Method for optimizing formula of medicament for preventing and treating powdery mildew

Abstract

The invention provides a powdery mildew prevention and treatment agent formula optimization method, and belongs to the technical field of computational chemistry, the powdery mildew prevention and treatment agent formula optimization method comprises the following steps: firstly, constructing performance vectors containing 11 parameters such as tripotassium phosphate concentration, fatty acid concentration and the like, and establishing an experimental database; then collecting training data and carrying out standardization processing; constructing a deep neural network model based on an attention mechanism, wherein the deep neural network model comprises multiple attention layers and three hidden layers; and finally, through experimental verification and iterative optimization, precise regulation and control of a medicament formula are realized. According to the method, a systematic optimization strategy is adopted, intelligence and automation of the formula optimization process are achieved, and the technical problem that in the prior art, automatic optimization and precise regulation and control of a powdery mildew prevention and treatment agent formula cannot be achieved is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of computational chemistry, and in particular relates to a method for optimizing a powdery mildew control agent formula. Background Art

[0002] Powdery mildew is one of the common diseases of crops, which is mainly controlled by chemical agents. The development of traditional powdery mildew control agent formulas mainly relies on empirical experimental methods and single-factor experimental methods, and screening is carried out by setting different concentration gradients. In practice, researchers usually use orthogonal experimental design to determine the distribution ratio of each group, and then conduct indoor antibacterial and field tests to evaluate the efficacy. Currently, the more commonly used control agents include a compound system of tripotassium phosphate and fatty acids, which have the advantages of environmental protection and low residue.

[0003] However, traditional methods for developing pharmaceutical formulations have obvious shortcomings. First, the empirical test method requires a large number of repeated experiments, which is time-consuming, labor-intensive and inefficient; second, the single-factor test method makes it difficult to examine the interactions between the components and is prone to miss the optimal ratio; third, although orthogonal experiments can reduce the number of experiments, they cannot achieve dynamic optimization and precise regulation of the formula. In addition, existing methods often only focus on a single performance indicator, such as antibacterial effect or preventive effect, and ignore the overall evaluation of the comprehensive performance of the agent. This leads to problems such as unstable efficacy and risk of drug damage in the developed pharmaceutical formula.

[0004] Based on the above problems, there is an urgent need for a method that can realize the automated optimization of powdery mildew control agent formulas. The existing technology makes it difficult to establish an accurate mapping relationship between agent performance and formula parameters, and lacks a systematic formula optimization strategy. At the same time, due to the influence of variables such as environmental factors and crop varieties, the formula optimization process is more complicated. How to build an intelligent formula optimization system to achieve multi-dimensional evaluation of agent performance and precise regulation of formulas has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] In view of this, the present invention provides a method for optimizing the formula of a powdery mildew control agent, which can solve the technical problem in the prior art that it is impossible to achieve automated optimization and precise regulation of the powdery mildew control agent formula.

[0006] The present invention is implemented as follows: The present invention provides a method for optimizing the formula of a powdery mildew control agent, comprising the following steps: constructing a powdery mildew control agent performance vector including parameters such as a tripotassium phosphate concentration value and a fatty acid concentration value; collecting multiple groups of experimental data of different ratios of tripotassium phosphate and fatty acids to establish a performance vector database; collecting a training data set; standardizing the training data set and dividing it into a training sample set and a verification sample set; constructing a deep neural network model including an input layer, a first attention layer, a second attention layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer, and using the training sample set to train and generate a powdery mildew control agent formula optimization model; preparing experimental samples according to the ratio range of tripotassium phosphate and fatty acids and performing performance tests; inputting the test data into the powdery mildew control agent formula optimization model; calculating a performance gain vector including a tripotassium phosphate concentration adjustment amount; optimizing and adjusting the ratio of tripotassium phosphate and fatty acids based on the performance gain vector until the performance gain vector meets a preset threshold requirement.

[0007] This also includes collecting multiple sets of experimental data of different ratios of tripotassium phosphate and fatty acids, and establishing a performance vector database. The experimental data include antibacterial rate values, prevention efficiency values ​​and phytotoxicity grade values.

[0008] Among them, it includes a group of equations for evaluating the performance of the agent, and the group of equations for evaluating the performance of the agent includes an equation for evaluating antibacterial activity, an equation for evaluating the protective effect, an equation for evaluating the toxicity of the agent, and an equation for calculating the comprehensive performance.

[0009] Among them, the deep neural network model adopts a deep learning architecture based on a sparse attention mechanism. The first attention layer performs feature extraction and weight calculation on tripotassium phosphate concentration values ​​and fatty acid concentration values ​​based on a multi-head attention mechanism, and the second attention layer uses a self-attention mechanism to perform feature extraction and correlation analysis on environmental parameters.

[0010] The performance vector is analyzed by principal component analysis to determine key performance indicators, including antibacterial effect, prevention and control effect and phytotoxicity index, and a multidimensional evaluation system is established by analytic hierarchy process.

[0011] The training data set uses a stratified sampling method to extract samples from the experimental database, uses a 3-times standard deviation method to delete outliers, and uses a mean filling method to handle missing values.

[0012] Among them, the minimum and maximum value standardization method is used to normalize the continuous variables of the training data set, and the one-hot encoding method is used to encode the discrete variables.

[0013] The deep neural network model training adopts a small batch stochastic gradient descent method, introduces an early stopping mechanism, and adds a batch normalization layer and a dropout layer after each hidden layer.

[0014] Among them, the experimental samples were prepared according to the ratio range of 1 to 5 parts of tripotassium phosphate and 1 to 2 parts of fatty acids, and the antibacterial activity test, control effect test and phytotoxicity evaluation were carried out on the experimental samples.

[0015] Among them, the performance gain vector includes the adjustment amount of tripotassium phosphate concentration, the adjustment amount of fatty acid concentration, the predicted antibacterial rate increment value, the predicted prevention efficiency increment value and the predicted phytotoxicity level improvement value.

[0016] Compared with the prior art, the present invention provides a method for optimizing the formulation of a powdery mildew control agent, which systematically collects and manages experimental data by constructing performance vectors and establishing a database; adopts a deep neural network model to establish a nonlinear mapping relationship between formulation parameters and agent performance; and realizes automatic extraction and weight calculation of key features through an attention mechanism, providing intelligent decision support for formulation optimization.

[0017] The technical solution of the present invention overcomes the limitations of traditional methods. Through a multidimensional performance evaluation system, a comprehensive evaluation of multiple indicators such as antibacterial effect, prevention and control effect, and pesticide damage is achieved; based on the prediction function of the deep learning model, the number of experiments is significantly reduced and the efficiency of formula development is improved; the introduction of the attention mechanism and multi-layer feature extraction structure enhances the model's adaptability to environmental factors and crop characteristics; and the adoption of an iterative optimization strategy realizes dynamic adjustment and precise control of formula parameters.

[0018] The present invention solves the problem of automated optimization of powdery mildew control agent formulas, which is mainly reflected in the following aspects: first, a standardized performance evaluation method is established to achieve quantitative characterization of agent performance; second, an intelligent optimization model is constructed, which can automatically learn formula rules and give optimization suggestions; third, a complete optimization process is designed to ensure the scientificity and reliability of formula adjustment. Through the organic combination of the above-mentioned technical means, accurate optimization of the agent formula is achieved, the control effect is improved and the risk of pesticide damage is reduced. In summary, the present invention solves the technical problem that the prior art cannot achieve automated optimization and precise regulation of powdery mildew control agent formulas. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0021] like Figure 1 FIG. 1 is a flow chart of a method for optimizing a powdery mildew control agent formulation provided by the present invention, and the method comprises the following steps: S01, constructing a performance vector of a powdery mildew control agent, including establishing a multidimensional evaluation matrix of antibacterial effect index, control effect index and pesticide damage index, wherein the performance vector of the powdery mildew control agent includes a tripotassium phosphate concentration value, a fatty acid concentration value, an ambient temperature value, a relative humidity value, an agent pH value, a crop variety number, an incidence degree value, a pesticide application number value, a crop growth period value, a leaf area value and an agent contact time value; S02, collecting multiple groups of experimental data of different ratios of tripotassium phosphate and fatty acids, and establishing a performance vector database, wherein the experimental data includes antibacterial rate value, anti-efficiency value and phytotoxicity grade value; S03, collecting a training data set, wherein the training data set includes tripotassium phosphate concentration value, fatty acid concentration value, ambient temperature value, relative humidity value, agent pH value, crop type number, disease severity value, application frequency value, crop growth period value, leaf area value, agent contact time value, antibacterial rate value, prevention efficiency value and pesticide damage grade value; S04, performing standardization processing on the training data set to generate a standardized training data set, and dividing the standardized training data set into a training sample set and a verification sample set in a ratio of 8 to 2; S05, constructing a deep neural network model, the deep neural network model comprising an input layer, a first attention layer, a second attention layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer; using the training sample set to train the deep neural network model to generate a powdery mildew control agent formula optimization model; S06. Prepare multiple groups of experimental samples in a ratio range of 1 to 5 parts of tripotassium phosphate and 1 to 2 parts of fatty acids; perform antibacterial activity tests on the experimental samples, measure the diameter of powdery mildew lesions, and calculate the average inhibition rate of lesions; perform control effect tests on the experimental samples, record the disease index values ​​according to the plant disease and insect pest classification standards, and calculate the relative control effect values; perform pesticide damage evaluation on the experimental samples, record the plant leaf pesticide damage level values; and use the pesticide performance evaluation equation group to calculate the comprehensive performance index value; S07, inputting the average lesion inhibition rate value, relative prevention effect value, pesticide damage level value and comprehensive performance index value into the powdery mildew control agent formula optimization model; S08, calculating a performance gain vector through the powdery mildew control agent formulation optimization model, wherein the performance gain vector includes a tripotassium phosphate concentration adjustment amount, a fatty acid concentration adjustment amount, a predicted antibacterial rate increment value, a predicted control efficiency increment value, and a predicted pesticide damage grade improvement value; S09, optimizing and adjusting the ratio of tripotassium phosphate to fatty acids based on the performance gain vector; repeating steps S06 to S08 until the performance gain vector meets a preset threshold requirement; and outputting an optimized powdery mildew control agent formula.

[0022] Optionally, the method further includes verifying the optimized powdery mildew control agent formula using the least squares method.

[0023] The agent performance evaluation equation group includes an antibacterial activity evaluation equation, a protective effect evaluation equation, a pesticide damage evaluation equation and a comprehensive performance calculation equation; the antibacterial activity evaluation equation is used to calculate the relationship function between the lesion inhibition rate and the concentration of tripotassium phosphate and the concentration of fatty acids, and the input includes the concentration value of tripotassium phosphate, the concentration value of fatty acids, the value of ambient temperature, the value of relative humidity, the pH value of the agent and the value of the number of times the pesticide is applied, and the output is the predicted value of the lesion inhibition rate; the protective effect evaluation equation is used to calculate the relationship function between the relative protective effect and the concentration of tripotassium phosphate and the concentration of fatty acids, and the input includes the concentration value of tripotassium phosphate, the concentration value of fatty acids, the crop type number, the disease severity value, The numerical value of the number of pesticide applications and the crop growth period value are used, and the output is the relative prevention effect prediction value; the pesticide damage evaluation equation is used to calculate the relationship function between the pesticide damage grade and the tripotassium phosphate concentration and the fatty acid concentration. The input includes the tripotassium phosphate concentration value, the fatty acid concentration value, the crop growth period value, the leaf area value, the pesticide contact time value and the pesticide pH value, and the output is the pesticide damage grade prediction value; the comprehensive performance calculation equation is used to integrate the antibacterial activity, prevention effect and pesticide damage evaluation results. The input includes the spot inhibition rate prediction value, the relative prevention effect prediction value, the pesticide damage grade prediction value, the first weight coefficient value, the second weight coefficient value and the target threshold value, and the output is the comprehensive performance index value.

[0024] The deep neural network model adopts a deep learning architecture based on a sparse attention mechanism. The overall structure includes an input layer, a first attention layer, a second attention layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer. The input layer receives standardized tripotassium phosphate concentration values, fatty acid concentration values, ambient temperature values, relative humidity values, agent pH values, crop variety numbers, disease severity values, application times values, crop growth period values, leaf area values ​​and agent contact time values. The first attention layer performs feature extraction and weight calculation on tripotassium phosphate concentration values ​​and fatty acid concentration values ​​based on a multi-head attention mechanism. The second attention layer performs feature extraction and weight calculation on tripotassium phosphate concentration values ​​and fatty acid concentration values ​​based on a multi-head attention mechanism. The attention layer uses a self-attention mechanism to extract features and perform correlation analysis on environmental parameters, including ambient temperature, relative humidity, and pH value of the agent. The first hidden layer uses a fully connected structure of 256 neurons to perform preliminary transformation on the extracted features. The second hidden layer uses a fully connected structure of 128 neurons to extract middle-level features. The third hidden layer uses a fully connected structure of 64 neurons to perform high-level feature fusion. The first hidden layer, the second hidden layer, and the third hidden layer all use ReLU activation functions for nonlinear feature transformation. The output layer uses a linear activation function to output The performance gain vector includes a concentration adjustment amount of tripotassium phosphate, a concentration adjustment amount of fatty acids, a predicted antibacterial rate increment, a predicted anti-efficiency increment, and a predicted phytotoxicity grade improvement value; the training data set of the deep neural network model is constructed by a multi-source data fusion method, including literature record data and field test data. After being standardized, the training data set is divided into a training sample set and a verification sample set in a ratio of 8 to 2. The training process adopts a small batch stochastic gradient descent method with a batch size of 32, and the initial value of the learning rate is set to 0.001. The training of the deep neural network model adopts Adam The optimization algorithm iteratively updates the network parameters. The loss function uses the mean square error function to calculate the error between the predicted value and the true value. An early stopping mechanism is introduced into the training process. When the loss value on the validation set does not decrease for 10 consecutive epochs, the training is stopped. At the same time, a batch normalization layer is added after each hidden layer to reduce the internal covariate shift and improve the training stability. A Dropout layer with a dropout rate of 0.2 is introduced after each hidden layer to prevent overfitting. The deep neural network model completes training when the mean square error on the validation sample set is less than 0.01, which is used in the subsequent powdery mildew control agent formulation optimization process.

[0025] The specific implementation of the above steps is described in detail below.

[0026] The specific implementation method of step S01 is to construct the performance vector of powdery mildew control agent. First, the principal component analysis method is used to determine the key performance indicators, including antibacterial effect, control effect and drug damage index. A multidimensional evaluation system is established through the hierarchical analysis method, and the weight coefficient of each index is set. The antibacterial effect weight is 0.4, the control effect weight is 0.4, and the drug damage index weight is 0.2. Then, 11 parameters such as tripotassium phosphate concentration value and fatty acid concentration value are normalized by data standardization method, and each parameter value is mapped to the interval of 0 to 1 to ensure the comparability of different dimension indicators. Finally, a performance vector data structure is established, and a sparse matrix storage method is used to reduce storage space occupancy and improve data processing efficiency. The purpose of this step is to provide a standardized input data format for the subsequent optimization model.

[0027] The specific implementation method of step S02 is to collect experimental data and establish a performance vector database. First, an orthogonal test scheme is designed to determine the experimental combination of tripotassium phosphate and fatty acid ratio. The concentration range of tripotassium phosphate is 1% to 5%, and the concentration range of fatty acids is 0.5% to 2%. The number of experimental groups is not less than 20. Then a laboratory antibacterial test is carried out, and the antibacterial rate is determined by the colony counting method. The test temperature is 25 degrees Celsius, the relative humidity is 60%, and the incubation time is 72 hours. At the same time, a greenhouse prevention test is carried out, and the prevention and control effect is evaluated according to the 0 to 4 disease index standard, and the observation period is 14 days. In addition, a drug damage evaluation test is carried out, and the degree of leaf damage is recorded using the 0 to 3 drug damage classification standard. Finally, a structured query language is used to establish an experimental database, which includes fields such as experimental number, formula parameters, performance indicators, and data integrity constraints are set to ensure data quality. The role of this step is to accumulate basic data required for optimization.

[0028] The specific implementation method of step S03 is to collect a training data set. First, a stratified sampling method is used to extract representative samples from the experimental database, and the sample size is not less than 200 groups. Then data cleaning is performed, and the 3 times standard deviation method is used to delete outliers, and the mean filling method is used to handle missing values. Then feature engineering is performed, the correlation coefficient matrix between each parameter is calculated, and the variance inflation factor method is used to detect multicollinearity and delete redundant features. Finally, the data set is randomly shuffled to eliminate data order dependency. The purpose of this step is to build a high-quality model training data set.

[0029] The specific implementation method of step S04 is to perform standardization on the training data set. First, the minimum and maximum value standardization method is used to normalize the continuous variables, including numerical features such as tripotassium phosphate concentration value and fatty acid concentration value. Then, the unique hot encoding method is used to encode the discrete variables, such as the crop type number. Then, the amplitude scaling method is used to uniformly process the indicators of different dimensions so that the values ​​of each indicator are distributed in a similar interval. Finally, the training sample set and the verification sample set are randomly divided according to the ratio of 8 to 2, and the distribution consistency of each type of sample is maintained. The role of this step is to improve the model training effect.

[0030] The specific implementation method of step S05 is to construct a deep neural network model. First, the network structure is designed. The number of nodes in the input layer is 11, corresponding to 11 input features. The first attention layer adopts an 8-head attention mechanism, and the number of attention heads is determined by a grid search method. The second attention layer adopts a self-attention mechanism, and the attention weight is learned by a back propagation algorithm. The first hidden layer uses 256 neurons, the second hidden layer uses 128 neurons, and the third hidden layer uses 64 neurons. Batch normalization layers and random inactivation layers are used between each hidden layer. The momentum parameter of the batch normalization layer is set to 0.9, and the random inactivation rate is set to 0.2. The output layer uses a linear activation function and outputs 5 performance gain indicators. Then the model is trained, using a small batch stochastic gradient descent method, the batch size is set to 32, the initial learning rate is set to 0.001, and the cosine annealing strategy is used to dynamically adjust the learning rate. The early stopping strategy is used during the training process, and the training is stopped when the validation set loss does not decrease for 10 consecutive rounds. Finally, the model is evaluated, the mean square error is calculated on the validation set, and the training is completed when the mean square error is less than 0.01. The purpose of this step is to establish a performance optimization prediction model.

[0031] The specific implementation method of step S06 is to prepare experimental samples and perform performance tests. First, a formula combination is designed according to the ratio range of 1 to 5 parts of tripotassium phosphate and 1 to 2 parts of fatty acids. The experimental gradient is determined by the single factor method. The concentration gradient of tripotassium phosphate is 1%, 2%, 3%, 4%, and 5%, and the concentration gradient of fatty acids is 1%, 1.5%, and 2%. 15 groups of experimental samples are formed. Then, an antibacterial activity test is performed, and the growth of pathogenic bacteria colonies is determined by the plate method. The samples are cultured in a constant temperature incubator at 25 degrees Celsius for 72 hours, and the diameter of the lesions is measured by a vernier caliper. Each sample is tested repeatedly 3 times, and the average value is taken to calculate the lesion inhibition rate. Then, a control effect test is performed, and typical powdery mildew-infected crops are selected. The disease is recorded according to the disease index standard of 0 to 4 levels, and the statistics are made every 7 days. Observe 3 times continuously and calculate the relative prevention value. In addition, a pesticide damage evaluation test is performed, and the damage to plant leaves is observed and recorded according to the pesticide damage classification standard of 0 to 3 levels, and the evaluation period is 14 days. Finally, the comprehensive performance index is calculated using the drug performance evaluation equation group, where the first weight coefficient is 0.4, the second weight coefficient is 0.4, and the target threshold is set to 0.8. The purpose of this step is to obtain experimental verification data.

[0032] The specific implementation method of step S07 is to input the test data into the optimization model, first pre-process the test data, and normalize the average inhibition rate value of the lesions, the relative prevention value, the pesticide level value and the comprehensive performance index value using the same standardization method as the training data. Then convert the processed data into the model input format, organize the data in batch mode, and improve the data processing efficiency. Then perform data verification to check the data integrity and validity to ensure the quality of the input data. Finally, input the verified data into the powdery mildew control agent formula optimization model. The purpose of this step is to prepare the input data required for model prediction.

[0033] The specific implementation method of step S08 is to calculate the performance gain vector. First, the input data is forward propagated through the deep neural network model. The first attention layer extracts the characteristics of tripotassium phosphate concentration and fatty acid concentration, and the multi-head attention mechanism is used to calculate the feature weight. The second attention layer processes the environmental parameter characteristics and uses the self-attention mechanism to analyze the parameter correlation. Then, after nonlinear transformation of three hidden layers, each layer uses a modified linear unit activation function for feature extraction. Finally, the performance gain vector is calculated through the output layer, including the recommended formula adjustment amount and the expected performance improvement index. When the predicted performance gain value is greater than the preset threshold, the optimization suggestion is output, and the threshold reference value is set to 0.1. The purpose of this step is to generate a formula optimization plan.

[0034] The specific implementation method of step S09 is to optimize the formula according to the performance gain vector. First, the performance gain vector is analyzed to calculate the adjustment direction and amplitude of the ratio of tripotassium phosphate to fatty acids. Then the formula is adjusted, the adjustment step length of tripotassium phosphate is 0.5%, and the adjustment step length of fatty acids is 0.2%. Experimental verification is performed after each adjustment. Then the optimization effect is evaluated. When the increment of the antibacterial rate is greater than 5%, the increment of the anti-efficiency is greater than 5%, and the phytotoxicity level decreases by more than 0.5 levels, the optimization is considered to be effective. If the performance gain does not meet expectations, return to step S08 to recalculate the optimization scheme. Finally, when the change in the performance gain vector is less than 0.05 after three consecutive optimizations, the optimization process is stopped and the final formula is output. Optionally, the optimized formula is verified by the least squares method, and the residual sum of squares between the predicted value and the actual value is calculated. When the residual sum of squares is less than 0.1, it is confirmed that the optimization result is reliable. The purpose of this step is to iteratively optimize and obtain the optimal formula.

[0035] The mathematical model or calculation process involved in the present invention is described in detail below.

[0036] 1. Pharmaceutical performance vector construction equation: The performance vector of the powdery mildew control agent is specifically expressed as follows: ; In the formula, is the performance vector; is the concentration value of tripotassium phosphate; is the fatty acid concentration value; is the ambient temperature value; is the relative humidity value; is the pH value of the agent; Number the crop types; is the disease severity value; is the number of times the pesticide is applied; is the crop growth period value; is the leaf area value; is the agent contact time value; to is the weight coefficient.

[0037] 2. Antibacterial activity evaluation equation: The antibacterial activity evaluation equation is specifically expressed as follows: ; In the formula, is the predicted value of lesion inhibition rate; is the maximum concentration of tripotassium phosphate; is the maximum concentration of fatty acids; is the optimal temperature value; is the maximum relative humidity value; is the optimal pH value; is the maximum number of times of application; is the temperature standard deviation; is the pH standard deviation; to is the weight coefficient; is the error term.

[0038] 3. Prevention effectiveness evaluation equation: The prevention effect evaluation equation is specifically expressed as follows: ; In the formula, is the predicted value of relative protection efficacy; is the power coefficient; to is the weight coefficient; is the error term.

[0039] 4. Phytotoxicity assessment equation: The phytotoxicity assessment equation is specifically expressed as follows: ; In the formula, is the predicted value of pesticide damage level; to is the weight coefficient; is the error term.

[0040] 5. Comprehensive performance calculation equation: The comprehensive performance calculation equation is specifically expressed as follows: ; In the formula, is the comprehensive performance index value; is the weight coefficient; is the error term.

[0041] Parameter acquisition method description: 1. Weight coefficient to , to , to , to It is obtained through principal component analysis. The specific steps include: Step 1: Construct a sample matrix; Step 2: Standardization; Step 3: Calculate the correlation coefficient matrix; Step 4: Eigenvalue decomposition; Step 5: Calculate the contribution rate; Step 6: Determine the number of principal components; Step 7: Calculate the weight coefficient.

[0042] 2. Error term to The range is -0.1~0.1.

[0043] 3. Power coefficient It is obtained by nonlinear regression method and ranges from 0.5 to 2.0.

[0044] Optional, data normalization equation: The standardization process is specifically expressed as follows: ; In the formula, is the standardized value; is the original value; is the minimum value; is the maximum value.

[0045] Optionally, the performance gain vector is calculated using the equation: The performance gain vector is specifically expressed as follows: ; In the formula, is the performance gain vector; is the concentration adjustment amount of tripotassium phosphate; is the fatty acid concentration adjustment amount; To predict the incremental value of the inhibition rate; To predict the incremental value of defense efficiency; To predict the improvement value of pesticide injury level.

[0046] Optionally, the loss function calculation equation is: The mean square error loss function is specifically expressed as follows: ; In the formula, is the loss function value; is the sample size; is the true value; is the predicted value.

[0047] Specifically, the principle of the present invention is: the technical principle of the present invention is based on deep learning and multidimensional evaluation theory. The deep neural network model can effectively capture the complex relationship between formulation parameters and drug performance through multi-layer nonlinear transformation. The introduction of the attention mechanism enables the model to automatically identify important features and improve the accuracy of feature extraction. At the same time, the establishment of a multidimensional evaluation system provides a scientific decision-making basis for formulation optimization.

[0048] The technical solution adopted by the present invention has sufficient theoretical support. First, the deep learning model has a strong nonlinear fitting ability and can handle complex mapping relationships in high-dimensional feature space; second, the attention mechanism realizes adaptive screening of key factors by learning feature weights; third, the use of technologies such as batch normalization and random inactivation improves the training stability and generalization ability of the model; finally, the iterative optimization strategy ensures the continuity and convergence of the formula adjustment.

[0049] The technical solution of the present invention forms a complete logical chain. From data collection to model construction, from performance evaluation to optimization and adjustment, each link is closely connected to form a closed-loop optimization system. Through experimental verification and theoretical analysis, the feasibility and effectiveness of the scheme in optimizing the formula of powdery mildew control agents are proved.

[0050] A specific embodiment 1 of the present invention is provided below, and the specific implementation method of each step in this embodiment 1 is described in detail as follows.

[0051] The specific implementation of step S01 is to construct a performance vector of the powdery mildew control agent, and determine key performance indicators by principal component analysis, including antibacterial effect indicators, control effect indicators and pesticide damage indicators, wherein the expression of the performance vector is: , a multidimensional evaluation system was established using the analytic hierarchy process, and the weight coefficients of each index were set. The weight of the antibacterial effect was 0.4, the weight of the control effect was 0.4, and the weight of the phytotoxicity index was 0.2. Eleven parameters such as the concentration of tripotassium phosphate and the concentration of fatty acids were normalized. The normalization formula was: , map each parameter value to the interval of 0 to 1, ensure the comparability of indicators of different dimensions, establish the performance vector data structure, use sparse matrix storage to reduce storage space occupancy, and determine the weight coefficient of each indicator through principal component analysis. First, construct the original data matrix, perform centralization, calculate the covariance matrix, solve the eigenvalues ​​and eigenvectors, select the number of principal components according to the cumulative contribution rate, calculate the weight coefficient of each indicator, and finally establish the standardized expression of the performance vector.

[0052] The specific implementation method of step S02 is to collect multiple groups of experimental data of different ratios of tripotassium phosphate and fatty acids to establish a performance vector database. First, an orthogonal test scheme is designed to determine the experimental combination of tripotassium phosphate and fatty acid ratios, and the antibacterial activity evaluation equation is used: , design the experimental parameter range, the concentration range of tripotassium phosphate is 1% to 5%, the concentration range of fatty acids is 0.5% to 2%, the number of experimental groups is not less than 20, and the laboratory antibacterial test is carried out. The colony counting method is used to determine the antibacterial rate. The test temperature is 25 degrees Celsius, the relative humidity is 60%, and the incubation time is 72 hours. At the same time, a greenhouse prevention test is carried out. The control effect is evaluated according to the 0 to 4 disease index standard. The observation period is 14 days. In addition, a drug damage evaluation test is carried out, and the degree of leaf damage is recorded using the 0 to 3 drug damage grading standard. Finally, a structured query language is used to establish an experimental database, which includes fields such as experiment number, formula parameters, and performance indicators, and data integrity constraints are set to ensure data quality.

[0053] The specific implementation method of step S03 is to collect a training data set. First, a stratified sampling method is used to extract representative samples from the experimental database, with a sample size of no less than 200 groups. Then, data cleaning is performed, and the 3-times standard deviation method is used to delete outliers. The mean filling method is used to process missing values. Then, feature engineering is performed to calculate the correlation coefficient matrix between the parameters, and the variance inflation factor method is used to detect multicollinearity, and redundant features are deleted. Finally, the data set is randomly shuffled to eliminate data order dependence. The training data set includes tripotassium phosphate concentration values, fatty acid concentration values, ambient temperature values, relative humidity values, agent pH values, crop variety numbers, disease severity values, application times values, crop growth period values, leaf area values, agent contact time values, antibacterial rate values, prevention efficiency values, and pesticide damage grade values.

[0054] The specific implementation method of step S04 is to standardize the training data set, use the minimum and maximum value standardization method to normalize the continuous variables, including numerical features such as tripotassium phosphate concentration value and fatty acid concentration value, use the one-hot encoding method to encode discrete variables, such as crop type number, and use the amplitude scaling method to unify the indicators of different dimensions so that the values ​​of each indicator are distributed in a similar interval. The training sample set and the verification sample set are randomly divided according to the ratio of 8 to 2, and the distribution consistency of each type of sample is maintained. The standardization processing formula is: ,in is the standardized value, is the original value, is the minimum value, is the maximum value.

[0055] The specific implementation method of step S05 is to construct a deep neural network model and use a deep learning architecture based on a sparse attention mechanism to optimize the powdery mildew control agent formula. First, the network structure is designed. The input layer is set with 11 input nodes, corresponding to the concentration value of tripotassium phosphate. , fatty acid concentration , Ambient temperature value , relative humidity value , pH value of the reagent , Crop Type No. , severity of disease 、Number of times of application , crop growth period value , Leaf area value and drug contact time value The first attention layer uses an 8-head attention mechanism. The optimal number of attention heads is determined by the grid search method. Each attention head independently learns different feature representations. The inter-head attention weight matrix Through the back propagation algorithm optimization, the attention calculation formula is: ,in is the query matrix, is the key matrix, is the value matrix, is the dimension of the key vector. The second attention layer uses the self-attention mechanism to capture the relationship between environmental parameters. The self-attention weight calculation formula is: ,in is the input feature matrix, is the attention weight matrix. The first hidden layer is configured with 256 neurons and uses the rectified linear unit activation function After nonlinear transformation, the neuron output calculation formula is: ,in is the weight matrix, is the bias vector. The second hidden layer is configured with 128 neurons and also uses the rectified linear unit activation function. The output calculation formula is The third hidden layer is configured with 64 neurons and continues to use the rectified linear unit activation function. The output calculation formula is Add a batch normalization layer between each hidden layer. The batch normalization calculation formula is ,in is the batch mean, is the batch variance, and is a learnable parameter, is a small constant with a value of 0.00001. At the same time, a random deactivation layer is added after each hidden layer, and the deactivation rate is set to 0.2 to prevent the model from overfitting. The output layer uses a linear activation function to output a 5-dimensional performance gain vector, including the potassium phosphate concentration adjustment amount , Fatty acid concentration adjustment , predict the incremental value of the inhibition rate , predict the incremental value of defense efficiency And predict the improvement value of pesticide damage level The network training adopts the small batch stochastic gradient descent method, the batch size is set to 32, the initial learning rate is set to 0.001, and the cosine annealing strategy is used to dynamically adjust the learning rate. The learning rate update formula is ,in For the The learning rate of the step, and are the minimum and maximum learning rates, respectively, is the total number of steps. The loss function uses the mean square error function ,in is the sample size, is the true value, is the predicted value. To prevent overfitting, Regularization term, the regularization coefficient is set to 0.0001. The early stopping strategy is used during training, and the patience parameter is set to 10. When the validation set loss does not decrease for 10 consecutive rounds, the training is stopped. The model parameter is updated using the Adam optimizer, and the momentum parameter Set to 0.9, Set to 0.999. The parameter update formula is ,in is the first-order momentum estimate, is the second-order momentum estimate. During the model training process, performance indicators are calculated on the validation set after each epoch, including mean square error, mean absolute error, and Score. When the mean square error on the validation set is less than 0.01, the model training is considered complete. The model evaluation uses a 5-fold cross-validation method to calculate the performance stability of the model under different data partitions. When the final model is saved, the network structure configuration and the trained parameter weights are saved at the same time to facilitate subsequent calls and iterative optimization. The training and optimization process of the entire deep neural network model is implemented based on the PyTorch deep learning framework, using GPU accelerated computing to improve training efficiency.

[0056] The specific implementation method of step S06 is to prepare experimental samples and perform performance tests. First, a formula combination is designed according to the ratio range of 1 to 5 parts of tripotassium phosphate and 1 to 2 parts of fatty acids. The experimental gradient is determined by the single factor method. The concentration gradient of tripotassium phosphate is 1%, 2%, 3%, 4%, and 5%, and the concentration gradient of fatty acids is 1%, 1.5%, and 2%. 15 groups of experimental samples are formed, and then the antibacterial activity test is performed. The growth of pathogenic bacteria colonies is determined by the plate method. The samples are cultured in a constant temperature incubator at 25 degrees Celsius for 72 hours, and the diameter of the lesions is measured with a vernier caliper. Each sample is tested 3 times, and the average value is taken to calculate the lesion inhibition rate. According to the antibacterial activity evaluation equation: Calculate the predicted inhibition rate, and then conduct the control effect test according to the control effect evaluation equation: To evaluate, select typical powdery mildew-infected crops, record the disease according to the disease index standard of 0 to 4, count once every 7 days, observe for 3 consecutive times, calculate the relative protection value, and conduct pesticide damage evaluation test according to the pesticide damage evaluation equation: Calculate the pesticide damage level, observe and record the plant leaf damage according to the pesticide damage grading standard from 0 to 3, the evaluation period is 14 days, and finally use the comprehensive performance calculation equation: The comprehensive performance index is calculated, where the first weight coefficient is 0.4, the second weight coefficient is 0.4, and the target threshold is set to 0.8.

[0057] The specific implementation method of step S07 is to input the test data into the optimization model, first pre-process the test data, and use the standardized equation: The average inhibition rate of lesions, relative control efficiency, pesticide damage level and comprehensive performance index were normalized, and then the processed data were converted into the model input format. The data were organized in batch mode to improve data processing efficiency. Then the data was verified to check the integrity and validity of the data to ensure the quality of the input data. Finally, the verified data was input into the powdery mildew control agent formulation optimization model. The data input format was the performance vector: .

[0058] The specific implementation of step S08 is to calculate the performance gain vector, perform forward propagation calculation on the input data through the deep neural network model, extract the characteristics of tripotassium phosphate concentration and fatty acid concentration in the first attention layer, calculate the feature weights using the multi-head attention mechanism, process the environmental parameter characteristics in the second attention layer, and analyze the parameter correlation using the self-attention mechanism, then go through the nonlinear transformation of the three hidden layers, each layer uses the rectified linear unit activation function for feature extraction, and finally calculate the performance gain vector through the output layer: ,in is the concentration adjustment amount of tripotassium phosphate, is the fatty acid concentration adjustment amount, To predict the incremental value of the inhibition rate, To predict the incremental value of defense efficiency, To predict the improvement value of the pesticide damage level, when the predicted performance gain value is greater than the preset threshold of 0.1, an optimization suggestion is output.

[0059] The specific implementation of step S09 is to optimize the formula according to the performance gain vector. First, the performance gain vector is analyzed, the adjustment direction and amplitude of the ratio of tripotassium phosphate to fatty acids are calculated, and then the formula is adjusted. The adjustment step of tripotassium phosphate is 0.5%, and the adjustment step of fatty acids is 0.2%. After each adjustment, experimental verification is performed. According to the comprehensive performance calculation equation: Evaluate the optimization effect. When the increment of the antibacterial rate is greater than 5%, the increment of the antibacterial efficiency is greater than 5%, and the phytotoxicity level decreases by more than 0.5 levels, the optimization is considered effective. If the performance gain does not meet expectations, return to step S08 to recalculate the optimization plan. When the change in the performance gain vector is less than 0.05 after three consecutive optimizations, stop the optimization process and output the final formula. Optionally, use the least squares method to verify the optimized formula and calculate the residual sum of squares between the predicted value and the actual value. The residual sum of squares calculation formula is: ,in is the residual sum of squares, is the sample size, is the actual value, is the predicted value. When the residual sum of squares is less than 0.1, the optimization result is confirmed to be reliable.

[0060] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below, which is described in detail as follows.

[0061] 1. Research background and objectives The researchers conducted a systematic study on the optimization of the formula of tobacco powdery mildew control agents. Powdery mildew is one of the main diseases in tobacco cultivation. Traditional control methods have problems such as unstable efficacy and easy to develop drug resistance. This study aims to optimize the ratio of tripotassium phosphate and fatty acids through a deep neural network model to develop an efficient and low-toxic powdery mildew control agent formula. The researchers used a multidimensional performance vector model, an attention mechanism deep learning architecture, and a systematic experimental method to achieve precise optimization of the tobacco powdery mildew control agent formula.

[0062] 2. Materials and Methods 2.1 Experimental Materials Main reagent components: analytical grade tripotassium phosphate ( ), analytically pure oleic acid ( ) Test crops: three-year-old tobacco plants, 60 plants in total Test pathogen: Pure culture of tobacco powdery mildew (Uncinula necator) Main instruments and equipment: Constant temperature incubator (25±0.5℃), digital pH meter (accuracy ±0.01), high performance liquid chromatograph (Shimadzu LC-20AT), vernier caliper (accuracy 0.02mm), electronic balance (accuracy 0.001g), supercomputing workstation (128-core CPU, 4×NVIDIA Tesla V100 GPU) 2.2 Experimental methods 2.2.1 Performance Vector Construction The researchers first constructed a performance vector for powdery mildew control agents, used principal component analysis to determine key performance indicators, and established a multidimensional evaluation matrix of antibacterial effect indicators, control effect indicators, and pesticide damage indicators. The performance vector expression is as follows: ; in: : Tripotassium phosphate concentration value (%); : Fatty acid concentration value (%); : Ambient temperature value (℃); : Relative humidity value (%); : pH value of the agent; : Crop type number; : Disease severity value; : Number of times of application; : Crop growth period value; : Leaf area value (cm²); : Drug contact time value (min); ~ : Weight coefficient of each parameter.

[0063] The weight coefficient of each indicator is determined by the hierarchical analysis method, as shown in Table 1: Table 1 Weight coefficients of performance indicators of powdery mildew control agents

[0064] 2.2.2 Experimental data collection The researchers designed an orthogonal test scheme and collected experimental data for different ratios of tripotassium phosphate and oleic acid. The range of experimental parameters is shown in Table 2: Table 2 Parameter ranges of powdery mildew control agent formulation test

[0065] According to the above parameter range, 25 groups of drug samples with different proportions were prepared, and the antibacterial rate, anti-efficiency rate and drug damage level data of each group of samples were recorded. Some experimental data are shown in Table 3: Table 3 Partial experimental data of powdery mildew control agent formula

[0066] 2.2.3 Data preprocessing and partitioning After the collected experimental data is cleaned, it is normalized to map all eigenvalues ​​to the interval [0,1]. The normalization formula is: ; in, is the standardized value, is the original value, is the minimum value, is the maximum value.

[0067] The researchers divided the total 256 sets of data into training sets and validation sets in a ratio of 8:2, obtaining 205 sets of training samples and 51 sets of validation samples, and ensuring that each type of sample in the data set was evenly distributed.

[0068] 2.2.4 Deep Neural Network Model Construction The researchers constructed a deep neural network model based on the sparse attention mechanism. The network structure is shown in Table 4: Table 4 Structure of deep neural network model for optimizing powdery mildew control agent formulation

[0069] The attention mechanism of this model is designed as follows: (1) First attention layer: A multi-head attention mechanism is used to extract features and calculate weights for tripotassium phosphate concentration and fatty acid concentration. The attention calculation formula is: ; in, , , denote query, key, and value matrices respectively, is the dimension of the key vector.

[0070] (2) Second attention layer: The self-attention mechanism is used to extract features and perform correlation analysis on environmental parameters (ambient temperature, relative humidity, pH value), and calculate the mutual influence weights between each parameter.

[0071] 2.2.5 Model training process The model training adopts the small batch stochastic gradient descent method. The main training parameters are shown in Table 5: Table 5. Deep neural network model training parameter settings

[0072] The change of loss value during the training process is shown in Table 6: Table 6 Loss value record of model training process

[0073] The model reached the early stopping condition in the 487th round, and the mean square error on the validation set was 0.0086, meeting the training termination condition of less than 0.01.

[0074] 3. Pharmaceutical formulation optimization experiment 3.1 Initial formulation performance test The researchers selected an initial formulation (2% tripotassium phosphate, 1.5% oleic acid) for performance testing. The experimental conditions are shown in Table 7: Table 7 Experimental conditions of initial drug formulation

[0075] The performance test results of the initial formulation are shown in Table 8: Table 8 Performance test results of initial pharmaceutical formulation

[0076] 3.2 Performance Evaluation Equations The researchers used the following set of performance evaluation equations for formulation optimization: (1) Antibacterial activity evaluation equation: ; (2) Prevention effectiveness evaluation equation: ; (3) Phytotoxicity assessment equation: ; (4) Comprehensive performance calculation equation: ; Among them, the equation parameters are obtained through regression analysis, as shown in Table 9: Table 9 Performance evaluation equation group parameter table

[0077] 3.3 Optimization process implementation The researchers input the performance test data of the initial pharmaceutical formula into the deep neural network model for iterative optimization. The optimization process is shown in Table 10: Table 10 Optimization process of powdery mildew control agent formulation

[0078] At the sixth iteration, the performance gain vector change was less than 0.001, meeting the optimization termination condition. The final optimized powdery mildew control agent formula was: tripotassium phosphate 2.78%, oleic acid 1.72%.

[0079] 3.4 Optimized formula verification test To verify the actual effect of the optimized formula, the researchers compared the performance of the initial formula and the optimized formula. The experimental conditions were kept consistent and each group of experiments was repeated 5 times. The results are shown in Table 11: Table 11 Comparative test results of pharmaceutical formulation performance before and after optimization

[0080] The researchers also used the least squares method to verify the optimization formula and calculated the residual sum of squares RSS: ; in, is the sample size, is the actual value, is the predicted value. The calculated RSS=0.068 is less than the preset threshold 0.1, which proves that the optimization result is reliable.

[0081] 3.5 Optimized formula performance test under different environmental conditions In order to verify the stability of the optimized formula under different environmental conditions, the researchers conducted multi-environmental condition tests. The test results are shown in Table 12: Table 12 Performance test results of optimized formula under different environmental conditions

[0082] The test results show that the optimized formula maintains good prevention and control effects and safety under different environmental conditions, and has strong environmental adaptability and stability.

[0083] 4. Field trial verification To further verify the effect of the optimized formula in practical application, the researchers conducted field trials in tobacco gardens in three different regions. Three treatments were set up at each test site: the optimized formula treatment group, the initial formula treatment group and the clear water control group. Each treatment was repeated 4 times, with a randomized block design, a plot area of ​​20m², and an experimental period of 60 days. The results of the field trial are shown in Table 13: Table 13 Optimized formula field test results

[0084] The results of field tests showed that the average pest control efficiency of the optimized formula in actual application reached 95.67%, 2.44 percentage points higher than the initial formula, and no pesticide damage occurred, which confirmed the effectiveness and safety of the optimized formula in actual application.

[0085] It should be noted that the variables involved in the present invention are explained in detail as shown in Table 14 below.

[0086] Table 14 Variable explanation table : Performance vector of powdery mildew control agents

[0087] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for optimizing the formulation of a powdery mildew control agent, characterized in that: The following steps are involved: Construct a powdery mildew control agent performance vector including tripotassium phosphate concentration values ​​and fatty acid concentration values; collect multiple sets of experimental data with different ratios of tripotassium phosphate and fatty acids to establish a performance vector database; collect a training data set; standardize the training data set and divide it into a training sample set and a verification sample set; construct a deep neural network model including an input layer, a first attention layer, a second attention layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer, and use the training sample set to train and generate a powdery mildew control agent formulation optimization model; prepare experimental samples according to the range of tripotassium phosphate and fatty acid ratios and conduct performance tests; input the test data into the powdery mildew control agent formulation optimization model; calculate a performance gain vector including the tripotassium phosphate concentration adjustment amount; optimize and adjust the tripotassium phosphate and fatty acid ratio based on the performance gain vector until the performance gain vector meets the preset threshold requirements.

2. The method for optimizing the formula of a powdery mildew control agent according to claim 1, characterized in that: It also includes collecting multiple groups of experimental data of different ratios of tripotassium phosphate and fatty acids, and establishing a performance vector database, wherein the experimental data include antibacterial rate value, prevention efficiency value and phytotoxicity grade value.

3. The method for optimizing the powdery mildew control agent formulation according to claim 2, characterized in that: It comprises a group of equations for evaluating the performance of an agent, which comprises an equation for evaluating antibacterial activity, an equation for evaluating the protective effect, an equation for evaluating the toxicity of the agent and an equation for calculating the comprehensive performance.

4. The method for optimizing the formula of a powdery mildew control agent according to claim 3, characterized in that: The deep neural network model adopts a deep learning architecture based on a sparse attention mechanism. The first attention layer performs feature extraction and weight calculation on tripotassium phosphate concentration values ​​and fatty acid concentration values ​​based on a multi-head attention mechanism, and the second attention layer uses a self-attention mechanism to perform feature extraction and correlation analysis on environmental parameters.

5. The method for optimizing the formula of a powdery mildew control agent according to claim 4, characterized in that: The performance vector determines key performance indicators through principal component analysis, including antibacterial effect, prevention and control effect and phytotoxicity index, and establishes a multidimensional evaluation system through hierarchical analysis.

6. The method for optimizing the formula of a powdery mildew control agent according to claim 5, characterized in that: The training data set uses a stratified sampling method to extract samples from the experimental database, uses the 3 times standard deviation method to delete outliers, and uses the mean filling method to handle missing values.

7. The method for optimizing the formula of a powdery mildew control agent according to claim 6, characterized in that: The training data set is normalized using the minimum and maximum value standardization method for continuous variables, and the one-hot encoding method is used to encode discrete variables.

8. The method for optimizing the formula of a powdery mildew control agent according to claim 7, characterized in that: The deep neural network model training adopts a small batch stochastic gradient descent method, introduces an early stopping mechanism, and adds a batch normalization layer and a dropout layer after each hidden layer.

9. The method for optimizing the formula of a powdery mildew control agent according to claim 8, characterized in that: The experimental samples were prepared in a ratio range of 1 to 5 parts of tripotassium phosphate and 1 to 2 parts of fatty acids, and the antibacterial activity test, control effect test and phytotoxicity evaluation were carried out on the experimental samples.

10. The method for optimizing the formula of a powdery mildew control agent according to claim 9, characterized in that: The performance gain vector includes a tripotassium phosphate concentration adjustment amount, a fatty acid concentration adjustment amount, a predicted antibacterial rate increment value, a predicted prevention efficiency increment value, and a predicted phytotoxicity grade improvement value.

Images