Interactive solid propellant performance prediction platform and method based on machine learning
Through the interactive solid propellant performance prediction platform, a matrix for machine learning modeling is automatically generated, solving the complexity and user-friendliness of solid propellant performance prediction in the prior art, and achieving a simple and efficient performance prediction process .
Patent Information
- Application Number
- CN202410375448.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-29
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-03-29
AI Technical Summary
In the prior art, the use of machine learning to predict solid propellant performance requires multiple attempts in data set construction, method parameter search and cross-validation, and the conversion of solid propellant formula into a data set that can be modeled by machine learning exists. Multiple processes are not friendly to users without a foundation in machine learning and program processing and cannot meet the diverse needs of users.
Provide an interactive solid propellant performance prediction platform based on machine learning, including user input and result display modules, model training and prediction modules. Users can select machine learning methods and parameter ranges, automatically generate a matrix for machine learning modeling, and implement it. Non-programmed performance prediction, and supports interrupts and optimal parameter preservation of training processes, automatically completes the combination of multi-parameters and cross-validation, providing visualization of modeling processes.
Users can perform machine learning modeling and performance prediction through simple choices, eliminating the data conversion process, and automatically completing the time-consuming and labor-intensive multi-parameter and cross-verification combination, improving user experience and processing efficiency and reducing the threshold for use.
Smart Images

Figure CN118155752B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data mining and energetic material technology, in particular to the field of energetic material solid propellant formulation screening and prediction technology, specifically, to an interactive solid propellant performance prediction platform and method based on machine learning. Background Art
[0002] With the popularization of machine learning methods and empirical formulas as a theoretical basis, it has gradually become possible to predict the burning rate, safety and other properties of solid propellants using only solid propellant formulas, provided that there is sufficient data. However, despite the large number of machine learning methods, multiple attempts are required in terms of data set construction, method parameter search, cross-validation, etc., and there are also several steps to convert the solid propellant formula into a data set that can be used for machine learning modeling. In particular, the two steps of parameterizing the solid propellant formula and the experimental environment are very unfriendly to users with a chemical background but no machine learning and program processing foundation. In order to meet the diverse needs of users and improve user experience, it is necessary to provide an interactive platform to achieve non-programmed solid propellant performance prediction. However, such an interactive platform does not currently exist. Summary of the invention
[0003] The purpose of the present invention is to provide an interactive solid propellant performance prediction platform and method based on machine learning, which is used to solve the problem that the prior art uses machine learning to predict the performance of solid propellants, which requires multiple attempts in data set construction, method parameter search, cross-validation, etc., and there are also several steps in converting the solid propellant formula into a data set that can be used for machine learning modeling. In particular, the two steps of parameterizing the solid propellant formula and the experimental environment are very unfriendly to users with a chemical background but no machine learning and program processing foundation, and cannot meet the diverse needs of users.
[0004] The present invention solves the above problems through the following technical solutions:
[0005] An interactive solid propellant performance prediction platform based on machine learning, including:
[0006] The user input and result display module is used for users to select machine learning methods and parameter ranges and send them to the model training and prediction module; it is used to send the user's start training, interrupt parameter search instructions and save optimal parameter instructions to the model training and prediction module; it is used for users to select recipes, trained models and experimental parameters and send them to the model training and prediction module, and it is also used to visualize the feedback results of the model training and prediction module;
[0007] The model training and prediction module is used to retrieve the corresponding data in the database according to the machine learning method and parameter range selected by the user, form a recipe information matrix after preprocessing, and feed back the matrix size to the user input and result display module for display; and after receiving the user's start training instruction, it searches for parameters and builds models, and after receiving the user's interrupt instruction, stops calculating and feeds back the current optimal parameters; after receiving the user's save optimal parameter instruction, saves the current optimal parameters as the model parameters; and predicts the recipe performance according to the user's selected recipe, the trained model and the experimental parameters, and feeds back to the user input and result display module.
[0008] Furthermore, the user input and result display module includes:
[0009] A recipe information selection unit, used for allowing users to select recipe information and send it to the model training and prediction module;
[0010] Model and parameter selection unit, used for users to select machine learning methods and model parameters for modeling, and send them to the model training and prediction module;
[0011] The instruction issuing unit is used for users to issue instructions for starting model training, stopping parameter search, and saving optimal parameters and corresponding model instructions, and sending them to the model training and prediction module;
[0012] The recipe evaluation input unit is used to provide the user with the trained model and experimental parameters, and send them to the model training and prediction module;
[0013] The visualization display unit is used to display the model training process information and model prediction results fed back by the model training and prediction modules.
[0014] Furthermore, the model training and prediction module includes:
[0015] A database is used to store specific data corresponding to the recipe, the trained model, and the optimal parameters of the model;
[0016] A data sample preprocessing unit, used to receive the recipe information sent by the recipe information selection unit, retrieve specific data corresponding to the recipe from the database, perform data preprocessing to form a recipe information matrix and store it in a cache, and feed back the matrix scale to the visualization display unit;
[0017] A parameter-searching and modeling unit is used to, after receiving a model training start instruction sent by the instruction issuing unit, call the recipe information matrix in the cache according to the machine learning method and model parameters sent by the model and parameter selection unit, start parameter-searching and modeling, and feed back model training process information to the visualization display unit in real time until the global optimal parameters are fed back when the parameter search is exhausted; if a stop parameter-searching instruction is received from the instruction issuing unit during the parameter-searching process, the parameter-searching is stopped and the current optimal parameters are fed back; and when a save optimal parameter and corresponding model instruction is received from the instruction issuing unit, the current optimal parameter or the global optimal parameter and the model are saved to the database;
[0018] The prediction unit is used to select the corresponding model and the optimal parameters corresponding to the model from the database according to the trained model and experimental parameters sent by the recipe evaluation input unit, call the cached recipe information matrix for prediction, and send the prediction result to the visualization display unit.
[0019] Furthermore, the data preprocessing method includes:
[0020] A1. Sort the formula information retrieved from the database, transform different components into features through topological sorting, and fix the feature positions to generate feature indexes;
[0021] A2. Determine the value of each feature in the formula according to the content of the specific components in the formula, fill in the missing components in the formula with zeros, and generate a feature matrix;
[0022] A3. Logarithmization is performed on the attributes set in the generated feature matrix, such as pressure, which are widely distributed and non-negative. The formula is:
[0023]
[0024] A4. If it is a modeling process, the corresponding label information is generated. The label information is the performance index of the solid propellant. If it is a prediction process, only features are generated.
[0025] Furthermore, the method for parameter-finding and modeling includes:
[0026] B1. Determine whether there is environmental information in the experimental parameters. If so, increase the number of features of the feature matrix in the preprocessing process;
[0027] B2. According to the number of parameters passed and the number of cross-validation, the sub-datasets are split and a sub-task list is further generated. The number of sub-datasets is the number of cross-validation times N. k , that is, the data set will be divided into N kEach set of parameters is used as a test set to evaluate the modeling performance. The evaluation index of each set of parameters is the value of the set of parameters in these N k The average of the evaluation indicators, and the total number of tasks is the product of all parameters after full combination and the number of cross-validation:
[0028] N task =N k ∏ p∈θ |p|
[0029] Where p is the set of all values of a parameter, and θ is the set of all parameters involved in parameter search. The list of these parameters and sub-datasets is the task list.
[0030] B3. After determining the task list, the model is initialized according to the input information. After determining the computing resources to be used and other environmental information, the modeling and calculation are started one by one according to the task list. After each modeling, the test set in the sub-dataset is used to evaluate the model performance and the information is saved.
[0031] Furthermore, the prediction method includes:
[0032] The features in the recipe information matrix are compared with the features used when the model was built. If the features match, the model is initialized again, the best model is loaded into the initialized model, and the model is used to predict the labels of the recipe information matrix.
[0033] Interactive solid propellant performance prediction method based on machine learning, including:
[0034] Step S1, the user terminal selects recipe information, the front end selects the information and sends it to the back end, the back end retrieves the specific data through the database, forms a recipe information matrix after preprocessing, and feeds back the matrix scale to the front end for display, completing the first interaction;
[0035] Step S2: The user selects the machine learning method and parameters used for modeling according to the matrix size. The parameters include model parameters and experimental environment parameters, and sends a model training start instruction. The front end transmits the model training start instruction to the back end. After receiving the instruction, the back end starts to search for parameters and model building, completing the second interaction.
[0036] Step S3: During the parameter search process, the backend feeds back the model training process information to the frontend in real time. When the user sends a command to stop the parameter search, the backend terminates the modeling process. After receiving the signal, the model stops calculating and feeds back the current optimal parameters. If the parameter search is not stopped, the model stops calculating and feeds back the global optimal parameters after exhausting all current parameter combinations. After the parameter search is completed, the user chooses whether to save the optimal parameters and the corresponding model to the database for prediction.
[0037] Step S4: After the modeling is completed, the user selects the recipe, the trained model and the experimental parameters, and sends them to the backend through the frontend. The backend can make predictions based on the information. The frontend will display the prediction results to the user in the form of a chart, and the user chooses whether to save the prediction chart.
[0038] The recipe information data selected by the user is preprocessed into a recipe information matrix as follows:
[0039] First, the backend organizes the recipe information retrieved from the database in the previous step, transforms different components into features through topological sorting, and fixes the feature positions to generate feature indexes;
[0040] Then, according to the specific component content in the formula, the value of the feature in each formula is determined, and the missing components in the formula are filled with zero values;
[0041] Logarithmization is performed on the widely distributed and non-negative attributes such as pressure in the generated feature matrix. The formula is:
[0042] If it is a modeling process, corresponding label information will also be generated. The labels are solid propellant performance indicators such as burning rate and safety, while the prediction process only generates features;
[0043] Feedback the data scale to the front end, and store the generated matrix in the cache, waiting to be called during subsequent modeling.
[0044] The specific method of parameter-finding modeling is:
[0045] First, the background will obtain the corresponding data matrix from the cache according to the information transmitted by the foreground, and determine whether there is environmental information such as pressure in the parameters. If so, the number of features of the feature matrix in the preprocessing process will be further increased.
[0046] According to the number of parameters passed and the number of cross-validation, the sub-dataset is split and further generated into a sub-task list. The number of sub-datasets is the number of cross-validation times N. k , that is, the data set will be divided into N k Each set of parameters is used as a test set to evaluate the modeling performance. The evaluation index of each set of parameters is the value of the set of parameters in these N k The total number of tasks is the product of all parameters after full combination and the number of cross-validation: N task =N k ∏ p∈θ |p|;
[0047] Where p is the set of all values of a parameter, and θ is the set of all parameters involved in parameter search. The list of these parameters and sub-datasets is the task list.
[0048] After determining the task list, the model will initialize the model according to the input information, and after determining the computing resources used and other environmental information, it will start modeling and calculation according to the list one by one. After each modeling, the test set in the sub-dataset will be used to evaluate the model performance, and the information will be saved. After each sub-task is completed, the model will feed back the calculated parameters and evaluation indicators to the front end for progress display, and will check whether a stop signal is received. If received, the calculation process will be terminated immediately. If no termination information is received, it will automatically terminate after traversing the task list.
[0049] After the process is completed, the model will count the parameters corresponding to the model with the highest evaluation index in the cross-validation, and store the optimal model in the cache. If the user chooses to save it, it will be stored in the database later, otherwise it will be destroyed.
[0050] The specific method of using the model to predict the recipe is:
[0051] After receiving the parameters from the frontend, the backend first generates the prediction recipe (without labels) into a numerical matrix through the preprocessing process.
[0052] The features of the generated numerical matrix are compared with the features used when building the model, and the prediction process can only continue if the features match.
[0053] Similar to modeling, the model is initialized, but only the optimal parameters during modeling are loaded into the initialized model. The model is then used to predict labels for the established numerical matrix.
[0054] Feed the prediction results back to the front end for subsequent processes such as visualization and saving.
[0055] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0056] (1) The present invention provides an interactive solid propellant performance prediction platform based on machine learning. Users only need to make the most basic selections to perform machine learning modeling and performance prediction on the selected formula and experimental environment. The training process can also be interrupted in time during the training process.
[0057] (2) The present invention automatically generates a matrix for machine learning modeling, eliminating the need for users to convert data by themselves, and automatically completes the most time-consuming and labor-intensive combination of multi-parameters and n-fold cross-validation to form a task list; at the same time, it provides visualization of the modeling process. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 is a flow chart of the present invention;
[0059] Figure 2A flowchart of data and processing;
[0060] Figure 3 Flowchart for parameter-finding modeling;
[0061] Figure 4 Forecasting flow chart. DETAILED DESCRIPTION
[0062] The present invention is further described in detail below in conjunction with examples, but the embodiments of the present invention are not limited thereto.
[0063] Embodiment 1:
[0064] Combined with Figure 1 As shown in the figure, the interactive solid propellant performance prediction platform based on machine learning includes:
[0065] The user input and result display module is used for users to select machine learning methods and parameter ranges and send them to the model training and prediction module; it is used to send the user's start training, interrupt parameter search instructions and save optimal parameter instructions to the model training and prediction module; it is used for users to select recipes, trained models and experimental parameters and send them to the model training and prediction module, and it is also used to visualize the feedback results of the model training and prediction module;
[0066] The model training and prediction module is used to retrieve the corresponding data in the database according to the machine learning method and parameter range selected by the user, form a recipe information matrix after preprocessing, and feed back the matrix size to the user input and result display module for display; and after receiving the user's start training instruction, it searches for parameters and builds models, and after receiving the user's interrupt instruction, stops calculating and feeds back the current optimal parameters; after receiving the user's save optimal parameter instruction, saves the current optimal parameters as the model parameters; and predicts the recipe performance according to the user's selected recipe, the trained model and the experimental parameters, and feeds back to the user input and result display module.
[0067] Furthermore, the user input and result display module includes:
[0068] A recipe information selection unit, used for allowing users to select recipe information and send it to the model training and prediction module;
[0069] Model and parameter selection unit, used for users to select machine learning methods and model parameters for modeling, and send them to the model training and prediction module;
[0070] The instruction issuing unit is used for users to issue instructions for starting model training, stopping parameter search, and saving optimal parameters and corresponding model instructions, and sending them to the model training and prediction module;
[0071] The recipe evaluation input unit is used to provide the user with the trained model and experimental parameters, and send them to the model training and prediction module;
[0072] The visualization display unit is used to display the model training process information and model prediction results fed back by the model training and prediction modules.
[0073] Furthermore, the model training and prediction module includes:
[0074] A database is used to store specific data corresponding to the recipe, the trained model, and the optimal parameters of the model;
[0075] A data sample preprocessing unit, used to receive the recipe information sent by the recipe information selection unit, retrieve specific data corresponding to the recipe from the database, perform data preprocessing to form a recipe information matrix and store it in a cache, and feed back the matrix scale to the visualization display unit;
[0076] A parameter-searching and modeling unit is used to, after receiving a model training start instruction sent by the instruction issuing unit, call the recipe information matrix in the cache according to the machine learning method and model parameters sent by the model and parameter selection unit, start parameter-searching and modeling, and feed back model training process information to the visualization display unit in real time until the global optimal parameters are fed back when the parameter search is exhausted; if a stop parameter-searching instruction is received from the instruction issuing unit during the parameter-searching process, the parameter-searching is stopped and the current optimal parameters are fed back; and when a save optimal parameter and corresponding model instruction is received from the instruction issuing unit, the current optimal parameter or the global optimal parameter and the model are saved to the database;
[0077] The prediction unit is used to select the corresponding model and the optimal parameters corresponding to the model from the database according to the trained model and experimental parameters sent by the recipe evaluation input unit, call the cached recipe information matrix for prediction, and send the prediction result to the visualization display unit.
[0078] Furthermore, the data preprocessing method includes:
[0079] A1. Sort the formula information retrieved from the database, transform different components into features through topological sorting, and fix the feature positions to generate feature indexes;
[0080] A2. Determine the value of each feature in the formula according to the content of the specific components in the formula, fill in the missing components in the formula with zeros, and generate a feature matrix;
[0081] A3. Logarithmization is performed on the attributes set in the generated feature matrix, such as pressure, which are widely distributed and non-negative. The formula is:
[0082]
[0083] A4. If it is a modeling process, the corresponding label information is generated. The label information is the performance index of the solid propellant. If it is a prediction process, only features are generated.
[0084] Furthermore, the method for parameter-finding and modeling includes:
[0085] B1. Determine whether there is environmental information in the experimental parameters. If so, increase the number of features of the feature matrix in the preprocessing process;
[0086] B2. According to the number of parameters passed and the number of cross-validation, the sub-datasets are split and a sub-task list is further generated. The number of sub-datasets is the number of cross-validation times N. k , that is, the data set will be divided into N k Each set of parameters is used as a test set to evaluate the modeling performance. The evaluation index of each set of parameters is the value of the set of parameters in these N k The average of the evaluation indicators, and the total number of tasks is the product of all parameters after full combination and the number of cross-validation:
[0087] N task =N k ∏ p∈θ |p|
[0088] Where p is the set of all values of a parameter, and θ is the set of all parameters involved in parameter search. The list of these parameters and sub-datasets is the task list.
[0089] B3. After determining the task list, the model is initialized according to the input information. After determining the computing resources to be used and other environmental information, the modeling and calculation are started one by one according to the task list. After each modeling, the test set in the sub-dataset is used to evaluate the model performance and the information is saved.
[0090] Furthermore, the prediction method includes:
[0091] The features in the recipe information matrix are compared with the features used when the model was built. If the features match, the model is initialized again, the best model is loaded into the initialized model, and the model is used to predict the labels of the recipe information matrix.
[0092] Embodiment 2:
[0093] Combination Figure 1 As shown, the interactive solid propellant performance prediction method based on machine learning includes:
[0094] Step S1, the user selects the recipe information, the front end selects the information and sends it to the back end, the back end retrieves the specific data through the database, forms a recipe information matrix after preprocessing, and feeds back the matrix scale to the front end for display, completing the first interaction; this step automatically generates a matrix for machine learning modeling, eliminating the need for users to convert data by themselves.
[0095] Step S2: The user selects the machine learning method and parameters used for modeling according to the matrix size. The parameters include model parameters and experimental environment parameters, and sends a model training start instruction. The front end passes the model training start instruction to the back end. After receiving the instruction, the back end starts to search for parameters and modeling, completing the second interaction. This interaction is equivalent to generating a configuration file for the modeling process in machine learning, which can greatly reduce the threshold for using machine learning methods.
[0096] Step S3: During the parameter search process, the backend feeds back the model training process information to the frontend in real time. When the user sends a command to stop parameter search, the backend terminates the modeling process. After receiving the signal, the model stops calculating and feeds back the current optimal parameters. If the parameter search is not stopped, the model stops calculating and feeds back the global optimal parameters after exhausting all current parameter combinations. After the parameter search is completed, the user chooses whether to save the optimal parameters and the corresponding model to the database for prediction. This step is the core of the platform processing. It not only needs to automatically complete the most time-consuming and labor-intensive multi-parameter and n-fold cross-validation combinations to form a task list, but also provides an interruption option for the visualization of the modeling process for the long modeling process, which improves the transparency and practicality of the processing process and facilitates users to adjust their configurations in a timely manner.
[0097] Step S4: After the modeling is completed, the user selects the recipe, the trained model and the experimental parameters, and sends them to the backend through the frontend. The backend can make predictions based on the information. The frontend will display the prediction results to the user in the form of a chart, and the user chooses whether to save the prediction chart. The built model is utilized, and the feature comparison, report generation and other processes in the prediction step are integrated into a unified process, ultimately achieving the purpose of automatic modeling without computer foundation.
[0098] like Figure 2 As shown, the recipe information data selected by the user is preprocessed into a recipe information matrix as follows:
[0099] First, the backend organizes the recipe information retrieved from the database in the previous step, transforms different components into features through topological sorting, and fixes the feature positions to generate feature indexes;
[0100] Then, according to the specific component content in the formula, the value of the feature in each formula is determined, and the missing components in the formula are filled with zero values;
[0101] Logarithmization is performed on the widely distributed and non-negative attributes such as pressure in the generated feature matrix. The formula is:
[0102] If it is a modeling process, corresponding label information will also be generated. The labels are solid propellant performance indicators such as burning rate and safety, while the prediction process only generates features;
[0103] Feedback the data scale to the front end, and store the generated matrix in the cache, waiting to be called during subsequent modeling.
[0104] like Figure 3 As shown in the figure, the specific method of parameter-finding modeling is:
[0105] First, the background will obtain the corresponding data matrix from the cache according to the information transmitted by the foreground, and determine whether there is environmental information such as pressure in the parameters. If so, the number of features of the feature matrix in the preprocessing process will be further increased.
[0106] According to the number of parameters passed and the number of cross-validation, the sub-dataset is split and further generated into a sub-task list. The number of sub-datasets is the number of cross-validation times N. k , that is, the data set will be divided into N k Each set of parameters is used as a test set to evaluate the modeling performance. The evaluation index of each set of parameters is the value of the set of parameters in these N k The total number of tasks is the product of all parameters after full combination and the number of cross-validation: N task =N k ∏ p∈θ |p|;
[0107] Where p is the set of all values of a parameter, and θ is the set of all parameters involved in parameter search. The list of these parameters and sub-datasets is the task list.
[0108] After determining the task list, the model will initialize the model according to the input information, and after determining the computing resources used and other environmental information, it will start modeling and calculation according to the list one by one. After each modeling, the test set in the sub-dataset will be used to evaluate the model performance, and the information will be saved. After each sub-task is completed, the model will feed back the calculated parameters and evaluation indicators to the front end for progress display, and will check whether a stop signal is received. If received, the calculation process will be terminated immediately. If no termination information is received, it will automatically terminate after traversing the task list.
[0109] After the process is completed, the model will count the parameters corresponding to the model with the highest evaluation index in the cross-validation, and store the optimal model in the cache. If the user chooses to save it, it will be stored in the database later, otherwise it will be destroyed.
[0110] like Figure 4 As shown in the figure, the specific method of using the model to predict the recipe is:
[0111] After receiving the parameters from the frontend, the backend first generates the prediction recipe (without labels) into a numerical matrix through the preprocessing process.
[0112] The features of the generated numerical matrix are compared with the features used when building the model, and the prediction process can only continue if the features match.
[0113] Similar to modeling, the model is initialized, but only the optimal parameters during modeling are loaded into the initialized model. The model is then used to predict labels for the established numerical matrix.
[0114] Taking the burning rate performance prediction of solid propellant as an example, combined with Figure 1-Figure 4 The process is described below.
[0115] Preparation: First, the formula and burning rate information of solid propellant have been stored in a database in advance, and the two associations of "solid propellant formula-component" and "solid propellant formula-burning rate" have been made. The information extracted when used is in the form of an association list. For example, for solid propellant formula 1, its component combination is "A: 80%, B: 10%, C: 5%, D: 5%", and the corresponding burning rate is 2.6mm / s, then the extracted information can be expressed as
[0116] Recipe ID property Numeric 1 Component A content (%) 80 1 Component B content (%) 10 1 Component C content (%) 5 1 Component D content (%) 5 1 Pressure(MPa) 6 1 Burning rate (mm / s) 2.6
[0117] This list format.
[0118] Step 1: User selects a recipe for modeling
[0119] Assuming that the user has a chemistry background and no computer and data mining background, the user first filters the formula through the database interface where the information has been entered. For example, the user can use the filter provided by the user end to select all formulas that "contain component A and at least one of components D, E, and F, and the burning rate is not less than 2mm / s" for model building through simple data. After the user completes the selection, the information will be passed to the background main control program through the front end, and the background will automatically generate a search table based on the screening conditions. Assuming that there are 50 formulas that meet the conditions, here is a part of them, and the components are shown in Table 1:
[0120] Table 1 Components
[0121] Recipe ID property Proportion(%) 1 Component A content (%) 80 1 Component B content (%) 10 1 Component C content (%) 5 1 Component D content (%) 5 1 Pressure(MPa) 6 1 Burning rate (mm / s) 2.6 2 A 90 2 E 10 2 Pressure(MPa) 14 2 Burning rate (mm / s) 3.1 … … … 50 A 70 50 D 15 50 F 15 50 Pressure(MPa) 4 50 Burning rate (mm / s) 2.1
[0122] The background will automatically process the component table further. The program will automatically count which components are used. For the convenience of description, assume that these 50 formulas use six components, A, B, C, D, E, and F. Then, combined with pressure and combustion rate, the following information matrix will be generated, as shown in Table 2.
[0123] Table 2 Information matrix
[0124] Recipe ID Component A Component B Component C Component D Component E Component F pressure Burning rate 1 80 10 5 5 0 0 6 2.6 2 90 0 0 0 10 0 14 3.1 … … … … … … … … … 50 70 0 0 15 15 0 4 2.1
[0125] That is, "if the formula contains a certain component, record its percentage, if not, take the value as 0". In particular, the background will automatically logarithmize the pressure
[0126] x=log(1+x)
[0127] The final information matrix is obtained, as shown in Table 3:
[0128] Table 3 Final information matrix
[0129] Recipe ID Component A Component B Component C Component D Component E Component F pressure Burning rate 1 80 10 5 5 0 0 log(7) 2.6 2 90 0 0 0 10 0 log(15) 3.1 … … … … … … … … … 50 70 0 0 15 15 0 log(5) 2.1
[0130] After the information matrix is generated, the background will transmit the length and width of the matrix (corresponding to the total number of selected recipes and the number of components contained in these recipes) back to the front end, and store the component information matrix and the burning rate list into the cache, waiting for subsequent instructions.
[0131] Step 2: User selects machine learning algorithm and related parameters
[0132] After knowing the data scale, users can select machine learning methods and which columns (i.e. attributes) to use as features and which columns to use as prediction labels based on the data scale. This example is for burning rate prediction, so users will select component AF and pressure as features and burning rate as prediction label. The platform integrates machine learning methods in the scikit-learn library in Python. Here, assuming that support vector regression (SVR) is selected, you can choose which built-in parameters to search for according to the prompts. Here, assuming that the user clicks to search for two parameters gamma (kernel function coefficient) and C (regularization parameter). The front end will automatically generate a parameter search box, which contains four input boxes: "Starting value", "Change step", "Ending value" and "Whether to index". The first three are numerical inputs, and the last one is a selection box. Assuming that the starting value of gamma is -5, the change step is 2, and the ending value is 5, the backend will set gamma to -5, -3, -1, 1, 3, 5 respectively according to these values;
[0133] If you check "Exponential", the value of gamma in the background will become e -5 ,e-3 ,e -1 ,e 1 ,e 3 ,e 5 ;
[0134] After selecting the method and parameter search range, the user can choose to view the current parameter search scale. Assuming that the gamma selection is as shown above, and the C selection is the same as gamma, a total of 6×6=36 sets of parameters will be generated for modeling. In addition, the user also needs to select the number of n-fold cross-validation. Here, assuming that the user selects 5-fold cross-validation, the final number of modeling will become 6×6×5=180 times. If the user thinks the scale is acceptable, he can click the "Start Modeling" button. The front end will pass the selected method, parameter search range, and number of cross-validations to the back end. After receiving the signal, the back end will start parameter search and modeling. (Note: If the user has no basic knowledge, the system will also make some template suggestions based on the data scale. For example, if the data volume is less than 100, it will recommend the use of SVR. If the data volume is more than 100, it will recommend the use of kernel ridge regression (kernelRidge) or random forest (RandomForest), and provide a simple parameter search template.)
[0135] Step 3: Background parameter search and modeling
[0136] When modeling begins, the backend receives the following information: the ID of the matrix in the cache, the machine learning method and parameters used for parameter search, the number of cross-validation, and the number of columns used as features and labels. The backend first combines the parameters in a many-to-many manner to generate a parameter search list, as shown in Table 4.
[0137] Table 4 Parameter search list
[0138]
[0139]
[0140] Then, the backend will group and reassemble the data according to the number of cross-validation times. In this case, it is a 5-fold cross-validation. The backend will take out the information matrix in the cache, divide it into 5 equal parts by row, and then reassemble it according to Table 5:
[0141] Table 5 Dataset partition table
[0142] Data Group (Each Fold) Training set Test Set i 1,2,3,4 5 ii 1,2,3,5 4 iii 1,2,4,5 3 iv 1,3,4,5 2 v 2,3,4,5 1
[0143] That is, in each fold, each of the five sub-datasets is used as a test set, and the remaining four are combined into a training set. For each sub-dataset, the backend will also divide the sub-dataset into information matrix and label column by column according to the column numbers of the features and labels passed in by the frontend. Generally, the information matrix is x and the label is y. Considering the grouping and training and test sets, the sub-information matrix is generally recorded as The label is Where k represents the number of data groups, and its value is {i, ii, iii, iv, v}, and m represents the training set or test set, and its value is {train, test}, for example It means "the test set composed of AF components and pressure of the 1st, 2nd, 4th and 5th groups of data grouped by row in the original information matrix". It can be seen that there are 36 sets of parameters and 5 sets of sub-datasets in this example, and each set of parameters needs to be calculated with all sub-datasets, so a total of 36×5=180 calculations are required. In each calculation, the background will first instantiate a new SVR model (all parameters and hyperparameters are initial values), and update the parameters that need to be found, which are recorded as model. Then, it is trained with the training set in the sub-dataset, that is, it is executed
[0144]
[0145] After the training, the parameters of the model are optimized, and then the trained model is used to predict the corresponding test set, that is,
[0146]
[0147] Since the task of burning rate prediction is a regression task, the mean squared error (MSE) can be used as an evaluation indicator to measure the prediction accuracy. and the original The value between (the smaller the better), that is
[0148]
[0149] After each modeling, the background will record the score, similar to the following Table 6:
[0150] Table 6 Dataset score table
[0151]
[0152]
[0153] That is, for each set of parameters, there are 5 scores (there are n scores for n folds), then the average of these scores can be used as the final score, similar to the following Table 7:
[0154] Table 7 Final score table
[0155]
[0156] Therefore, after calculating all modeling tasks and obtaining the mean, the background will record the parameter combination with the smallest score (the smaller the MSE, the better) as the optimal parameter. This record will be fed back to the user through the front end, and the optimal parameter combination and MSE value, combined with the model method name (i.e. SVR), the column name used for modeling (i.e. column AF, and pressure), will be stored in the cache. If the user chooses to save it later, it will be stored in the database. If the user chooses not to save it, the corresponding information will be destroyed from the cache.
[0157] In particular, since the actual modeling efficiency may be much higher than expected, after the calculation of each set of parameters (corresponding to 5 sub-models in this example) is completed, the background will check once whether it has received a stop signal from the user. If a signal is received, the background will automatically stop the current modeling process and find the best combination among the trained parameter combinations for subsequent processes.
[0158] Step 4: Make predictions for untested recipes
[0159] If a trained model already exists in the database, prediction operations can be performed. Unlike modeling, prediction operations do not require label vectors, only feature matrices. In fact, since no experiments have been conducted, non-recipe information such as pressure will not be recorded. Therefore, for prediction, the user needs to provide a pressure range after selecting the model and recipe. The selection of the pressure range is similar to the parameter range. The initial pressure value, end value, and incremental step size can be set. Here it is assumed that the user has selected a new recipe, including component A 70%, component B 20%, component F 10%, and the pressure is set to an initial 0MPa, an end value of 100MPa, and a step size of 5MPa. The model selects the model trained in the previous step. Assume that the optimal parameter is gamma and e -5 , C is e 3 , then the model will first detect whether the components in the formula to be predicted are used in modeling. In this example, the components used in the formula to be predicted are A, B and F, which meet the AF used in training. Therefore, the model will generate pressure values, take the logarithm and combine them with the components to generate a numerical matrix for prediction, as shown in Table 8:
[0160] Table 8 Numerical matrix for prediction
[0161] Recipe ID Component A Component B Component C Component D Component E Component F pressure 1 70 20 0 0 0 10 log(1) 2 70 20 0 0 0 10 log(6) … … … … … … … … 50 70 20 0 0 0 10 log(101)
[0162] That is, in this matrix, there is no label column, and only the pressure is different (for other attributes, there may be multiple attribute columns, but the components are still consistent, and the attributes are arranged and combined). This matrix can be recorded as x pred After getting the matrix, the background will initialize an SVR model and assign gamma and C to e -5 and e 3 , and then make a prediction
[0163] y pred =model.predict(x pred )
[0164] Finally, the model will feed back the predicted values to the user in the form of a list to complete the prediction.
[0165] The above four steps are introduced using the burning rate prediction as an example. The prediction of other attributes is similar, with only slight differences in the matrix generation of the subsequent prediction part.
[0166] Although the present invention is described herein with reference to the illustrative embodiments of the present invention, the above embodiments are only preferred embodiments of the present invention, and the embodiments of the present invention are not limited to the above embodiments. It should be understood that those skilled in the art can design many other modifications and embodiments, which will fall within the scope and spirit of the principles disclosed in this application.
Claims
1. An interactive solid propellant performance prediction platform based on machine learning, characterized by: include: The user input and result display module is used for users to select machine learning methods and parameter ranges and send them to the model training and prediction module; It is used to send the user's start training, interrupt parameter search instructions and save optimal parameter instructions to the model training and prediction module; it is used for the user to select the recipe, trained model and experimental parameters and send them to the model training and prediction module, and it is also used to visualize the feedback results of the model training and prediction module; The model training and prediction module is used to retrieve the corresponding data in the database according to the machine learning method and parameter range selected by the user, form a recipe information matrix after preprocessing, and feed back the matrix scale to the user input and result display module for display; After receiving the user's instruction to start training, it searches for parameters and builds models, and after receiving the user's interruption instruction, it stops calculating and feeds back the current best parameters; after receiving the user's instruction to save the best parameters, it saves the current best parameters as the parameters of the model; and according to the user's selected formula, the trained model and experimental parameters, it predicts the formula performance and feeds back to the user input and result display module; The user input and result display module includes: A recipe information selection unit, used for allowing users to select recipe information and send it to the model training and prediction module; Model and parameter selection unit, used for users to select machine learning methods and model parameters for modeling, and send them to the model training and prediction module; The instruction issuing unit is used for users to issue instructions for starting model training, stopping parameter search, and saving optimal parameters and corresponding model instructions, and sending them to the model training and prediction module; The recipe evaluation input unit is used to provide the user with the trained model and experimental parameters, and send them to the model training and prediction module; Visual display unit, used to display the model training process information and model prediction results fed back by the model training and prediction modules; The model training and prediction module includes: A database is used to store specific data corresponding to the recipe, the trained model, and the optimal parameters of the model; A data sample preprocessing unit, used to receive the recipe information sent by the recipe information selection unit, retrieve specific data corresponding to the recipe from the database, perform data preprocessing to form a recipe information matrix and store it in a cache, and feed back the matrix scale to the visualization display unit; A parameter-searching and modeling unit is used to, after receiving a model training start instruction sent by the instruction issuing unit, call the recipe information matrix in the cache according to the machine learning method and model parameters sent by the model and parameter selection unit, start parameter-searching and modeling, and feed back model training process information to the visualization display unit in real time until the global optimal parameters are fed back when the parameter search is exhausted; if a stop parameter-searching instruction is received from the instruction issuing unit during the parameter-searching process, the parameter-searching is stopped and the current optimal parameters are fed back; and when a save optimal parameter and corresponding model instruction is received from the instruction issuing unit, the current optimal parameter or the global optimal parameter and the model are saved to the database; A prediction unit, used to select a corresponding model and optimal parameters corresponding to the model from a database according to the trained model and experimental parameters sent by the recipe evaluation input unit, call the cached recipe information matrix for prediction, and send the prediction result to the visualization display unit; The data preprocessing method comprises: A1. Sort the formula information retrieved from the database, transform different components into features through topological sorting, and fix the feature positions to generate feature indexes; A2. Determine the value of each feature in the formula according to the specific content of the components in the formula, fill in the missing components in the formula with zeros, and generate a feature matrix; A3. Perform logarithmic operation on the attributes set in the generated feature matrix. The formula is: A4. If it is a modeling process, the corresponding label information is generated. The label information is the performance index of the solid propellant. If it is a prediction process, only features are generated.
2. The interactive solid propellant performance prediction platform based on machine learning according to claim 1, characterized in that: The method for parameter-finding and modeling comprises: B1. Determine whether there is environmental information in the experimental parameters. If so, increase the number of features of the feature matrix in the preprocessing process; B2. According to the number of parameters passed and the number of cross-validation, the sub-datasets are split and a sub-task list is further generated. The number of sub-datasets is the number of cross-validation times N. k , that is, the data set will be divided into N k Each set of parameters is used as a test set to evaluate the modeling performance. The evaluation index of each set of parameters is the value of the set of parameters in these N sets. k The average of the evaluation indicators, and the total number of tasks is the product of all parameters after full combination and the number of cross-validation: N task =N k ∏ p∈θ |p| Where p is the set of all values of a parameter, and θ is the set of all parameters involved in parameter search. The list of these parameters and sub-datasets is the task list. B3. After determining the task list, the model is initialized according to the input information. After determining the computing resource environment information to be used, modeling and calculation are started one by one according to the task list. After each modeling, the test set in the sub-dataset is used to evaluate the model performance and the information is saved.
3. The interactive solid propellant performance prediction platform based on machine learning according to claim 1, characterized in that: The prediction method includes: The features in the recipe information matrix are compared with the features used when the model was built. If the features match, the model is initialized again, the best model is loaded into the initialized model, and the model is used to predict the labels of the recipe information matrix.
4. An interactive solid propellant performance prediction method based on machine learning implemented by the interactive solid propellant performance prediction platform based on machine learning as claimed in claim 1, characterized in that: include: Step S1, the user terminal selects recipe information, the front end selects the information and sends it to the back end, the back end retrieves the specific data through the database, forms a recipe information matrix after preprocessing, and feeds back the matrix scale to the front end for display, completing the first interaction; Step S2: The user selects the machine learning method and parameters used for modeling according to the matrix size. The parameters include model parameters and experimental environment parameters, and sends a model training start instruction. The front end transmits the model training start instruction to the back end. After receiving the instruction, the back end starts to search for parameters and model building, completing the second interaction. Step S3: During the parameter search process, the backend feeds back the model training process information to the frontend in real time. When the user sends a command to stop the parameter search, the backend terminates the modeling process. After receiving the signal, the model stops calculating and feeds back the current optimal parameters. If the parameter search is not stopped, the model stops calculating and feeds back the global optimal parameters after exhausting all current parameter combinations. After the parameter search is completed, the user chooses whether to save the optimal parameters and the corresponding model to the database for prediction. Step S4: After completing the modeling, the user selects the recipe, the trained model and the experimental parameters, and sends them to the backend through the frontend. The backend can make predictions based on the information. The frontend will display the prediction results to the user in the form of a chart, and the user chooses whether to save the prediction chart.
Citation Information
Patent Citations
A universal machine learning hyper-parametric black box optimization method and system
CN109447277A
Alloy heat treatment technology optimization method based on machine learning
CN111286599A
System, method and device for realizing data mining and storage medium
CN113392151A
Biphase material composite pipe performance prediction method based on machine learning
CN115563824A
Machine learning platform
CN116108928A
