Formula-based whole-line research and development production management method and system

Abstract

The invention discloses a formula-based whole-line research and development production management method and system, and the method comprises the steps: generating a market vector and a supply chain vector according to market data and supply chain data; inputting the market vector and the supply chain vector into a preset formula generation model, and selecting a matched historical formula vector in a process parameter vector space; determining a balance formula vector of the market vector, the supply chain vector and the historical formula vector by using a Pareto frontier method; according to the balance formula vector, obtaining implementation process parameters and inputting the implementation process parameters into a production execution subsystem; and according to the efficacy vector and a preset target efficacy vector, adjusting the formula generation model, and generating new implementation process parameters. Through data vectorization, AI generation, multi-objective optimization, medical knowledge constraint and closed-loop feedback technologies, the problems of data islands, multi-objective conflicts, security risks, iteration lag and the like in traditional research, development and production are solved.

Description

Technical Field

[0001] The present invention relates to the field of medical data processing, and in particular, to a full-line R & D production management method and system based on a formula. Background Art

[0002] In the R & D and production process of the traditional cosmetics / daily chemical industry, the process parameters of the production environment, market data, and supply chain data usually exist in isolation and are not integrated by a unified technical means: the process parameters are restricted by physical / chemical laws (for example, high temperature may cause the decomposition of active ingredients), but the traditional method cannot optimize them in coordination with market data (such as seasonal demand) and supply chain data (such as raw material inventory).

[0003] In addition, the existing formula generation method cannot be updated in real time according to the actual production effect (such as the efficacy compliance rate), resulting in the accumulation of prediction deviations. If the efficacy does not meet the standard after production, it is necessary to restart the R & D process, increasing the experimental cost and time. Summary of the Invention

[0004] Embodiments of the present invention provide a full-line R & D production management method and system based on a formula, which solve specific problems such as data islands, difficulty in multi-objective balance, and lag in formulation iteration in the R & D and production of the cosmetics / daily chemical industry through technical means such as data integration modeling, multi-objective optimization, and periodic closed-loop feedback adjustment model.

[0005] To achieve the above object, the first aspect of the embodiments of the present application provides a full-line R & D production management method based on a formula, including:

[0006] Obtain the process parameter range from the production execution subsystem to generate a process parameter vector space; the process parameter range includes a temperature range, a pH range, and raw material inventory;

[0007] Generate a market vector and a supply chain vector according to market data and supply chain data;

[0008] Input the market vector and the supply chain vector into a preset formula generation model, and select a matching historical formulation vector in the process parameter vector space; the formula generation model is a conditional generative adversarial network structure;

[0009] Use the Pareto front method to determine the balanced formulation vector of the market vector, the supply chain vector, and the historical formulation vector; the difference vector between the balanced formulation vector and the historical formulation vector is in the process parameter vector space;

[0010] Obtain the implementation process parameters according to the balanced formulation vector and input the implementation process parameters into the production execution subsystem;

[0011] After the production execution subsystem runs for one production cycle, it counts the efficacy of the cosmetics produced during the production cycle and the corresponding efficacy vectors.

[0012] According to the efficacy vector and the preset target efficacy vector, adjust the formula generation model to generate new implementation process parameters.

[0013] In a possible implementation manner of the first aspect, obtaining the process parameter range from the production execution subsystem to generate a process parameter vector space; the process parameter range includes a temperature range, a pH range, and raw material inventory, specifically including:

[0014] Use the chi-square test to screen multiple relevant process parameters whose correlation values with the preset target efficacy vector are greater than the preset correlation threshold.

[0015] Obtain the temperature range, pH range, parameter ranges of multiple said relevant process parameters, and raw material inventory from the process parameter range in the production execution subsystem; the number of said relevant process parameters is equal to the number of dimensions of the raw material inventory minus two.

[0016] Vectorize the temperature range, pH range, parameter ranges of multiple said relevant process parameters, and raw material inventory to obtain a temperature vector, a pH vector, multiple relevant process parameter vectors, and a raw material inventory vector.

[0017] Use PCA to reduce the dimensions of the temperature vector, the pH vector, multiple said relevant process parameter vectors, and the raw material inventory vector, and construct a process parameter vector space according to the remaining vectors after dimensionality reduction.

[0018] In a possible implementation manner of the first aspect, the using PCA to reduce the dimensions of the temperature vector, the pH vector, multiple said relevant process parameter vectors, and the raw material inventory vector, and constructing a process parameter vector space according to the remaining vectors after dimensionality reduction, specifically includes:

[0019] Convert the temperature vector, the pH vector, the relevant process parameter vectors, and the raw material inventory vector into row vectors and merge them into a data matrix, where each row of the data matrix represents the process parameter vector of a process.

[0020] Calculate the eigenvalues and the corresponding eigenvectors according to the covariance matrix of the data matrix, and the eigenvalues are arranged in descending order of magnitude.

[0021] Select the minimum number of components such that the cumulative variance is greater than or equal to the preset variance threshold, and select the same number of eigenvectors as the minimum number of components according to the descending order of magnitude result to form a transformation matrix.

[0022] Perform dimensionality reduction transformation on the data matrix according to the conversion matrix to obtain a process parameter vector matrix, and construct a process parameter vector space according to the process parameter vector matrix.

[0023] In a possible implementation manner of the first aspect, the generating the market vector and the supply chain vector according to the market data and the supply chain data specifically includes:

[0024] Use the window sliding method to extract the historical sales volume sequence in the market data as the first market vector, and use the seasonal index, promotion impact factor, regional bias value, and customer rating in the market data as the respective dimension values of the second market vector;

[0025] Convert the supply chain data into a supply chain vector including dimensions of inventory level, supplier delivery time, logistics cost, and warehouse distribution efficiency.

[0026] In a possible implementation manner of the first aspect, the inputting the market vector and the supply chain vector into a preset formula generation model and selecting a matching historical formula vector in the process parameter vector space specifically includes:

[0027] Write the process parameter vector space as a conditional constraint into the cGAN model;

[0028] Input the market vector and the supply chain vector into the cGAN model to obtain a historical formula vector.

[0029] In a possible implementation manner of the first aspect, the using the Pareto front method to determine the balanced formula vector of the market vector, the supply chain vector, and the historical formula vector; the difference vector between the balanced formula vector and the historical formula vector in the process parameter vector space includes:

[0030] Form an objective function for three-objective optimization according to the market demand matching degree, the supply chain feasible value, and the difference value from the historical formula; the market demand matching degree is the similarity between the market vector and the balanced formula vector; the supply chain feasible value is the weighted sum of the modulus value of the balanced formula vector and the inventory constraint violation degree; the difference value from the historical formula is the Euclidean distance between the balanced formula vector and the historical formula vector;

[0031] In the process parameter vector space, use the decomposition multi-objective evolutionary algorithm to solve the objective function to obtain the balanced formula vector.

[0032] In a possible implementation manner of the first aspect, the obtaining the implementation process parameters according to the balanced formula vector and inputting the implementation process parameters into the production execution subsystem specifically includes:

[0033] Map each dimension in the balance distribution vector to specific process parameters, and push all the process parameters to the production execution subsystem through the REST API.

[0034] In a possible implementation manner of the first aspect, after the production execution subsystem runs a production cycle, it statistically analyzes the efficacy of the produced cosmetics and the corresponding efficacy vectors during the production cycle, specifically including:

[0035] Define the dimensions of the efficacy vector according to the type of cosmetics and market demand; each dimension value is a standardized efficacy index;

[0036] During the production execution subsystem running a production cycle, calculate the average efficacy vector of all batches.

[0037] In a possible implementation manner of the first aspect, adjusting the formula generation model according to the efficacy vector and the preset target efficacy vector specifically includes:

[0038] Obtain the medical knowledge base in real time, and construct or update the medical knowledge graph including multiple triples; the first triple of the medical knowledge graph is the ingredient, user age and efficacy, the second triple is the ingredient, user age and property, and the third triple is the concentration, user age and usage constraint;

[0039] Introduce the medical knowledge embedding layer formed by the medical knowledge graph into the generator of the formula generation model, and introduce the knowledge of the first triple of the medical knowledge graph into the attention mechanism;

[0040] According to the difference between the efficacy vector and the preset target efficacy vector, and in combination with the first triple of the medical knowledge graph, dynamically adjust the generation probability of the corresponding ingredient in the generator;

[0041] In the backpropagation of the formula generation model, according to the second and third triples of the medical knowledge graph, use the sum of the weights of the gradients of the medical safety loss and the efficacy loss gradient as the medical constraint gradient of the formula generation model.

[0042] The second aspect of the embodiments of the present application provides a full-line R & D and production management system based on a formula, including:

[0043] An acquisition module, configured to acquire the process parameter range from the production execution subsystem and generate a process parameter vector space; the process parameter range includes a temperature range, a pH range, and raw material inventory;

[0044] A generation module, configured to generate a market vector and a supply chain vector according to market data and supply chain data;

[0045] A matching module, configured to input the market vector and the supply chain vector into a preset formula generation model, and select a matching historical formula vector within the process parameter vector space; the formula generation model is a conditional generative adversarial network structure;

[0046] A determination module, configured to use the Pareto front method to determine a balanced formula vector of the market vector, the supply chain vector, and the historical formula vector; the difference vector between the balanced formula vector and the historical formula vector is within the process parameter vector space;

[0047] A production module, configured to obtain implementation process parameters according to the balanced formula vector and input the implementation process parameters into a production execution subsystem;

[0048] A statistics module, configured to, after the production execution subsystem runs a production cycle, statistics the efficacy of the cosmetics produced during the production cycle and the corresponding efficacy vector;

[0049] An adjustment module, configured to adjust the formula generation model according to the efficacy vector and a preset target efficacy vector to generate new implementation process parameters.

[0050] Compared with the prior art, the formula-based full-line R & D production management method and system provided in this embodiment unify process parameters, market, and supply chain data into a vector space, and combine cGAN to generate formulas, exceeding the simple data storage function of traditional PLM systems and reducing the time for manual parameter screening; under the constraint of process parameters, while balancing market demand, supply chain costs, and historical differences, solving the problem that it is difficult to balance the three in traditional methods. Dynamically adjust the model through the difference between the efficacy vector and the target vector to form a real-time closed loop of "design → production → feedback".

[0051] In addition, by linking with the production execution subsystem, automatic push of production parameters is realized, and production environment feedback information is obtained from the production execution subsystem to adjust the local model in real time. Brief Description of the Drawings

[0052] Figure 1 is a schematic flowchart of a formula-based full-line R & D production management method provided by an embodiment of the present invention;

[0053] Figure 2 is a schematic structural diagram of a formula-based full-line R & D production management system provided by an embodiment of the present invention. Detailed Embodiment

[0054] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts belong to the scope of protection of the present invention.

[0055] To solve the above problems, please refer to Figure 1 , an embodiment of the present invention provides a full-line R & D production management method based on a formula, including:

[0056] S10. Obtain the process parameter range from the production execution subsystem and generate a process parameter vector space; the process parameter range includes a temperature range, a pH range, and raw material inventory.

[0057] S11. Generate a market vector and a supply chain vector according to market data and supply chain data.

[0058] S12. Input the market vector and the supply chain vector into a preset formula generation model, and select a matching historical formula vector in the process parameter vector space; the formula generation model is a conditional generative adversarial network structure.

[0059] S13. Use the Pareto frontier method to determine the balanced formula vector of the market vector, the supply chain vector, and the historical formula vector; the difference vector between the balanced formula vector and the historical formula vector is in the process parameter vector space.

[0060] S14. Obtain the implementation process parameters according to the balanced formula vector and input the implementation process parameters into the production execution subsystem.

[0061] S15. After the production execution subsystem runs a production cycle, count the efficacy of the cosmetics produced during the production cycle and the corresponding efficacy vector.

[0062] S16. Adjust the formula generation model according to the efficacy vector and a preset target efficacy vector to generate new implementation process parameters.

[0063] S10 is the process of constructing a process parameter vector space. Parameters such as temperature range, pH range, and raw material inventory are extracted from a production execution system (such as MES) to construct a multi-dimensional vector space; S11 unifies scattered market and supply chain data into computable vectors to support subsequent input to the AI model; S12 uses a conditional adversarial generation network (cGAN) to quickly match historical successful formulas under process parameter constraints, reducing manual screening time; uses historical formula vectors (such as past high-efficacy formulas) to reduce R & D risks and avoid repetitive and inefficient experiments; S13 determines the balanced formula vector through Pareto optimization, finding the optimal solution among market demand (such as high efficacy), supply chain cost (low cost), and historical differences (such as avoiding excessive deviation from mature formulas), avoiding the extreme tendencies of "emphasizing efficacy over cost" or "emphasizing cost over quality" in traditional R & D. Through the Pareto front method, select the non-dominated solution with the lowest cost or the highest efficacy to reduce the trial-and-error cost; S14 directly pushes the optimized process parameters to the production execution system (MES), reducing the delay and error of manual transfer, realizing an automated closed-loop of "design-production", and shortening the cycle from R & D to production. S15 converts subjective efficacy (such as "moisturizing", "anti-aging") into a standardized vector for objective analysis and comparison, providing real production data for subsequent model optimization and avoiding relying on results under ideal laboratory conditions. S16 adjusts the AI model parameters according to the actual production results (the gap between the efficacy vector and the target efficacy, and the target efficacy vector is the vector corresponding to the preset cosmetic efficacy index), improving the prediction accuracy of future formulas; furthermore, through medical knowledge graph embedding, new discovered ingredient-efficacy relationships, safety constraints, etc. are converted into model knowledge to avoid the loss of experience.

[0064] Through the above steps, this solution not only solves problems such as data islands, multi-objective conflicts, and safety risks in traditional R & D, but also significantly improves R & D efficiency and product market competitiveness through full-process automation and data-driven methods.

[0065] Exemplarily, obtaining the process parameter range from the production execution subsystem to generate a process parameter vector space; the process parameter range includes a temperature range, a pH range, and raw material inventory, specifically including:

[0066] Using the chi-square test to screen multiple relevant process parameters whose correlation values with the preset target efficacy vector are greater than the preset correlation threshold;

[0067] Obtaining the temperature range, pH range, parameter ranges of multiple said relevant process parameters, and raw material inventory from the process parameter range obtained from the production execution subsystem; the number of said relevant process parameters is equal to the number of raw material inventory dimensions minus two;

[0068] Vectorize the temperature range, pH range, parameter ranges of multiple said relevant process parameters, and raw material inventory to obtain a temperature vector, a pH vector, multiple relevant process parameter vectors, and a raw material inventory vector;

[0069] Use PCA to reduce the dimensions of the temperature vector, the pH vector, multiple said relevant process parameter vectors, and the raw material inventory vector, and construct a process parameter vector space based on the remaining vectors after dimensionality reduction.

[0070] First, preset cosmetic efficacy indicators (such as moisturizing rate, SPF value, etc.). Then, historical production records that can be extracted from the Manufacturing Execution System (MES), including parameters such as temperature, pH, raw material concentration, stirring speed, etc. The correlation between each process parameter and the target efficacy vector is tested through the chi-square test. Process parameters with a p-value less than a preset threshold (such as 0.05) in the chi-square test can be selected, that is, parameters significantly correlated with the target efficacy.

[0071] The above steps are to screen out redundant parameters irrelevant to the target efficacy (for example, the stirring time of some raw materials may have no significant impact on the efficacy), and reduce the dimensions of subsequent calculations. For example: If the "pH value" is significantly correlated with the "anti-aging efficacy" (p = 0.03), it is retained; while if the "stirring speed" has a p = 0.12, it is excluded.

[0072] Then, extract the range of relevant process parameters after screening from the production execution system: Temperature range: such as [25°C, 45°C]. pH range: such as [5.0, 7.0]. Other relevant parameters: such as the concentration of raw material A [0.5%, 2.0%], the inventory of raw material B [100 kg, 500 kg]. Dimension of raw material inventory: Assume that the raw material inventory contains 3 raw materials (number of dimensions = 3), then the number of relevant process parameters is 3 - 2 = 1 (according to the user description).

[0073] Example: If the inventory dimension is 3 (raw materials X, Y, Z), then select 1 parameter strongly correlated with the inventory (such as the concentration of raw material X). Vectorization: Temperature vector: [min_temp, max_temp] → [25, 45]. pH vector: [min_pH, max_pH] → [5.0, 7.0]. Relevant process parameter vector: such as [min and max of the concentration of raw material X] → [0.5%, 2.0%]. Raw material inventory vector: the current inventory of each raw material → [150 kg (X), 300 kg (Y), 400 kg (Z)].

[0074] The above steps are to convert the parameter range into a numerical vector to provide standardized data for subsequent dimensionality reduction and model input. For example: After vectorizing the "concentration range of raw material X", it can be directly involved in the PCA calculation.

[0075] Next, all process parameter vectors are combined into a data matrix, where each row represents the parameter combination of a process step.

[0076] Example: Assume data for 3 process steps:

[0077]

[0078]

[0079] The matrix can be standardized using Z-score, the covariance matrix can be eigen-decomposed, and sorted in descending order of eigenvalues. If the cumulative variance contribution rate ≥ a preset threshold (e.g., 95%), determine the number of principal components to retain. Example: If the cumulative variance of the first 2 principal components is 96%, then retain 2 components. This is to project the original data into the low-dimensional space formed by the principal components, creating a process parameter vector space. If reducing from the original 10 dimensions (assuming 10 parameters including temperature, pH, concentration, inventory) to 3 dimensions, it reduces the computational complexity and ensures that the vectors after dimensionality reduction can still explain most of the variance of the process parameters.

[0080] In summary, through data screening (chi-square test) → vectorization → dimensionality reduction (PCA), the original high-dimensional and complex process parameters are transformed into a low-dimensional and interpretable vector space, providing an efficient data basis for subsequent formulation generation and optimization. Its core value lies in: eliminating irrelevant parameters, focusing on key process factors; unifying the data format to support AI model input; reducing computational complexity while retaining the main information. R & D personnel can quickly locate the "effective process parameter combinations", significantly improving the efficiency and accuracy of formulation generation.

[0081] Exemplarily, the PCA is used to reduce the dimensionality of the temperature vector, the pH vector, multiple relevant process parameter vectors, and the raw material inventory vector, and a process parameter vector space is constructed based on the remaining vectors after dimensionality reduction. Specifically, it includes:

[0082] Convert the temperature vector, the pH vector, the relevant process parameter vectors, and the raw material inventory vector into row vectors and combine them into a data matrix, where each row of the data matrix represents the process parameter vector of a process step;

[0083] Calculate the eigenvalues and corresponding eigenvectors based on the covariance matrix of the data matrix, and the eigenvalues are sorted in descending order of magnitude;

[0084] Select the minimum number of components such that the cumulative variance is greater than or equal to the preset variance threshold, and select the same number of eigenvectors as the minimum number of components according to the descending order result of the magnitude values to form a transformation matrix;

[0085] Perform dimensionality reduction transformation on the data matrix according to the transformation matrix to obtain a process parameter vector matrix, and construct a process parameter vector space according to the process parameter vector matrix.

[0086] Select the smallest number of components k such that the cumulative variance ≥ a preset threshold (e.g., 95%). Select the first k eigenvectors (corresponding to the largest eigenvalues) to form the transformation matrix W, whose dimension is the number of features × k. Usually, a cumulative variance ≥ 95% is selected to balance information retention and dimensionality compression. For example, if the original data has 4 features and the cumulative variance of the first 2 principal components is 96%, then k = 2. The columns of the transformation matrix W are the principal component directions, which are used to project the original data into a low-dimensional space.

[0087] Project the standardized data matrix X′ onto the transformation matrix W to obtain the dimensionality-reduced process parameter vector matrix. Formula: Z = X′W. The projected data Z retains the main variances of the original data in the new coordinate system, and the principal components are orthogonal (no redundancy). The vector space is used as the input for the subsequent formulation generation model (such as the cGAN in step S12) to ensure that the generated formulation meets the process feasibility constraints. The principal components after dimensionality reduction are linear combinations of the original parameters, and the contribution of each parameter to the principal components can be analyzed through the eigenvector weights (e.g., a high temperature weight indicates that it has a significant impact on the process).

[0088] It should be noted that the original data must be centered and standardized, otherwise the covariance matrix may be dominated by high-dimensional parameters. The cumulative variance threshold needs to be adjusted according to actual needs (e.g., for the R & D of cosmetics with high precision requirements, it can be set above 95%). If the correlation between features is weak, more components may need to be retained to avoid information loss.

[0089] In summary, through PCA, the high-dimensional process parameter vectors (such as temperature, pH, inventory) are reduced to a low-dimensional space, and its core value lies in: eliminating redundant information between parameters (such as strongly correlated temperature and pH); reducing the input dimension of the subsequent model (such as cGAN) and accelerating the calculation; ensuring that the variances of key process parameters are retained through cumulative variance control.

[0090] Exemplarily, generating a market vector and a supply chain vector according to market data and supply chain data specifically includes:

[0091] Use the window sliding method to extract the historical sales volume sequence in the market data as the first market vector, and use the seasonal index, promotion impact factor, regional bias value, and customer rating in the market data as the respective dimension values of the second market vector;

[0092] Convert the supply chain data into a supply chain vector including dimensions of inventory level, supplier delivery time, logistics cost, and warehouse distribution efficiency.

[0093] The window sliding method captures short-term trends and seasonal fluctuations by moving a fixed-length window (such as 7 days, 30 days) over the time series and calculating the average or total sales volume within the window. For each date t, the average sales volume over the past 30 days can be calculated, and the first market vector = {sales volume series t} reflects the recent sales trend.

[0094] Time series decomposition (such as STL decomposition) can be used to separate the trend, seasonal, and residual components and extract seasonal coefficients.

[0095] Seasonal index = average annual sales volume / average monthly sales volume;

[0096] Compare the sales volume differences between the promotion period and the non-promotion period and calculate the lift rate:

[0097] Promotion factor = (sales volume during the promotion period - sales volume during the benchmark period) / (sales volume during the benchmark period);

[0098] Calculate the degree to which the sales volume proportion in different regions deviates from the national average:

[0099] Regional deviation i = (sales volume proportion of region i - national average proportion) / (national standard deviation);

[0100] Standardize the scoring data (such as 0-1 normalization).

[0101] Example: The second market vector of a certain product is: Second market vector = [1.2 (seasonal), 0.3 (promotion), 0.8 (region), 0.9 (score)].

[0102] Then the supply chain data can be mapped to four dimensions to form a standardized vector.

[0103] Inventory level dimension: Inventory level = safety inventory / current inventory × 100%;

[0104] Supplier delivery time dimension: Statistically calculate the average or quantile (such as 95% delivery time) of the supplier's historical delivery time: Delivery time = average delivery days in the past 30 days;

[0105] Logistics cost dimension: Calculate the unit logistics cost (such as the transportation cost per item):

[0106] Logistics cost = total logistics cost / total sales volume;

[0107] Warehouse distribution efficiency dimension: Comprehensively score through warehouse utilization rate and distribution coverage rate: Efficiency = 0.5 × utilization rate + 0.5 × coverage rate;

[0108] In the above embodiments, by converting complex business data into structured vectors, it provides standardized inputs for subsequent AI models (such as cGAN), supporting automated formula generation and resource allocation decisions.

[0109] Exemplarily, inputting the market vector and the supply chain vector into a preset formula generation model, and selecting a matching historical formula vector in the process parameter vector space specifically includes:

[0110] Writing the process parameter vector space as a conditional constraint into the cGAN model;

[0111] Inputting the market vector and the supply chain vector into the cGAN model to obtain a historical formula vector.

[0112] Define the boundaries of the process parameter vector space (such as temperature range, pH value range) as the process parameter vector cprocess. Introduce a conditional layer in the generator and use the process parameter vector cprocess as an additional input.

[0113] It is also possible to add a linear layer before the output layer of the generator to project the generated formula vector into the process parameter feasible region (process parameter vector space). For example: restrict the generated pH value to the range of 5 - 7, which can be achieved through the Clip function or a constraint loss function.

[0114] Exemplarily, using the Pareto front method to determine the balanced formula vector of the market vector, the supply chain vector, and the historical formula vector; the difference vector between the balanced formula vector and the historical formula vector in the process parameter vector space includes:

[0115] Form an objective function for three - objective optimization based on the market demand matching degree, the supply chain feasible value, and the difference value from the historical formula; the market demand matching degree is the similarity between the market vector and the balanced formula vector; the supply chain feasible value is the weighted sum of the modulus of the balanced formula vector and the inventory constraint violation degree; the difference value from the historical formula is the Euclidean distance between the balanced formula vector and the historical formula vector;

[0116] In the process parameter vector space, use the decomposed multi - objective evolutionary algorithm to solve the objective function to obtain the balanced formula vector.

[0117] Take the development of a moisturizing cream formula as an example:

[0118] Input conditions:

[0119] Market vector: User preference for "natural ingredients" (such as hyaluronic acid, glycerin).

[0120] Supply chain vector: Sufficient inventory (glycerol inventory = 500 kg), cost constraint (total cost ≤ 10 yuan / bottle).

[0121] Historical formulation vector: A certain classic formulation (hyaluronic acid 5%, glycerol 10%, pH = 5.5).

[0122] Objective function calculation:

[0123] Formulation A (generated formulation vector): Hyaluronic acid 6%, glycerol 8%, pH = 5.0

[0124] f1 = 0.9 * (High market matching degree)

[0125] f2 = 0.7·14% + 0.3·0 = 0.98 (No inventory violation)

[0126]

[0127] Pareto front solution:

[0128] Solution 1: High market matching (f1 = 0.9), Medium supply chain feasibility (f2 = 1.0), Large difference (f3 = 2.5).

[0129] Solution 2: Low market matching (f1 = 0.7), Optimal supply chain (f2 = 0.8), Small difference (f3 = 0.5).

[0130] Select the balanced solution:

[0131] Select Solution 1 or Solution 2 according to business priorities, or generate a new solution through weight adjustment.

[0132] Through the Pareto front method and the MOEA / D algorithm, this solution can find the optimal solution set among market demand, supply chain feasibility, and historical formulation differences; prioritize optimizing key components and constraints to improve resource utilization efficiency; ensure that the generated formulation meets production requirements through hard constraints or penalty functions.

[0133] Exemplarily, obtaining the implementation process parameters according to the balanced formulation vector and inputting the implementation process parameters into the production execution subsystem specifically includes:

[0134] Map each dimension in the balanced formulation vector to specific process parameters, and push all process parameters to the production execution subsystem through the REST API.

[0135] The optimized balanced formulation vector can be converted into specific process parameters and pushed to the production execution subsystem (such as the MES system) through the REST API.

[0136] Rule design needs to be carried out in advance to define parameter names, physical ranges, and conversion formulas for each dimension of the balance allocation vector. Then, the actual parameter values are calculated dimension by dimension according to the mapping table (or other formats of rule design). The balance allocation vector is [0.7, 0.3, 0.6, 0.8]:

[0137] Temperature: 25 + 20×0.7 = 39°C;

[0138] pH value: 5.0 + 2.0×0.3 = 5.6;

[0139] Raw material concentration: 0.5 + 1.5×0.6 = 1.4%;

[0140] Stirring time: 5 + 10×0.8 = 13 minutes;

[0141] Then, the parameters are passed to the manufacturing execution subsystem (MES) through the standard HTTP protocol (such as REST API) to achieve data interaction between systems.

[0142] The above process ensures the accuracy and reliability of parameter push in the following ways: decoding the abstract vector into specific physical parameters to avoid ambiguity; ensuring transmission security and stability through authentication, encryption, and retry mechanisms. Parameter verification and confirmation by the MES system ensure compliance of manufacturing execution.

[0143] Through this step, the formulated parameters optimized in research and development can be quickly and securely transmitted to the production side, achieving seamless connection from the laboratory to the production line, and significantly improving the automation level and efficiency of cosmetics production.

[0144] Exemplarily, after the manufacturing execution subsystem runs a production cycle, the efficacy of the cosmetics produced during the production cycle and the corresponding efficacy vectors are statistically analyzed, specifically including:

[0145] Define the dimensions of the efficacy vector according to the type of cosmetics and market demand; each dimension value is a standardized efficacy index;

[0146] During the operation of the manufacturing execution subsystem for a production cycle, calculate the average efficacy vector of all batches.

[0147] In this embodiment, the dimensions of the efficacy vector can be defined according to the type of cosmetics and market demand, and the detection methods and indicators for each dimension can be determined. Adjust the weights according to market demand or product positioning. For example, the weight of "L value" may be higher than that of "the area of age spots" in whitening products.

[0148] Example: Moisturizing cream: The core efficacy is "moisturizing", and it is necessary to pay attention to water content, trans-epidermal water loss (TEWL), skin barrier repair, etc. Anti-aging essence: The core efficacy is "anti-wrinkle" and "elasticity improvement", and it is necessary to pay attention to wrinkle depth, skin elasticity, collagen content, etc. Whitening lotion: The core efficacy is "skin brightening", and it is necessary to pay attention to the area of skin pigmentation, Lab value (especially L value and a value), melanin content, etc.

[0149] Through the above steps, the production execution subsystem can systematically convert efficacy data into structured vectors, support closed-loop optimization from R & D to production, and improve the quality of cosmetics.

[0150] Exemplarily, adjusting the formula generation model according to the efficacy vector and the preset target efficacy vector specifically includes:

[0151] Obtain the medical knowledge base in real time and construct or update a medical knowledge graph including multiple triples; the first triple of the medical knowledge graph is ingredient, user age and efficacy, the second triple is ingredient, user age and property, and the third triple is concentration, user age and usage constraint;

[0152] Introduce a medical knowledge embedding layer formed by the medical knowledge graph into the generator of the formula generation model, and introduce the knowledge of the first triple of the medical knowledge graph into the attention mechanism;

[0153] According to the difference between the efficacy vector and the preset target efficacy vector, and in combination with the first triple of the medical knowledge graph, dynamically adjust the generation probability of the corresponding ingredient in the generator;

[0154] In the backpropagation of the formula generation model, according to the second triple and the third triple of the medical knowledge graph, use the weighted sum of the gradients of the medical safety loss and the efficacy loss gradient as the medical constraint gradient of the formula generation model.

[0155] By obtaining the medical knowledge base in real time (such as drug instructions, clinical trial data, medical literature), construct a medical knowledge graph containing triples to guide the constraint and optimization of the formula generation model.

[0156] Definition of triples:

[0157] The first triple (ingredient - age - efficacy): Represents the efficacy impact of a specific ingredient on users of different ages.

[0158] Example: (Hyaluronic acid, 25 - 30 years old, enhanced moisturizing), (Retinol, 30 - 40 years old, anti-wrinkle and elasticity improvement)

[0159] The second triple (ingredient - age - property): Represents the change of the impact of an ingredient on the properties of the product (such as texture, stability) with age.

[0160] Examples: (mineral oil, children, prone to pore clogging), (ceramide, sensitive skin, enhance the barrier)

[0161] The third tuple (concentration - age - constraint): Defines the safe usage limits of ingredients at different concentrations in specific age groups.

[0162] Examples: (salicylic acid concentration > 2%, under 12 years old, prohibited), (niacinamide concentration > 5%, sensitive skin, need to be used with a moisturizer)

[0163] Introduce a medical knowledge embedding layer in the generator network, encode the semantic information of the knowledge graph into the model parameters, and strengthen the parameter values of key ingredients in the model through the attention mechanism.

[0164] Construction of the knowledge embedding layer: Use knowledge graph embedding models such as TransE or DistMult to convert the tuples into low - dimensional vectors. For example, concatenate the ingredient embedding and the age embedding, and map them to a knowledge vector through a fully - connected layer: Knowledge vector i＝ FC([ingredient i , age i ), where j is a natural number.

[0165] Enhancement of the attention mechanism: In the self - attention layer of the Transformer generator, use the knowledge vector as an additional Query or Key. This can enhance the generation probability of efficacy ingredients (such as ingredients that are highly effective for the target age group).

[0166] Then, according to the difference between the current efficacy vector and the target vector, dynamically adjust the generation probability of ingredients in the generator, and preferentially select ingredients that meet the constraints of the knowledge graph. Select the top k dimensions with the largest differences (such as anti - wrinkle, moisturizing). Based on the rules of the first tuple: For the efficacy corresponding to the different dimensions (such as anti - wrinkle), query the high - efficacy ingredients in the knowledge graph that match the target age group. Example: If the target age group is 30 - 40 years old and the anti - wrinkle difference is large, then preferentially select ingredients such as retinol and peptides.

[0167] Probability correction formula: P(ingredient i ) ∝ exp(β · efficacy relevance i ), where efficacy relevance i is the query result of the first tuple from the knowledge graph, β is the adjustment coefficient, and i is a natural number.

[0168] Finally, in the backpropagation, combine the constraints of the second and third tuples of the knowledge graph to design the gradient of the loss function including safety.

[0169] Efficacy loss L eff = the absolute value between the target efficacy vector and the predicted efficacy vector;

[0170] Safety loss γ is a penalty coefficient, which is triggered when the component concentration or property violates the knowledge graph constraint.

[0171] Weighted sum of medical constraint gradients α is a weight parameter (e.g., α = 0.8, giving priority to optimizing efficacy).

[0172] The above method ensures that the generated formula meets clinical evidence and regulatory requirements (such as high-concentration salicylic acid being prohibited for children) through a medical knowledge graph; adjusts the component probability according to real-time efficacy feedback to accelerate formula iteration; and avoids safety risks while enhancing efficacy through gradient weighting.

[0173] In summary, the formula generation model can combine medical knowledge with actual efficacy data to generate personalized formulas that are both effective and safe, significantly improving R & D efficiency and compliance.

[0174] Compared with the prior art, the full-line R & D and production management method based on formula provided in this embodiment unifies process parameters, market, and supply chain data into a vector space, and combines cGAN to generate formulas, exceeding the simple data storage function of traditional PLM systems and reducing the time for manual parameter screening; under the constraint of process parameters, it balances market demand, supply chain costs, and historical differences at the same time, solving the problem that it is difficult to balance the three in traditional methods. Dynamically adjust the model through the difference between the efficacy vector and the target vector to form a real-time closed loop of "design → production → feedback".

[0175] In addition, it also realizes the automatic push of production parameters by linking with the production execution subsystem, and obtains production environment feedback information from the production execution subsystem to adjust the local model in real time.

[0176] See Figure 2 , an embodiment of the present application provides a full-line R & D and production management system based on formula, including: an acquisition module 20, a generation module 21, a matching module 22, a determination module 23, a production module 24, a statistics module 25, and an adjustment module 26.

[0177] The acquisition module 20 is used to obtain the process parameter range from the production execution subsystem and generate a process parameter vector space; the process parameter range includes temperature range, pH range, and raw material inventory.

[0178] The generation module 21 is used to generate a market vector and a supply chain vector according to market data and supply chain data;

[0179] The matching module 22 is used to input the market vector and the supply chain vector into a preset formula generation model, and select a matching historical formula vector within the process parameter vector space; the formula generation model is a conditional generative adversarial network structure.

[0180] A determination module 23, configured to determine a balanced allocation vector of the market vector, the supply chain vector, and the historical allocation vector by using the Pareto front method; a difference vector between the balanced allocation vector and the historical allocation vector is in the process parameter vector space.

[0181] A production module 24, configured to obtain implementation process parameters according to the balanced allocation vector and input the implementation process parameters into a production execution subsystem.

[0182] A statistics module 25, configured to, after the production execution subsystem runs a production cycle, count the efficacy of the cosmetics produced during the production cycle and the corresponding efficacy vector.

[0183] An adjustment module 26, configured to adjust the formula generation model according to the efficacy vector and a preset target efficacy vector to generate new implementation process parameters.

[0184] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working process of the above-described full-line R & D production management system based on formula can refer to the corresponding process in the foregoing method embodiment, and will not be elaborated herein.

[0185] Compared with the prior art, the full-line R & D production management system based on formula provided in this embodiment unifies process parameters, market, and supply chain data into a vector space, combines cGAN to generate a formula, transcends the simple data storage function of traditional PLM systems, and reduces the time for manual parameter screening; under the constraint of process parameters, it balances market demand, supply chain cost, and historical differences at the same time, and solves the problem that it is difficult to balance the three in traditional methods. By dynamically adjusting the model according to the difference between the efficacy vector and the target vector, a real-time closed loop of "design → production → feedback" is formed.

[0186] In addition, by linking with the production execution subsystem, automatic push of production parameters is realized, and production environment feedback information is obtained from the production execution subsystem to adjust the local model in real time.

[0187] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the full-line R & D production management method based on formula as described above.

[0188] The computer device may be a computing device such as a smart phone, a tablet computer, a desktop computer, and a cloud server. The computer device may include but is not limited to a processor and a memory. Those skilled in the art can understand that the figure is only an example of the computer device, and does not constitute a limitation on the computer device. It may include more or fewer components than shown in the figure, or combine some components, or different components. For example, it may also include input and output devices, network access devices, etc.

[0189] The so-called processor may be a Central Processing Unit (CPU), and the processor may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0190] In some embodiments, the memory may be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. equipped on the computer device. Further, the memory may also include both the internal storage unit and the external storage device of the computer device. The memory is used to store an operating system, application programs, a BootLoader, data, and other programs, such as the program code of the computer program. The memory may also be used to temporarily store data that has been output or is to be output.

[0191] The embodiments of the present application provide a computer program product. When the computer program product runs on a computer device, it causes the computer device to execute the steps in the above-mentioned method embodiments.

[0192] In several embodiments provided by the present application, it can be understood that each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the module, the program segment, or the part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved.

[0193] When the above-mentioned functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art or a part of this technical solution can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device to execute all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-On l y Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0194] The above is the preferred implementation manner of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A formula-based full-line R&D and production management method, characterized in that: include: Obtain the process parameter range from the production execution subsystem and generate the process parameter vector space; The process parameter ranges include temperature range, pH range and raw material inventory; Generate market vectors and supply chain vectors based on market data and supply chain data; Inputting the market vector and the supply chain vector into a preset recipe generation model, and selecting a matching historical recipe vector in the process parameter vector space; the recipe generation model is a conditional generative adversarial network structure; Using the Pareto frontier method, determining a balanced recipe vector of the market vector, the supply chain vector and the historical recipe vector; a difference vector between the balanced recipe vector and the historical recipe vector is in the process parameter vector space; According to the balanced recipe vector, an implementation process parameter is obtained and the implementation process parameter is input into a production execution subsystem; After the production execution subsystem runs a production cycle, the efficacy of the cosmetics produced in the production cycle and the corresponding efficacy vector are counted; According to the efficacy vector and a preset target efficacy vector, the recipe generation model is adjusted to generate new implementation process parameters.

2. A formula-based full-line R&D and production management method as claimed in claim 1, characterized in that: The process parameter range is obtained from the production execution subsystem to generate a process parameter vector space; The process parameter range includes temperature range, pH range and raw material inventory, specifically including: Using chi-square test to screen multiple relevant process parameters whose correlation values ​​with preset target efficacy vectors are greater than preset correlation thresholds; Obtaining a temperature range, a pH range, a plurality of parameter ranges of the related process parameters and a raw material inventory from the process parameter range obtained in the production execution subsystem; the number of the related process parameters is equal to the number of dimensions of the raw material inventory minus two; Quantizing the temperature range, the pH range, the parameter ranges of the plurality of related process parameters, and the raw material inventory vector to obtain a temperature vector, a pH vector, a plurality of related process parameter vectors, and a raw material inventory vector; PCA is used to reduce the dimension of the temperature vector, the pH vector, a plurality of the related process parameter vectors and the raw material inventory vector, and a process parameter vector space is constructed according to the remaining vectors after the dimension reduction.

3. A formula-based full-line R&D and production management method as claimed in claim 2, characterized in that: The PCA is used to reduce the dimension of the temperature vector, the pH vector, the plurality of related process parameter vectors and the raw material inventory vector, and the process parameter vector space is constructed according to the remaining vectors after the dimension reduction, specifically including: Convert the temperature vector, the pH vector, the related process parameter vector and the raw material inventory vector into row vectors and merge them into a data matrix, wherein each row of the data matrix represents a process parameter vector of a process; Calculate eigenvalues ​​and corresponding eigenvectors according to the covariance matrix of the data matrix, and arrange the eigenvalues ​​in descending order of magnitude; Select the minimum number of components that makes the cumulative variance greater than or equal to a preset variance threshold, and select the eigenvectors equal to the minimum number of components according to the descending order of the size values ​​to form a transformation matrix; The data matrix is ​​subjected to dimensionality reduction transformation according to the transformation matrix to obtain a process parameter vector matrix, and a process parameter vector space is constructed according to the process parameter vector matrix.

4. A formula-based full-line R&D and production management method as claimed in claim 1, characterized in that: The generating of the market vector and the supply chain vector according to the market data and the supply chain data specifically includes: A window sliding method is used to extract the historical sales volume sequence quantity in the market data as the first market vector, and the seasonal index, promotion impact factor, regional bias value and customer score in the market data are respectively used as the dimensional values ​​of the second market vector; The supply chain data is transformed into a supply chain vector including inventory level dimensions, supplier delivery time dimensions, logistics cost dimensions, and warehouse distribution efficiency dimensions.

5. A formula-based full-line R&D and production management method as claimed in claim 1, characterized in that: The step of inputting the market vector and the supply chain vector into a preset recipe generation model and selecting a matching historical recipe vector in the process parameter vector space specifically includes: Writing the process parameter vector space into the cGAN model as a conditional constraint; The market vector and the supply chain vector are input into the cGAN model to obtain a historical recipe vector.

6. A formula-based full-line R&D and production management method as claimed in claim 1, characterized in that: The Pareto frontier method is used to determine the equilibrium recipe vector of the market vector, the supply chain vector and the historical recipe vector; The difference vector between the balanced recipe vector and the historical recipe vector in the process parameter vector space includes: The objective function of three-objective optimization is formed based on the market demand matching degree, supply chain feasibility value and the difference value with the historical formula; The market demand matching degree is the similarity between the market vector and the equilibrium formula vector; the supply chain feasible value is the weighted sum of the equilibrium formula vector modulus and the inventory constraint violation degree; the difference value from the historical formula is the Euclidean distance between the equilibrium formula vector and the historical formula vector; In the process parameter vector space, a decomposition multi-objective evolutionary algorithm is used to solve the objective function to obtain a balanced recipe vector.

7. A formula-based full-line R&D and production management method as claimed in claim 1, characterized in that: The step of obtaining implementation process parameters according to the balanced recipe vector and inputting the implementation process parameters into the production execution subsystem specifically includes: Each dimension in the balanced recipe vector is mapped to a specific process parameter, and all process parameters are pushed to the production execution subsystem through the REST API.

8. A formula-based full-line R&D and production management method as claimed in claim 1, characterized in that: After the production execution subsystem runs a production cycle, the efficacy of the cosmetics produced in the production cycle and the corresponding efficacy vector are counted, specifically including: Define the dimensions of the efficacy vector according to the type of cosmetics and market demand; each dimension value is a standardized efficacy index; The production execution subsystem calculates the average efficacy vector of all batches during one production cycle.

9. A formula-based full-line R&D and production management method as claimed in claim 1, characterized in that: The adjusting the recipe generation model according to the efficacy vector and the preset target efficacy vector specifically includes: Acquire a medical knowledge base in real time, and construct or update a medical knowledge graph including a plurality of triples; the first tuple of the medical knowledge graph is ingredient, user age and efficacy, the second tuple is ingredient, user age and property, and the third tuple is concentration, user age and usage constraint; Introducing a medical knowledge embedding layer formed by the medical knowledge graph into the generator of the recipe generation model, and introducing the first tuple knowledge of the medical knowledge graph into the attention mechanism; According to the difference between the efficacy vector and the preset target efficacy vector, combined with the first tuple of the medical knowledge graph, dynamically adjusting the generation probability of the corresponding component in the generator; In the back propagation of the recipe generation model, according to the second tuple and the third tuple of the medical knowledge graph, the weighted sum of the gradient of the medical safety loss and the gradient of the efficacy loss is used as the medical constraint gradient of the recipe generation model.

10. A full-line R&D and production management system based on formula, characterized in that: include: An acquisition module is used to obtain the process parameter range from the production execution subsystem and generate a process parameter vector space; The process parameter ranges include temperature range, pH range and raw material inventory; A generation module, used for generating a market vector and a supply chain vector according to the market data and the supply chain data; A matching module, used for inputting the market vector and the supply chain vector into a preset recipe generation model, and selecting a matching historical recipe vector in the process parameter vector space; the recipe generation model is a conditional generative adversarial network structure; A determination module, configured to determine a balanced recipe vector of the market vector, the supply chain vector and the historical recipe vector using a Pareto frontier method; a difference vector between the balanced recipe vector and the historical recipe vector is in the process parameter vector space; A production module, used for obtaining implementation process parameters according to the balanced recipe vector and inputting the implementation process parameters into a production execution subsystem; A statistical module, used for counting the efficacy and corresponding efficacy vector of the cosmetics produced in a production cycle after the production execution subsystem runs a production cycle; The adjustment module is used to adjust the recipe generation model according to the efficacy vector and a preset target efficacy vector to generate new implementation process parameters.

Images