim 1, characterized in that, In step s23 , To support the centralized A sample set of similar materials To support the number of samples in each class, For the first Domain confusion features of each sample.

4. The method for synergistic optimization of the antifouling stability and safety of the antifouling material of the debris barrier net as described in claim 3, characterized in that, In step s23, the prediction formula for the category probability is: ； in, For query samples Belongs to the The probability of a class This is the function for calculating Euclidean distance.

5. The method for synergistic optimization of the antifouling stability and safety of the antifouling material of the debris barrier net as described in claim 4, characterized in that, In step s24, the total loss of the loss function The calculation formula is: ； in, For the source domain prototype loss, For the target domain prototype loss, Loss due to domain confusion To differentiate the loss between domains, , , and These are adaptive weights, , =1, 2, 3, or 4, For uncertain parameters; The calculation formula is: ； ； in, For the first One sample, For domain discriminator, This is a dot product feature fusion function based on dimensionality reduction of random matrices. For the Hadamard product operator, A randomly generated fixed matrix with dimension . , A randomly generated fixed matrix with dimension . , The feature dimensions after dimensionality reduction. Number of categories; The calculation formula is: , For domain discrimination; The calculation formula is: ; The calculation formula is: ; This represents the total number of samples in the query set within the source domain episode; The total number of samples in the query set within the target domain episode; For the sample Corresponding category The true label unique hot encoding.

6. The method for synergistic optimization of the antifouling stability and safety of the antifouling material of the debris barrier net as described in claim 1, characterized in that, In step 2, the model training and optimization steps are as follows: Will The dataset is divided into training, validation, and test sets. The SGD optimizer is used to optimize the parameters of the bi-objective prediction model, and the mean squared error (MSE) on the validation set is used as the convergence metric.

7. The method for synergistic optimization of the antifouling stability and safety of the antifouling material of the debris barrier net as described in claim 1, characterized in that, In step s31, the constraints include technical constraints, performance constraints, safety constraints, and economic constraints. In step s32, the combination of optimization variables includes nano-metal composition type, nanoparticle size, composition ratio, deposition temperature, reaction time, and coating thickness.

8. The method for synergistic optimization of the antifouling stability and safety of the antifouling material of the debris barrier net as described in claim 1, characterized in that, In step s32, the ALO update process is as follows: ； in, For the first The generation The solution of the first... The value of the variable, the first The solution is the first A combination of material parameters, Index for iteration count, For the index of the solution, For the index of the variable, For the first The lower bound of the values ​​of each optimization variable. For the first The upper limit of the values ​​of each optimization variable. For the first The optimal solution in the population of generations There are several variable values, where rand is a uniformly random number in the interval 0-1; The KHO update process is as follows: ； in, For the first The generation The solution of the first... The speed of each variable For inertial weights, For the first The generation The solution of the first... The speed of each variable As a learning factor, For the first The solution with the highest fitness in the population is the first One variable value, For the first The generation The solution of the first... One variable value, For the first All solutions of the generation population The average of the variables.

9. The method for synergistic optimization of the antifouling stability and safety of the antifouling material of the debris barrier net as described in claim 1, characterized in that, In step s32, the formula for calculating fitness is: ； in, For fitness, To score for antifouling stability, To score environmental safety For mechanical performance scores, Score the economic cost. To constrain the penalty coefficient for violations, For the first A function representing the degree of violation of a constraint. For the current solution, The total number of constraints. To constrain the effective penalties for violations, when hour, Output 0 when hour, Output .

Images