Formula screening method for multifunctional anti-corrosion and anti-fouling coating of trash holding net

Abstract

The invention belongs to the technical field of coating material formula optimization, and discloses a screening method for a multifunctional anti-corrosion and anti-fouling coating formula of a trash holding net, which comprises the following steps: step 1, collecting core data for preprocessing, and realizing deep fusion of'coating component-performance-environment-failure 'multi-modal data by using seed-guided neural metric learning; step 2, constructing an LAM-GRU encoder, and outputting a multi-dimensional feature vector E of the coating; 3, constructing a multi-task performance prediction network, taking E as input, outputting predicted values of four types of performance indexes, constructing a loss function to distribute differentiation weights to the four types of performance indexes, and optimizing parameters of the multi-task performance prediction network by adopting an Adam optimizer; 4, efficient optimization of a coating formula is carried out through HGWSFO; and step 5, constructing a weighted multi-objective optimization function, continuously iterating the HGWSFO, and outputting an optimal formula which meets a beam condition and has a maximum objective function. The performance prediction precision is high, and the formula optimization efficiency and reliability are both achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coating material formula optimization, in particular to a screening method for a multifunctional anti-corrosion and anti-fouling coating formula of a trash rack. BACKGROUND

[0002] As key equipment in marine engineering and water conservancy engineering, the trash rack is long-term served in the complex marine environment with high salt, high humidity and biological enrichment, and faces serious corrosion and biological fouling problems, which leads to the shortening of the service life of the equipment, the increase of the maintenance cost, and even affects the normal operation of the project. Therefore, the development of a multifunctional coating with excellent anti-corrosion and anti-fouling performance and good mechanical stability is the core means to solve the above problems.

[0003] The existing coating formula development of the trash rack mainly relies on the traditional experimental trial and error method, which needs to repeatedly adjust the component proportion and perform a large number of performance tests, and has defects such as long development cycle, high test cost and large blind optimization of formula. With the development of material genetic engineering and artificial intelligence technology, some research attempts to realize the performance prediction of the coating through a machine learning model, but there are still many deficiencies: first, the multi-modal data fusion effect is poor, it is difficult to effectively integrate heterogeneous data such as coating components, service environment, performance indicators and failure cases, leading to incomplete feature representation; second, the traditional encoder lacks the memory and correlation ability of local features, and cannot capture the complex mapping relationship between the subtle changes of component content, environmental parameter fluctuations and coating performance, so the prediction accuracy is limited; third, the formula optimization algorithm has the problem of exploration and utilization imbalance, and the single global search algorithm is easy to fall into local optimum, and the single local search algorithm is difficult to cover a wide range of formula space, resulting in low efficiency and insufficient reliability of the optimal formula screening.

[0004] In addition, there is a lack of a dedicated formula screening scheme for the service scenario of the trash rack in the prior art, which fails to fully combine key factors such as the physicochemical properties of the marine environment and the distribution of biological communities, and it is difficult to ensure the long-term effectiveness of the coating in the actual service process. Therefore, there is an urgent need for a technical solution that can efficiently fuse multi-modal data, accurately predict coating performance and quickly screen the optimal formula, in order to break through the bottleneck of the existing technology. SUMMARY

[0005] The purpose of the present application is to provide a screening method for a multifunctional anti-corrosion and anti-fouling coating formula of a trash rack, which overcomes the technical defects in the prior art.

[0006] The technical solution adopted by the present application to achieve the purpose is as follows: A screening method for a multifunctional anti-corrosion and anti-fouling coating formula of a trash rack, comprising the following steps: Step 1, collect core data to build a database, pre-process to obtain a pre-processed database, use seed-guided neural metric learning to realize deep fusion of "coating component-performance-environment-failure" multi-modal data, generate a seed sample set with similarity labels by screening representative seed samples, and construct a similarity matrix S ; Step 2, construct a grid-based memory tensor M Store local feature information of seed samples, local features of the similarity matrix S are written into the grid-based memory tensor M firstly, construct a LAM-GRU encoder, and input the coating component vector , the environment vector corresponding to the grid coordinates M in the grid-based memory tensor to output a coating multi-dimensional feature vector E , and the dimension of the coating multi-dimensional feature vector d ; Step 3, construct a multi-task performance prediction network, input the coating multi-dimensional feature vector E to output predicted values of four types of performance indicators, construct a loss function to assign different weights to the four types of performance indicators, and use an Adam optimizer to optimize parameters of the multi-task performance prediction network. Step 4, perform efficient optimization of coating formulations by HGWSFO: dynamically switch local search of grey wolf optimization and global search of sunflower optimization through index threshold value; Step 5, set optimization objectives and constraints, construct a weighted multi-objective optimization function, and HGWSFO iterates to output an optimal formula that meets the constraint conditions and has the maximum objective function.

[0007] In the above technical solution, in step 1, the data in the database includes coating formulation component data, performance data, environment data and failure data; the pre-processing includes data standardization, abnormal data elimination and data completion and enhancement.

[0008] In the above technical solution, in step 1, the similarity matrix , is the number of representative coating formulation components randomly sampled from the pre-processed database.

[0009] In the above technical solution, in step 2, the grid-based memory tensor , P , Q is the number of grid divisions of the component or environment parameter in the first dimension and the second dimension, respectively.

[0010] In the above technical solution, in step 2, the multidimensional feature vector of the coating is output through the read / write operation of the LAM-GRU encoder. E The specific steps include the following: (1) Reading operation: based on coating component vector or environment vector In grid-based memory tensors M The corresponding grid coordinates Scan local bandwidth w The neighborhood grid, through attention weights Fusion grid-based memory tensor Historical characteristics in the process of generating historical states ; (2) Write operation: using the bureau department Control the update intensity and update the current feature state. The weighted update to the memory tensor is calculated using the following formula: ; For bureau departments; This represents the current feature state; It is the sigmoid activation function; grid coordinates The corresponding original feature values; grid coordinates The corresponding updated feature values; (3) The final hidden layer output of the LAM-GRU encoder is used as the multidimensional feature vector. E To achieve a unified representation of multimodal data; ; This is the weight matrix. For bias vectors, This is the final time step after the LAM-GRU encoder has traversed all input sequences. The hidden layer state, This enhances the nonlinear representation of features for activation functions.

[0011] In the above technical solution, the formula for the multi-task performance prediction network in step 3 is expressed as follows: ; For the first k Predicted values ​​of class performance indicators, These correspond to four performance categories: corrosion resistance, antifouling performance, mechanical properties, and environmental adaptability. These are the weight matrix and bias vector of the fully connected layer 1 of the branch, respectively. For the hidden layer dimension; These are the weight matrix and bias vector of the branch output layer, respectively. For the first k The number of performance metrics for a class; For the hidden layer activation function, This is the activation function for the branch output layer.

[0012] In the above technical solution, the loss function in step 3 is: ; To calculate the weighted total loss across multiple tasks, For the first Class performance weights For the first Class performance The true value of each indicator For the first Class performance The predicted values ​​of each indicator.

[0013] In the above technical solution, in step 4, an index threshold is set. When the index value of the candidate recipe set When the index value of the recipe candidate set is reached, SFO is initiated to perform a global exploration. At that time, GWO will be launched to perform a local search.

[0014] In the above technical solution, the specific steps of SFO global exploration are as follows: (1) Initialize the "Sunflower" population and set the coating formula with the best fitness to "Sun". ; (2) Calculate the step size based on the inverse square law: ,in This is the maximum step size in the SFO stage, used to control the maximum magnitude by which the candidate formulation moves towards the "sun" formulation. These represent the upper and lower limits of the content of each component. This refers to the size of the sunflower population. (3) Generate new formulas through pollination operations Remove "dead plants" with poor adaptability and update the "sun". ; The specific steps of GWO local search are as follows: (1) Initialize the "gray wolf" population and set the top three coating formulas with the highest fitness as leaders. Secondary leaders Executor ; (2) Calculate the coefficient vector: the first coefficient vector of the GWO stage The second coefficient vector of the GWO stage , coefficient decay factor for GWO phase, linearly decreasing from 2 to 0, rand1, are random numbers within the interval; the first coefficient vector B is used to control the candidate matching direction to approach or move away from the optimal matching formula, and the second coefficient vector is used to enhance the randomness of the search; (3) calculating the updated general formula in the GWO phase .

[0015] In the above technical solution, the weighted multi-objective optimization function in step 5 is: ; Wherein: is the weighted multi-objective optimization function value; is the weight of the antifouling effective period; is the weight of the biological attachment coverage rate; is the 28-day biological attachment coverage rate; is the weight of the wear resistance; is the weight of the material cost; is the wear resistance; is the wear resistance constraint threshold; is the cost of coating materials per square meter; is the material cost constraint threshold.

[0016] Compared with the prior art, the beneficial effects of the present application are: 1. Excellent multi-modal data fusion effect: the seed-guided neural metric learning technology designed in the present application can effectively integrate multi-source heterogeneous data such as components, performance, environment and failure, establish data association through the seed sample similarity matrix, avoid the limitations of single data dimension, provide comprehensive and reliable data basis for coating feature characterization, and the integrity and effectiveness of feature fusion are significantly better than traditional methods; 2. High performance prediction accuracy: the local attention memory (LAM) enhanced encoder stores local features and historical information through memory tensor, and combines dynamic adjustment of attention weight to accurately capture the complex mapping relationship between component content change, environmental parameter fluctuation and coating performance; the multi-task performance prediction network simultaneously covers 4 types of core performance indicators, and the weighted loss function optimizes the training priority, so that the prediction error is significantly reduced, and the prediction accuracy of each type of performance indicator is improved by more than 30% compared with the prior art; 3. Formula optimization efficiency and reliability: mixed gray wolf-sunflower optimization algorithm (HGWSFO) switches global exploration and local search dynamically, which ensures comprehensive coverage of the formula space and precise mining of high-quality formula area, avoiding falling into local optimum. The optimal candidate formula can be output after 50 iterations, and the screening efficiency is improved by 50%-80% compared with traditional optimization algorithms, and the screened formula meets the comprehensive demand of "high corrosion resistance, strong antifouling, excellent mechanics and low cost"; 4. Reduced research and development costs and shortened cycle: the present application replaces a large number of repeated experiments through AI technology, reduces the number of physical tests, and reduces the test cost by more than 60%; the formula screening is fully automated, and the research and development cycle is shortened from 3-6 months to 2-4 weeks, significantly improving the research and development efficiency; 5. Strong adaptability and practicality: for the marine service environment of the trash rack, real-time and historical environmental data of typical sea areas are specially accessed, and the screened formula can accurately match the physicochemical properties and biological distribution characteristics of the marine environment, the antifouling effective period of the coating is extended to more than 24 months, the salt spray corrosion resistance time is ≥1000h, the mechanical properties meet the engineering application requirements, and the service life is 2-3 times longer than that of traditional coatings, which has significant engineering application value and economic and social benefits. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flow chart of a multifunctional corrosion and antifouling coating formula screening method for trash racks.

[0018] Figure 2 is a performance comparison chart of different algorithms. DETAILED DESCRIPTION

[0019] The present application will be further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. EMBODIMENT

[0020] As shown in Figure 1 , a multifunctional corrosion and antifouling coating formula screening method for trash racks comprises the following steps:

[0021] Step 1, collect core data to build a database, pre-process to obtain a pre-processed database, use seed-guided neural metric learning technology to realize deep fusion of "coating component-performance-environment-failure" multi-modal data, select representative seed samples to generate a seed sample set with similarity labels, and construct a similarity matrix , which provides high-quality data support for subsequent feature extraction and model training, and solves the problem of difficult effective integration of multi-source heterogeneous data.

[0022] Step s11, collect core data to build a database.

[0023] 1.1 Coating formulation component data collection:

[0024] Collect more than 20 commonly used coating components, such as resins (epoxy resin, polyurethane, fluorocarbon resin, etc.), antirust pigments (zinc powder, zinc phosphate, aluminum tripolyphosphate, etc.), antifouling agents (copper-based compounds, environmentally friendly isothiazolinone, tetrabutyl tin oxide, etc.), and additives (dispersants, leveling agents, defoamers, etc.), record the basic parameters of each component (including molecular weight, number of functional groups, density, melting point, solubility parameter) and the content ratio range in industrial application (such as resin ratio 30%-60%, antirust pigment ratio 15%-35%, antifouling agent ratio 5%-15%, additive ratio 1%-5%), form a component database.

[0025] 1.2 Performance data collection:

[0026] Through literature retrieval, historical experiment records and supplementary experiments, 12 key performance indicators of coatings are obtained: (1) Corrosion resistance: specific performance indicators include polarization resistance (test range ), salt spray test corrosion resistance time (test period 0-2000h), electrochemical impedance spectrum (low frequency impedance value, phase angle); (2) Antifouling performance: specific performance indicators include 28-day marine biological (diatom, barnacle, oyster) adhesion coverage (0%-100%), antifouling effective period (0-36 months), antifouling agent bleeding rate ; (3) Mechanical properties: specific performance indicators include adhesion (crosshatch method / pull-off method test, 0-10MPa), wear resistance (Taber wear index, mg / 1000 turns), flexibility (bending test diameter, mm), hardness (pencil hardness / HV hardness); (4) Environmental adaptability: specific performance indicators include performance retention rate after high and low temperature cycle (-40℃-80℃), performance decay rate after damp heat aging (40℃, 95%RH).

[0027] 1.3 Environmental data collection:

[0028] Access to real-time and historical environmental data in five typical pollution net deployment sea areas nationwide: (1) Physicochemical environmental parameters: seawater temperature (-2℃~30℃), salinity (28‰~35‰), pH value (7.5~8.5), dissolved oxygen (2~12mg / L), turbidity (1~50NTU), seawater flow rate (0~2m / s); (2) Biological environmental parameters: plankton abundance (ind / m³), benthic organism coverage rate (%), seawater nutrient salt concentration (nitrogen, phosphorus, silicon, mg / L); (3) Extreme environment data: typhoon season seawater impact force, tidal range, seawater corrosion potential (mV).

[0029] 1.4 Failure data collection:

[0030] Collect failure cases of different coating formulations during actual service as negative samples of the model, including: (1) Failure type: coating blistering (blistering density, diameter), cracking (crack length, width), peeling (peeling area ratio), massive bioattachment (time node when attachment amount ≥ 50%), antifouling agent depletion (antifouling failure time); (2) Failure-related data: record environmental conditions (temperature, salinity, biological abundance) at the time of failure, coating service time, and early performance degradation trend, and establish a "formulation-environment-failure" correlation data set.

[0031] Step s12, data preprocessing, to obtain a preprocessed database: (1) Data standardization: use Z-Score standardization method to eliminate the influence of different indicators on the same order of magnitude, which is convenient for model training; (2) Abnormal data elimination: use the Isolation Forest algorithm to isolate data by constructing multiple isolated trees, calculate the anomaly score of each sample, and eliminate abnormal data (such as abnormally high / low polarization resistance due to experimental error, negative bioattachment rate, etc.) with a score higher than the threshold (default 0.8); (3) Data completion and enhancement: use K-Nearest Neighbors (KNN) interpolation method to complete missing data, and use SMOTE (Synthetic Minority Over-sampling Technique) to generate virtual samples for small sample data, to ensure the balanced distribution of various formulations, performance, and environmental data in the database.

[0032] Step s13, screening seed samples, and constructing a similarity matrix .

[0033] Randomly sample representative coating formulations from the preprocessed database as seed samples (default ), calculate the comprehensive similarity between seed samples based on "component-performance-environment", and construct a similarity matrix (normalized to the interval [0, 1]). is the number of seed samples, i.e., the number of representative coating formulations randomly sampled from the preprocessed database, and the default value is 2000. is the similarity matrix of seed samples, with a dimension of , and the element value in the matrix ranges from , which is used to represent the comprehensive similarity between any two seed samples based on the "component-performance-environment", and has been normalized.

[0034] Step 2, constructing the grid-based memory tensor Store the local feature information of the seed sample, the similarity matrix The local features are preferentially written into the grid-based memory tensor , construct a local attention memory enhanced (LAM-GRU) encoder to coat the component vector , the environment vector In the grid-based memory tensor The corresponding grid coordinates As input, output the coating multi-dimensional feature vector , the dimension is . Store the local feature information through the grid-based memory tensor, combined with the read-write operation of the attention mechanism, capture the local association and historical feature influence of the coating components and environmental parameters, improve the unified representation ability of multi-modal data, and lay the foundation for accurate performance prediction.

[0035] Step s21, construct the grid-based memory tensor ( The number of grid divisions of the component or environmental parameter, The dimension of the coating multi-dimensional feature vector Store the local feature information of the seed sample; The grid-based memory tensor, the dimension is , used to store the local feature information of the seed sample. The number of grid divisions of the component or environmental parameter (the first dimension), that is, the first dimension division number when the coating component or environmental related parameter is divided into grid. The number of grid divisions of the component or environmental parameter (the second dimension), that is, the second dimension division number when the coating component or environmental related parameter is divided into grid. The default value is 128.

[0036] Step s22, design LAM-GRU as an encoder, input is the coating component vector (each component content), the environment vector (temperature, salinity, etc.) in the grid-based memory tensor The corresponding grid coordinates, output the coating multi-dimensional feature vector through the read-write operation of the LAM-GRU encoder, the specific steps are as follows: (1) Read operation: based on the current input grid coordinates , scan the neighborhood grid with a local bandwidth , through the attention weight Fusing historical features in the memory tensor to generate a historical state ; where, is the coating component vector, which is composed of the content of each type of coating component (resin, antirust pigment, antifouling agent, auxiliary agent, etc.). is the environment vector, which is composed of parameters related to the service environment of the coating (sea water temperature, salinity, pH value, etc.). is the current input grid coordinate, i.e., the coating component vector or the environment vector In the grid-based memory tensor , the corresponding grid coordinate. is the local bandwidth, i.e., the bandwidth size of scanning the neighborhood grid during the read operation, which is fixed at 2. A is the attention weight, which is used to measure the contribution of different neighborhood grid historical features in the memory tensor to the generation of the current feature. is the historical state, which is generated by fusing the historical features in the memory tensor through the read operation, and is used to enhance the representation ability of the current coating feature. Seed sample similarity matrix is used to select "high similarity seed sample pairs", whose local features are preferentially written into the memory tensor , providing a high-quality historical feature basis for the read operation; (2) Write operation: use the local department to control the update intensity, and update the current feature state to the memory tensor, with the formula: ; is the local department, which is used to control the update intensity of the memory tensor in the write operation, and determines the proportion of the influence of the current feature state on the memory tensor; is the current feature state, i.e., the coating feature state generated by the LAM-GRU encoder at the current time step; is the sigmoid activation function, which is used to map the value of the local department to the interval, realizing the normalized control of the update intensity of the memory tensor; is the grid coordinate corresponding to the original feature value, i.e., the local feature information of the seed sample stored in the grid before updating; is the grid coordinate corresponding to the updated feature value, i.e., the feature information after fusing the current feature state through the write operation; (3) Take the final hidden layer output of the LAM-GRU encoder as a multi-dimensional feature vector , realizing the unified representation of multi-modal data; Multi-dimensional feature vector Output by the final hidden layer of the LAM-GRU encoder for realizing unified representation of the coating "component-performance-environment" multi-modal data. is a weight matrix, is a bias vector, is the hidden layer state at the final time when the LAM-GRU encoder traverses all input sequences, is an activation function to enhance feature nonlinear representation.

[0037] Step 3, constructing a multi-task performance prediction network: constructing a multi-output fully connected network with the coating multi-dimensional feature vector E as input, and respectively outputting the predicted values of the corrosion resistance, antifouling, mechanics, and environmental adaptation of the coating: (1) Corrosion resistance performance branch: outputting the predicted value of the polarization resistance , the predicted value of the salt spray corrosion resistance time , and the predicted value of the electrochemical impedance spectrum parameters , ; (2) Antifouling performance branch: outputting the predicted value of the biofouling coverage , the predicted value of the antifouling effective period , and the predicted value of the antifouling agent exudation rate ; (3) Mechanics performance branch: outputting the predicted value of the adhesion , the predicted value of the wear resistance , the predicted value of the flexibility , and the predicted value of the hardness , ; (4) Environmental adaptation branch: outputting the predicted value of the performance retention rate after high-low temperature cycle / hot and humid aging .

[0038] Multi-dimensional feature vector as the input of the multi-output fully connected network, for unified representation of the coating "component-performance-environment" multi-modal data. is the predicted value of the polarization resistance, i.e., the estimated value of the coating polarization resistance output by the corrosion resistance performance branch of the multi-task performance prediction network. is the predicted value of the salt spray corrosion resistance time, i.e., the estimated value of the coating salt spray test corrosion resistance time output by the corrosion resistance performance branch of the multi-task performance prediction network. is the predicted value of the low-frequency electrochemical impedance, i.e., the estimated value of the low-frequency impedance in the coating electrochemical impedance spectrum output by the corrosion resistance performance branch of the multi-task performance prediction network. ​The phase angle prediction value is an estimated value of the phase angle in the coating electrochemical impedance spectrum output by the anti-corrosion performance branch of the multi-task performance prediction network. The bio-attachment coverage prediction value is an estimated value of the 28-day marine bio-attachment coverage of the coating output by the anti-fouling performance branch of the multi-task performance prediction network. The anti-fouling effective period prediction value is an estimated value of the effective duration of the anti-fouling function of the coating output by the anti-fouling performance branch of the multi-task performance prediction network. The anti-fouling agent leaching rate prediction value is an estimated value of the anti-fouling agent leaching rate in the coating output by the anti-fouling performance branch of the multi-task performance prediction network. The adhesion prediction value is an estimated value of the adhesion of the coating output by the mechanical performance branch of the multi-task performance prediction network (corresponding to the test by the cross-hatch method or the pull-off method). The abrasion resistance prediction value (Taber abrasion index) is an estimated value of the abrasion resistance of the coating output by the mechanical performance branch of the multi-task performance prediction network, with the unit of Turn. The flexibility prediction value is an estimated value of the bending test diameter of the coating output by the mechanical performance branch of the multi-task performance prediction network, with the unit of mm. The pencil hardness prediction value is an estimated value of the pencil hardness of the coating output by the mechanical performance branch of the multi-task performance prediction network. The HV hardness prediction value is an estimated value of the Vickers hardness of the coating output by the mechanical performance branch of the multi-task performance prediction network. The performance retention rate prediction value after high-low temperature cycling is an estimated value of the performance retention rate of the coating after high-low temperature cycling output by the environmental adaptation branch of the multi-task performance prediction network. The performance retention rate prediction value after high-low temperature cycling is an estimated value of the performance retention rate of the coating after high-low temperature cycling output by the environmental adaptation branch of the multi-task performance prediction network. The performance retention rate prediction value after damp-heat aging is an estimated value of the performance retention rate of the coating after 40°C, 95% RH damp-heat aging output by the environmental adaptation branch of the multi-task performance prediction network.

[0039] Step s31, all performance branches share the input multi-dimensional feature vector , but each has independent learnable parameters (weight matrix , bias vector ) to avoid interference between different performance indicators. The general calculation process is as follows: ; The prediction value of the kth performance is , respectively corresponding to the anti-corrosion performance, anti-fouling performance, mechanical performance, and environmental adaptation four performance branches; , , respectively, the weight matrix and bias vector of the first layer full connection of the branch, , the hidden layer dimension , These are the weight matrix and bias vector of the branch output layer, respectively. The number of performance metrics for the k-th class Use the activation function of the hidden layer to enhance the nonlinear representation capability. The activation function for the branch output layer (choose according to the physical range of the performance metric, such as Sigmoid for proportional metrics and Linear for continuous numerical metrics).

[0040] Step s32, loss function design.

[0041] Construct a weighted multi-task loss function and assign differentiated weights to different performance metrics (such as weights for anti-pollution effective options). Polarization resistance weight Adhesion weight Environmental adaptability weight The formula is: ; In the formula, The weighted total loss for multiple tasks is used to measure the overall prediction error of the multi-output fully connected network for four performance metrics, and to guide the optimization of model parameters. For the first The weights for different performance categories are used to adjust their priority during model training. Corresponding antifouling performance, Corrosion resistance, Corresponding mechanical properties The default values ​​for corresponding environmental adaptability are as follows: , ; For the first The number of performance metrics for a class, i.e., the first The number of specific performance metrics included in class performance (e.g.) hour, (corresponding to three indicators of antifouling performance). For the first Class performance The true value of an indicator, that is, the actual value of the performance indicator obtained through experimental testing or historical data recording. For the first Class performance The predicted value of each indicator, that is, the estimated value of the performance indicator output by the multi-output fully connected network.

[0042] Step s33: The Adam optimizer is used to optimize the multi-task performance prediction network.

[0043] Using the Adam optimizer, set the learning rate. , the number of iterations round, each round is based on the seed sample similarity matrix Design distance weighted sampling strategy, select discriminative sample pair (similar formula, dissimilar formula) for training, improve the sensitivity of multi-task performance prediction network to formula subtle differences.

[0044] Adam optimizer: optimization algorithm for model training, accelerate model convergence by adaptive learning rate adjustment, improve parameter optimization efficiency. is the learning rate, used to control the step size of each parameter update of the Adam optimizer, the default value is ; is the maximum number of iterations for model training, that is, the total number of model parameter optimization rounds, the default value is 100 rounds; is the seed sample similarity matrix (same as the previous meaning), used to guide the implementation of the distance weighted sampling strategy in each round of training, and select the sample pair with strong discriminative ability.

[0045] Adam optimizer adjusts the learning rate adaptively to minimize the weighted multi-task total loss , realizes the optimization of multi-task performance prediction network parameters (each branch weight , bias ).

[0046] Step 4, efficient optimization of coating formula by hybrid grey wolf-sunflower optimization algorithm (HGWSFO), dynamic switching of local search of grey wolf optimization and global search of sunflower optimization through index threshold, balancing exploration breadth and utilization accuracy of algorithm, avoiding the limitations of single optimization algorithm.

[0047] Set the index threshold , realize the dynamic switching of local search of grey wolf optimization (GWO) and global search of sunflower optimization (SFO), is the index value of the formula candidate set, used to compare with the index threshold , to determine the currently enabled optimization search mode:

[0048] When the index value of the formula candidate set is , start SFO for global exploration: (1) Initialize the "sunflower" population (formula candidate set, population size ), set the formula with the best fitness as "sun" ; is the sunflower population size, that is, the number of formula candidate sets in the SFO global exploration stage, fixed value is 50. is the "sun" formula, that is, the coating formula with the best fitness in the SFO stage, which is used as the search target of other candidate formulas; (2) Calculate the step size based on the inverse square law: ( (These are the upper and lower limits of the content of each component). The maximum step size in the SFO stage is calculated based on the inverse square law and is used to control the maximum magnitude of the movement of the candidate formulation towards the "sun" formulation. This represents the upper limit of the content of each component, i.e., the highest allowable content of various components (resin, anti-rust pigment, etc.) in the coating in industrial applications. This represents the lower limit of the content of each component, i.e., the minimum allowable content of each component in the coating for industrial applications; (3) Generate new formulas through pollination operations Remove "dead plants" with poor adaptability and update the position of the "sun". Expand the scope of formula search. The new formulations generated in the SFO stage are obtained by updating existing candidate formulations through pollination operations, thereby expanding the formulation search scope.

[0049] When the index value of the recipe candidate set At that time, start GWO to perform a local search: (1) Initialize the "Grey Wolf" population (select from high-quality recipes searched by SFO, scale) ), The gray wolf population size, i.e., the number of candidate formulations in the GWO local search phase, is selected from the high-quality formulations searched in the SFO, and is fixed at 30. The top 3 coating formulations in terms of fitness are set as follows: (leader ), (Secondary Leader) ), (Executor) ); The (leader) recipe, namely the best recipe with the highest fitness in the GWO phase, dominates the local search direction. (Secondary Boss) Recipe, namely the second-best recipe in terms of adaptability during the GWO phase, is an auxiliary... Recipe-guided search. (Executor) Formula, namely the high-quality formula ranked third in fitness during the GWO phase, combined with... , Optimize the search path for the formula; (2) Calculate the first coefficient vector Second coefficient vector ( Decrease linearly from 2 to 0. , for (random number); Coefficient decay factor for GWO stage, linearly decreasing from 2 to 0, used to adjust the balance between exploration and exploitation in the search process. rand1 is a random number in the interval [0, 1] used to calculate the coefficient vector rand1 is a random number in the interval [0, 1] used to calculate the coefficient vector , introducing search randomness. rand2 is a random number in the interval [0, 1] used to calculate the coefficient vector rand2 is a random number in the interval [0, 1] used to calculate the coefficient vector , enhancing search diversity. B is the first coefficient vector of the GWO stage, with the formula , used to control the quality of the candidate formula to approach or move away from the optimal formula, adjusting the intensity of local search. is the second coefficient vector of the GWO stage, with the formula , used to further enhance search randomness and avoid falling into local optima; (3) Other "gray wolves" (ordinary formulas ) update their formulas according to the updates of the alpha , beta , and delta : When , focus on local optimization and fine-tune component content. is the ordinary formula of the GWO stage (the formula corresponding to the th gray wolf), which is updated by referring to the positions of , , , is the updated ordinary formula of the GWO stage, with the formula , which is based on the average position adjustment of , ,

[0050] Step 5, set the optimization target and constraint conditions, and construct a weighted multi-objective optimization function.

[0051] Step s51, set the constraint conditions:

[0052] Set the minimum threshold of the core performance of the coating to filter out formulas that meet the engineering requirements: (1) Antifouling performance: 28-day biological attachment coverage , antifouling effective period months, antifouling agent bleeding rate ; wherein is the 28-day biological attachment coverage, i.e., the area ratio of marine organisms (diatoms, barnacles, etc.) attached to the coating after 28 days of service in a marine environment, which is a key constraint indicator of antifouling performance. Antifouling effective period, i.e., the duration of the coating's effective antifouling function, is a key constraint indicator of antifouling performance. Antifouling agent leaching rate, which is the rate at which the antifouling agent in the coating seeps into the external environment, is measured in units of... It needs to be within a reasonable range to ensure the anti-fouling effect and durability; (2) Mechanical properties: adhesion (Pull-out method), abrasion resistance Turn, flexibility ,hardness ; in, Adhesion, or the bonding strength between the coating and the substrate, is tested using the pull-off method and is a core constraint indicator of mechanical properties. Taber abrasion resistance is the coating's ability to resist wear, measured in feet (µm). Rotation is a key constraint indicator for mechanical properties. Flexibility refers to the coating's ability to withstand bending deformation without breaking. It is measured by the diameter of the bending test and is expressed in mm. It is an important constraint indicator of mechanical properties. Pencil hardness, or the coating's ability to resist scratching by hard objects, is one of the limiting indicators of mechanical properties. (3) Cost constraints: Cost of coating materials per square meter Yuan; in, The cost of materials per square meter of coating, i.e., the material cost required to prepare 1 square meter of coating, is an economic constraint indicator for formula screening.

[0053] Step s52, construct the weighted multi-objective optimization function: ; in, The weighted multi-objective optimization function value is used to comprehensively evaluate the quality of coating formulations, with the objective of maximizing this function value. The weight of the antifouling effectiveness period is used to adjust the importance of the antifouling effectiveness period in multi-objective optimization, and is fixed at 0.4. The weight of biofilm coverage is used to adjust the importance of biofilm coverage in multi-objective optimization, and is fixed at 0.2. The weight for wear resistance is used to adjust the importance of wear resistance in multi-objective optimization, and is fixed at 0.2. The weight of material cost is used to adjust the importance of material cost in multi-objective optimization, and is fixed at 0.2. The wear resistance constraint threshold, i.e., the maximum allowable wear of the coating, is fixed at a value. The normalization function is used to normalize the wear resistance index. The material cost constraint threshold, i.e. the maximum allowable material cost per square meter of coating, is fixed at 200 yuan and is used to normalize the cost index.

[0054] Step 6: The Hybrid Grey Wolf-Sunflower Optimization Algorithm (HGWSFO) iterates continuously, outputting a weighted multi-objective optimization function value that satisfies the constraints of Step 4. The largest and most optimal formula.

[0055] Algorithm Iteration After the round, the output satisfies the constraints and the objective function. The top 10 formulations are used as the candidate set, and the performance prediction for each formulation (including confidence intervals for each metric) is output. ) and recommended application sea areas (based on environmental adaptability matching). Among them, The number of iterations for algorithm optimization, i.e., the total number of iterations of the Hybrid Gray Wolf-Sunflower Optimization Algorithm (HGWSFO), is fixed at 50 rounds. This refers to the predicted value of a performance metric, which is an estimated value of a certain performance metric output by the multi-task performance prediction network. The standard deviation of the predicted values ​​of the performance index is used to measure the dispersion of the predicted values ​​and reflect the accuracy of the prediction. The confidence interval of a performance indicator is the range in which the actual value of the performance indicator is likely to fall, and it is used to assess the reliability of the prediction of the formulation performance.

[0056] Comparative Example

[0057] The model of this invention is compared with the Transformer, GRU, and LSTM models.

[0058] like Figure 2 As shown, the accuracy of the four models all showed a steady upward trend with the increase of training times, which is consistent with the convergence law of model training. However, the algorithm of this invention, with the synergistic advantages of seed-guided neurometric learning technology, Local Attention Memory (LAM) enhanced encoder and Hybrid Gray Wolf-Sunflower Optimization Algorithm (HGWSFO), maintained the highest accuracy in each training stage. After 100 training rounds, the accuracy reached 0.92, which is 9, 14 and 13 percentage points higher than Transformer (0.83), GRU (0.78) and LSTM (0.79), respectively. This fully verifies the technical superiority of this invention in multimodal data fusion, accurate feature representation and formulation optimization, and can provide more reliable support for the prediction of the performance of the pollution barrier coating.

[0059] 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 screening formulations of multifunctional anti-corrosion and anti-fouling coatings for debris-blocking nets, characterized in that, Includes the following steps: Step 1: Collect core data to build a database, perform preprocessing to obtain the preprocessed database, and use seed-guided neural metric learning to achieve deep fusion of multimodal data on "coating composition-performance-environment-failure". By selecting representative seed samples, generate a seed sample set with similarity labels and construct a similarity matrix. S ; Step 2, Construct the grid-based memory tensor M Stores local feature information of seed samples, similarity matrix S Local features are preferentially written into the grid base memory tensor. M A LAM-GRU encoder was constructed to obtain the coating component vector. Environment vectors In grid-based memory tensors M The corresponding grid coordinates As input, output is the multidimensional feature vector of the coating. E Its dimensions are d ; Step 3: Construct a multi-task performance prediction network using the multi-dimensional feature vector of the coating. E The network takes four performance metrics as input and outputs predicted values ​​for four types of performance metrics. A loss function is constructed to assign differentiated weights to the four types of performance metrics, and the Adam optimizer is used to optimize the parameters of the multi-task performance prediction network. Step 4: Efficiently optimize coating formulations using HGWSFO: Dynamically switch between gray wolf-optimized local search and sunflower-optimized global search by using index thresholds; Step 5: Set the optimization objective and constraints, construct a weighted multi-objective optimization function, and continuously iterate HGWSFO to output the optimal formula that satisfies the constraints and maximizes the objective function.

2. The method for screening the formulation of the multifunctional anti-corrosion and anti-fouling coating for the debris-blocking net as described in claim 1, characterized in that, In step 1, the data in the database includes coating formulation component data, performance data, environmental data, and failure data; preprocessing includes data standardization, outlier removal, and data completion and enhancement.

3. The method for screening the formulation of a multifunctional anti-corrosion and anti-fouling coating for a debris-blocking net as described in claim 1, characterized in that, In step 1, the similarity matrix , This refers to the number of representative coating formulation components randomly sampled from the pre-processed database.

4. The method for screening the formulation of the multifunctional anti-corrosion and anti-fouling coating for the debris-blocking net as described in claim 1, characterized in that, In step 2, the grid base memory tensor , P , Q These represent the number of grid divisions for the components or environmental parameters in the first and second dimensions, respectively.

5. The method for screening the formulation of the multifunctional anti-corrosion and anti-fouling coating for the debris-blocking net as described in claim 1, characterized in that, In step 2, the multidimensional feature vector of the coating is output through the read and write operations of the LAM-GRU encoder. E The specific steps include the following: (1) Reading operation: based on coating component vector or environment vector In grid-based memory tensors M The corresponding grid coordinates Scan local bandwidth w The neighborhood grid, through attention weights Fusion grid-based memory tensor Historical characteristics in the process of generating historical states ; (2) Write operation: using the bureau department Control the update intensity and update the current feature state. The weighted update to the memory tensor is calculated using the following formula: ； For bureau departments; This represents the current feature state; It is the sigmoid activation function; grid coordinates The corresponding original feature values; grid coordinates The corresponding updated feature values; (3) The final hidden layer output of the LAM-GRU encoder is used as the multidimensional feature vector. E To achieve a unified representation of multimodal data; ； This is the weight matrix. For bias vectors, This is the final time step after the LAM-GRU encoder has traversed all input sequences. The hidden layer state, This enhances the nonlinear representation of features for activation functions.

6. The method for screening the formulation of the multifunctional anti-corrosion and anti-fouling coating for the debris-blocking net as described in claim 1, characterized in that, In step 3, the formula for the multi-task performance prediction network is expressed as follows: ； For the first k Predicted values ​​of class performance indicators, These correspond to four performance categories: corrosion resistance, antifouling performance, mechanical properties, and environmental adaptability. These are the weight matrix and bias vector of the fully connected layer 1 of the branch, respectively. For the hidden layer dimension; These are the weight matrix and bias vector of the branch output layer, respectively. For the first k The number of performance metrics for a class; For the hidden layer activation function, This is the activation function for the branch output layer.

7. The method for screening the formulation of a multifunctional anti-corrosion and anti-fouling coating for a debris-blocking net as described in claim 1, characterized in that, In step 3, the loss function is: ; To calculate the weighted total loss across multiple tasks, For the first Class performance weights For the first Class performance The true value of each indicator For the first Class performance The predicted values ​​of each indicator.

8. The method for screening the formulation of the multifunctional anti-corrosion and anti-fouling coating for the debris-blocking net as described in claim 1, characterized in that, In step 4, the index threshold is set. When the index value of the candidate recipe set When the index value of the recipe candidate set is reached, SFO is initiated to perform a global exploration. At that time, GWO will be launched to perform a local search.

9. The method for screening the formulation of the multifunctional anti-corrosion and anti-fouling coating for the debris-blocking net as described in claim 8, characterized in that, The specific steps for global SFO exploration are as follows: (1) Initialize the "Sunflower" population and set the coating formula with the best fitness to "Sun". ; (2) Calculate the step size based on the inverse square law: ,in This is the maximum step size in the SFO stage, used to control the maximum magnitude by which the candidate formulation moves towards the "sun" formulation. These represent the upper and lower limits of the content of each component. This refers to the size of the sunflower population. (3) Generate new formulas through pollination operations Remove "dead plants" with poor adaptability and update the "sun". ; The specific steps of GWO local search are as follows: (1) Initialize the "gray wolf" population and set the top three coating formulas with the highest fitness as leaders. Secondary leaders Executor ; (2) Calculate the coefficient vector: the first coefficient vector of the GWO stage The second coefficient vector of the GWO stage , The coefficient decay factor for the GWO stage. The value decreases linearly from 2 to 0, rand1, All Random numbers within the interval; The first coefficient vector B is used to control whether the candidate formulation moves closer to or further away from the high-quality formulation. The second coefficient vector... Used to enhance search randomness; (3) Standard Formula According to the leader Secondary leaders Executor Update to obtain the updated standard recipe. .

10. The method for screening the formulation of the multifunctional anti-corrosion and anti-fouling coating for the debris-blocking net as described in claim 1, characterized in that, In step 5, the weighted multi-objective optimization function is: ; in: To optimize the function value for weighted multi-objectives; Weighting of the effective period of anti-fouling treatment; The weight for bioattachment coverage; The bio-attachment coverage rate is 28 days. Weighting for wear resistance; Weighting for material costs; For wear resistance; The wear resistance constraint threshold; Cost of coating materials per square meter; This is the threshold for material cost constraints.

Images