CN117766058B - Safety performance prediction method of RDX modified double-base propellant - Google Patents
Safety performance prediction method of RDX modified double-base propellant Download PDFInfo
- Publication number
- CN117766058B CN117766058B CN202311821308.8A CN202311821308A CN117766058B CN 117766058 B CN117766058 B CN 117766058B CN 202311821308 A CN202311821308 A CN 202311821308A CN 117766058 B CN117766058 B CN 117766058B
- Authority
- CN
- China
- Prior art keywords
- data set
- model
- optimal
- propellant
- sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 239000003380 propellant Substances 0.000 title claims abstract description 72
- 238000000034 method Methods 0.000 title claims abstract description 37
- 230000035945 sensitivity Effects 0.000 claims abstract description 73
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 41
- 238000010801 machine learning Methods 0.000 claims abstract description 38
- 238000005516 engineering process Methods 0.000 claims abstract description 18
- 238000002790 cross-validation Methods 0.000 claims abstract description 10
- 238000012549 training Methods 0.000 claims description 12
- 238000005457 optimization Methods 0.000 claims description 10
- 238000012360 testing method Methods 0.000 claims description 10
- 238000009826 distribution Methods 0.000 claims description 9
- 238000013528 artificial neural network Methods 0.000 claims description 6
- 230000009977 dual effect Effects 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 238000007637 random forest analysis Methods 0.000 claims description 5
- 238000011156 evaluation Methods 0.000 claims description 4
- 238000012417 linear regression Methods 0.000 claims description 4
- 238000012795 verification Methods 0.000 claims description 4
- 238000010238 partial least squares regression Methods 0.000 claims description 3
- 230000001351 cycling effect Effects 0.000 claims description 2
- 238000009827 uniform distribution Methods 0.000 claims description 2
- 230000007547 defect Effects 0.000 abstract description 3
- 238000012512 characterization method Methods 0.000 abstract description 2
- XTFIVUDBNACUBN-UHFFFAOYSA-N 1,3,5-trinitro-1,3,5-triazinane Chemical compound [O-][N+](=O)N1CN([N+]([O-])=O)CN([N+]([O-])=O)C1 XTFIVUDBNACUBN-UHFFFAOYSA-N 0.000 description 8
- 239000002360 explosive Substances 0.000 description 7
- 238000002474 experimental method Methods 0.000 description 6
- 239000000203 mixture Substances 0.000 description 5
- 238000011161 development Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 238000004880 explosion Methods 0.000 description 4
- 238000009472 formulation Methods 0.000 description 4
- 239000000463 material Substances 0.000 description 4
- 239000000126 substance Substances 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- LLVFMJQPUWTQTM-UHFFFAOYSA-M 2-carboxy-5-hydroxyphenolate;lead(2+) Chemical compound [Pb+2].OC1=CC=C(C([O-])=O)C(O)=C1 LLVFMJQPUWTQTM-UHFFFAOYSA-M 0.000 description 2
- OKTJSMMVPCPJKN-UHFFFAOYSA-N Carbon Chemical compound [C] OKTJSMMVPCPJKN-UHFFFAOYSA-N 0.000 description 2
- SNIOPGDIGTZGOP-UHFFFAOYSA-N Nitroglycerin Chemical compound [O-][N+](=O)OCC(O[N+]([O-])=O)CO[N+]([O-])=O SNIOPGDIGTZGOP-UHFFFAOYSA-N 0.000 description 2
- 239000000006 Nitroglycerin Substances 0.000 description 2
- 230000009471 action Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 239000000654 additive Substances 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 229910052799 carbon Inorganic materials 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000000295 complement effect Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000005284 excitation Effects 0.000 description 2
- 229960003711 glyceryl trinitrate Drugs 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 238000002156 mixing Methods 0.000 description 2
- 229920001220 nitrocellulos Polymers 0.000 description 2
- 239000002245 particle Substances 0.000 description 2
- PAWQVTBBRAZDMG-UHFFFAOYSA-N 2-(3-bromo-2-fluorophenyl)acetic acid Chemical compound OC(=O)CC1=CC=CC(Br)=C1F PAWQVTBBRAZDMG-UHFFFAOYSA-N 0.000 description 1
- 201000004569 Blindness Diseases 0.000 description 1
- 239000000020 Nitrocellulose Substances 0.000 description 1
- 230000003321 amplification Effects 0.000 description 1
- 239000011230 binding agent Substances 0.000 description 1
- 239000006229 carbon black Substances 0.000 description 1
- 239000003054 catalyst Substances 0.000 description 1
- 239000003795 chemical substances by application Substances 0.000 description 1
- 238000002485 combustion reaction Methods 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000001186 cumulative effect Effects 0.000 description 1
- 238000005315 distribution function Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000013213 extrapolation Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 239000007800 oxidant agent Substances 0.000 description 1
- 230000001590 oxidative effect Effects 0.000 description 1
- 239000004014 plasticizer Substances 0.000 description 1
- 238000011158 quantitative evaluation Methods 0.000 description 1
- 230000005610 quantum mechanics Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 230000000638 stimulation Effects 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
- 238000002759 z-score normalization Methods 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a safety performance prediction method of an RDX modified double-base propellant, which constructs an original data set based on a propellant sample; amplifying the original data set for multiple times by adopting a data enhancement technology to obtain a corresponding enhanced data set, and then carrying out standardization treatment; modeling and predicting standardized data, obtaining optimal super parameters of a machine learning algorithm by combining ten-fold cross validation with grid parameter searching, executing ten-fold cross validation under the optimal super parameters, and selecting an optimal friction sensitivity prediction model and an impact sensitivity prediction model; after the model is verified, the propellant friction sensitivity and the impact sensitivity are predicted. The invention solves the problems that the components of the propellant are complex, the influence factors are many, and the traditional molecular structure property characterization can not be used; the active learning and Mixup mixed interpolation technology are combined, the defect of insufficient sample size is overcome, and high-quality data enhancement is realized; the application range is wide, the operation is simple, and the prediction is rapid and accurate.
Description
Technical Field
The invention relates to the technical field of propellant safety performance prediction, in particular to a safety performance prediction method of an RDX modified double-base propellant.
Background
Because the formula contains a large amount of high-sensitivity energetic materials such as ammonium nitrate explosive, nitrocotton, nitroglycerin and the like, the propellant is extremely easy to be impacted, rubbed, heated and other external energy effects to cause rapid chemical reaction in the production, storage, transportation and use processes, thereby causing accidents such as combustion, explosion and the like. The safety performance of the propellant is characterized by the difficulty of violent reaction under the stimulation of certain external conditions such as heat, flame, mechanical action (friction and impact), electrostatic spark and the like, and the safety degree of the propellant can be measured by adopting different sensitivity according to different external excitation energy sources. The mechanical action is one of the main external excitation sources possibly encountered by the propellant in the life cycle, so the mechanical sensitivity (friction sensitivity and impact sensitivity) is a key index for representing the safety performance of the propellant, and the quantitative evaluation of the mechanical sensitivity is favorable for determining the application range and reliability of the propellant, and plays a vital role in the development of the solid rocket power technology.
At present, WL-1 vertical drop hammer instrument is commonly adopted in China to measure the impact sensitivity of the propellant, and the representation method mainly comprises a characteristic drop height method, an explosion percentage representation method and a 50% explosion critical drop height representation method; friction sensitivity is commonly measured using a WM-1 type pendulum friction meter, expressed in terms of the explosion percentage method. Although obtaining the mechanical sensitivity of the propellant through experiments is the mode with the highest reliability, the problems of long time consumption, high test cost, high risk, poor experimental repeatability and the like still exist, so that the implementation performance is low, and the development process of the propellant is hindered to a certain extent. If the safety performance of the propellant can be known in advance in the propellant formulation design stage, the blindness of subsequent experiments can be avoided, the time is shortened, and the economic cost is reduced. Therefore, it is important to develop a high-efficiency and rapid prediction method.
The sensitivity prediction theory in the prior art is mainly focused on the prediction of the mechanical sensitivity, especially the impact sensitivity, of the simple substance explosive, and the application method mainly comprises an empirical formula, quantum mechanical calculation and a neural network, wherein the empirical formula generally only needs a molecular structure, but is generally limited in explosive molecules; although the quantum mechanics has a strong theoretical foundation as a support, complex software is needed for calculation, and the time is long; the neural network requires a large amount of raw data as a training set, and training is difficult and long. Compared with simple substance explosive, the propellant has complex components, influence factors are more difficult to grasp, a single analysis microscopic molecular structure can not meet the condition of multiple components, the component proportion has great influence on the mechanical sensitivity of the propellant, the type and the content of main energetic materials, the granularity of a sample and even the content of different additives can also influence the sensitivity, and the traditional empirical formula method or quantum calculation is not applicable to a propellant system at all. With the rapid development of computer technology, machine learning has been widely used in the material field by virtue of its flexibility and strong learning ability. Machine learning is a science driven by data, the quantity and quality of the data are key factors influencing the robustness of a prediction model, and when the sample size of the data with labels is small, the model is extremely easy to be subjected to over fitting. However, the propellant sample is more difficult to obtain than the simple substance explosive, the measurement experiment cost of the safety performance is high, the time consumption is long, the available propellant safety performance sample is extremely lack, the problems of poor model prediction performance, serious overfitting and the like are caused by directly applying machine learning, and the corresponding technology must be developed to break the limitation of the data volume.
In summary, it is necessary to develop a fast and accurate safety performance prediction method aiming at the characteristics of complex components, multiple influencing factors and scarce sample size of the propellant, so that theoretical basis can be provided for the safe use of the propellant, and the formula optimization is guided.
Disclosure of Invention
The invention aims to provide a method for predicting the safety performance of an RDX modified double-base propellant, which is used for solving the problem that a method capable of rapidly and accurately predicting the safety performance of the propellant is not available in the prior art.
The invention solves the problems by the following technical proposal:
a method for predicting the safety performance of an RDX modified dual-based propellant comprising:
Step S100, analyzing propellant sample data, characterizing the sample by using component proportion information and granularity information, and constructing an original data set D 0;
Step S200, amplifying the original data set D 0 for a plurality of times by adopting a data enhancement technology to obtain corresponding enhancement data sets D 1、D2、D3 and D 4;
step S300, performing Z-score standardization processing on the amplified data in the data set by adopting a standardization device to obtain standardized data;
step S400, modeling and predicting standardized data by using a machine learning algorithm, obtaining optimal super parameters of the machine learning algorithm by combining ten-fold cross verification with grid parameter searching, executing ten-fold cross verification by each machine learning algorithm under the optimal super parameters, selecting an optimal machine learning model and an optimal enhancement data set, and constructing an RBF core SVR model for RBF core RDX friction sensitivity prediction and an ANN model for impact sensitivity prediction based on the optimal enhancement data set and the optimal machine learning model;
Step S500, verifying the SVR model and the ANN model by using an external sample, and storing a normalizer, the SVR model and the ANN model if the prediction errors are within a set threshold;
And S600, respectively inputting the propellant component proportion and the granularity information into the SVR model and the ANN model to obtain predicted values of friction sensitivity and impact sensitivity.
Further, the step S100 specifically includes:
Step S110, the components of different formulas of the propellant are combined to realize characteristic unification, and finally 13 components are contained, wherein each component corresponds to one characteristic, the component proportion is represented by mass percent, and if the original formula does not contain a certain component, the component proportion is represented by 0;
step S120, taking the logarithmic value of granularity as a characteristic value and adding the logarithmic value into a characteristic vector;
In step S130, each group of samples represents the safety performance by using the friction sensitivity and the impact sensitivity, each sample is characterized as a 14-dimensional feature vector, and contains 13-dimensional component proportion and 1-dimensional granularity information, and the original data set is marked as D 0 according to two label values of the friction sensitivity and the impact sensitivity.
Further, the step S200 specifically includes:
step S210, constructing a proxy model C f,0 and C i,0,Cf,0 as a friction sensitivity prediction model by using an original data set D 0, and C i,0 as an impact sensitivity prediction model;
step S220, randomly operating the Mixup mixed difference method for N times to generate N new data, so as to form a sample pool U 0 to be selected;
Step S230, predicting a sample pool U 0 to be selected by using C f,0 and C i,0, calculating an expected lifting value EI of each sample to be selected, sorting EI values of each sample to be selected in a descending order, adding samples M before the selection into an original dataset D 0 to form an enhanced dataset D 1, and taking N-M samples to be selected as sample pools U 1, wherein N and M are integers and M is smaller than N;
Step S240, retraining by using the enhanced dataset D 1 to obtain updated proxy models C f,1 and C i,1, predicting a sample pool U 1 by using C f,1 and C i,1, calculating an expected lifting value EI, sorting in descending order, selecting samples M before ranking, adding the samples M before ranking into the dataset D 1 to form the enhanced dataset D 2, and taking N-2*M samples to be selected as a sample pool U 2;
Step S250, retraining by using the enhancement data set D 2 to obtain updated proxy models C f,2 and C i,2, predicting a sample pool U 2, calculating an expected lifting value EI, sorting in a descending order, selecting samples M before ranking, adding the samples M to the enhancement data set D 2 to form an enhancement data set D 3, and taking N-3*M samples to be selected as the sample pool U 3;
Step S260, retraining with the enhanced data set D 3 to obtain updated proxy models C f,3 and C i,3, predicting the sample pool U 3, calculating the expected lifting value EI, sorting in descending order, selecting the samples M before ranking, adding them to the enhanced data set D 3 to form the enhanced data set D 4, and leaving N-4*M samples to be selected.
Further, the step S400 specifically includes:
Step S410, determining parameter optimization spaces of ten machine learning algorithms; the machine learning algorithm is a multiple linear regression MLR, a partial least squares regression PLSR, a kernel ridge regression KRR, a LASSO algorithm LASSO, a K nearest neighbor regression KNN, a support vector regression SVR, a random forest RF, a limit gradient lifting XGB, a lightweight gradient lifting LGB and an artificial neural network ANN respectively;
Step S420, randomly dividing the D 1 reinforced data set into 10 parts according to friction/impact sensitivity value distribution, ensuring uniform distribution of each part of data, wherein 9 parts of data are used as training sets, and the rest 1 part of data are used as test sets;
Step S430, executing a parameter optimization process: for each algorithm, respectively:
For each group of super parameters, constructing a corresponding algorithm model on a training set, and obtaining an evaluation index R 2 on a testing set; cycling for 10 times to ensure that each data is used as a training set and a testing set, correspondingly, 10R 2 values are generated, and the average value is taken as the performance of the model under the super-parameters of the group;
After traversing each group of super parameters, obtaining R 2 values under all super parameter combinations, and selecting a group of super parameters corresponding to the maximum R 2 value as the optimal super parameters of the machine learning algorithm.
Step S440: step S420 and step S430 are sequentially performed on the enhanced data set D 2、D3、D4, so as to obtain the optimal super parameters of each machine learning algorithm under different data sets.
Further, the step S500 specifically includes:
Step S510, based on the enhanced data set D 1, performing ten-fold cross validation on each machine learning algorithm according to the optimal super parameters in sequence to obtain ten-time average evaluation indexes R 2, RMSE and MAE, and selecting an optimal model according to the standards of the maximum R 2, the minimum RMSE and the minimum MAE;
Step S520, step S510 is sequentially executed on the enhanced data set D 2、D3、D4, and an optimal model corresponding to each data set is selected;
step S530, comparing the model performances under different enhancement data sets obtained in step S510 and step S520, and selecting an optimal enhancement data set and an optimal machine learning algorithm;
And S540, determining an optimal enhancement data set, determining an optimal friction sensitivity prediction algorithm as an rbf core SVR and an optimal impact sensitivity prediction algorithm as an ANN, and constructing a final rbf core SVR friction sensitivity prediction model and an ANN impact sensitivity prediction model by using optimal super parameters based on the optimal enhancement data set.
Compared with the prior art, the invention has the following advantages:
(1) The invention reasonably characterizes the mixed multicomponent system of the propellant, and solves the problems that the propellant has complex components, a plurality of influencing factors and cannot be represented by the traditional molecular structure property like the simple substance explosive molecules; aiming at the defect of insufficient sample size, an AL-Mixup data enhancement technology is provided for the first time, the technology skillfully combines the active learning technology and the Mixup mixed interpolation technology, the quality of a new sample generated by Mixup is monitored by utilizing the idea of active learning, meanwhile, the Mixup technology avoids experimental labeling required by active learning in the conventional sense, the advantages are complementary, and the resultant force realizes high-quality data enhancement; the application range is wide, the operation is simple, and the prediction is rapid and accurate.
(2) According to the invention, 4 external samples are used for testing the model, and the predicted values of friction sensitivity and impact sensitivity can be obtained rapidly only by inputting the component proportion and granularity information of the propellant, the predicted error is within 12%, and the prediction precision is high.
(3) The Mixup interpolation technology AL-Mixup based on active learning effectively solves the problem of small propellant sample size, can be used for constructing a machine learning model with strong robustness to rapidly realize the prediction of the propellant safety performance, avoids the limitation of time and labor waste in the traditional experiment measurement, and has theoretical guiding significance for the safe use and formula optimization of the propellant.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is an AL-Mixup data enhancement flow chart;
FIG. 3 is a graph showing a comparison of the prediction models R 2 of RDX propellants at different enhancement times, wherein A is a graph showing a comparison of the friction sensitivity prediction models; b is a comparison schematic diagram of an impact sensitivity prediction model;
FIG. 4 is a schematic diagram of the results of a ten fold cross-validation of an RDX-CMDB propellant predictive model, where A is a schematic diagram of the results of a ten fold cross-validation of a friction sensitivity predictive model and B is a schematic diagram of the results of a ten fold cross-validation of an impact sensitivity predictive model.
Detailed Description
The present invention will be described in further detail with reference to examples, but embodiments of the present invention are not limited thereto.
Example 1:
Referring to fig. 1, a method for predicting safety performance of an RDX modified double-based propellant includes:
Step S100, analyzing propellant sample data, characterizing the sample by using component proportion information and granularity information, and constructing an original data set D 0; the method specifically comprises the following steps:
Step S110, the components of different formulas of the propellant are combined to realize characteristic unification, and finally 13 components are contained, wherein each component corresponds to one characteristic, the component proportion is represented by mass percent, and if the original formula does not contain a certain component, the component proportion is represented by 0;
step S120, taking the logarithmic value of granularity as a characteristic value and adding the logarithmic value into a characteristic vector;
In step S130, each group of samples represents the safety performance by using the friction sensitivity and the impact sensitivity, each sample is characterized as a 14-dimensional feature vector, and contains 13-dimensional component proportion and 1-dimensional granularity information, and the original data set is marked as D 0 according to two label values of the friction sensitivity and the impact sensitivity. As shown in table 1.
Table 1 dataset
Step S200, referring to FIG. 2, performing 50 times, 100 times, 150 times and 200 times of amplification on the original data set by utilizing an AL-Mixup data enhancement technology to obtain corresponding enhanced data sets D 1、D2、D3 and D4; specifically:
The method of mixing differences is randomly operated 500 times Mixup, correspondingly 500 new data are generated, the sample pools U 0 and Mixup to be selected which are necessary for executing active learning are formed, and data enhancement is executed according to the following formula:
～Beta(α,α)
Wherein: x i and x j are two propellant samples randomly selected from the original dataset, and y i and y j are the labels corresponding to x i and x j. Is a new propellant sample generated after mixing by interpolation algorithm,/>Is/>A corresponding tag. Lambda obeys the Beta-distribution, alpha is an element of (0), ++ infinity A kind of electronic device.
Initial proxy models C f,0 and C i,0 are constructed by utilizing an original data set D 0, the safety performance value of a sample pool U 0 is predicted, the expected lifting value EI of each sample to be selected is calculated according to the following formula, the EI values of each sample to be selected are ordered in descending order, the sample of which the ranking is 50 is selected and added to the original data set D 0 to form an enhanced data set D 1, and at the moment, 450 samples to be selected remain in the sample pool U1.
z＝(μi-μ*)/σ
Wherein, sigma is the standard deviation,Is a probability density function, Φ (z) is a cumulative distribution function, and μ * is the maximum value of the security performance in the data set.
Retraining with the enhanced data set D 1 after 50 enhancements to obtain updated proxy models C f,1 and C i,1, predicting the sample pool U 1 using the proxy models C f,1 and C i,1 and calculating the expected enhancement value EI, After descending order, selecting the sample with the top ranking of 50, adding the sample into the data set D 1 to form an enhanced data set D 2, wherein 400 samples to be selected are left in the sample pool U 2; Retraining with the enhanced data set D 2 results in updated proxy models C f,2 and C i,2, and predicting the sample pool U 2 and calculating the desired elevation value EI, After descending order, selecting the sample with the top ranking of 50, adding the sample into the data set D 2 to form an enhanced data set D 3, wherein the sample pool U 3 still has 350 samples to be selected; retraining with the enhanced data set D 3 results in updated proxy models C f,3 and C i,3 and predicting the sample pool U 3 and calculating the desired elevation value EI, The top 50 samples are selected and added to the data set D 3 after descending order, so as to form an enhanced data set D 4, and 300 samples to be selected are left in the sample pool U 4. at this point, D 1、D2、D3、D4 is the new dataset enhanced 50 times, 100 times, 150 times, 200 times, respectively.
Step S300, performing Z-score standardization processing on the amplified data in the data set by adopting a standardization device to obtain standardized data;
Considering that the three components of nitrocellulose (N1), nitroglycerin (N2) and cyclotrimethylene trinitramine (R1) in the propellant formulation are more than 90% by mass, the components are the main components of the propellant, and the catalysts such as lead 2, 4-dihydroxybenzoate (BP) and carbon black (carbon) have low content but can have influence on the safety performance of the propellant. Therefore, the distribution range of the content of each component in the original sample data has larger difference, and if the original data is directly adopted for model training, the influence degree of the component with higher content on the model can be enhanced, and the effect of the component with low content is weakened or even ignored. In order to reduce the interference of the component content in different value scales to a large extent, balance the weight of each component on the influence of the safety performance, reduce the deviation of the prediction result caused by the excessive difference of the content of different components (characteristic variables), ensure the effectiveness of the model training fitting process, and perform Z-score standardization processing on the characteristic variables of the original sample data by means of a STANDARDSCALER module in python according to the following formula:
Wherein i is the ith feature variable; j is the j-th sample; x ij is the start value of the ith feature variable for the jth sample; x' ij is the normalized value of the ith feature variable for the jth sample; s i is the standard deviation of the ith feature variable. The characteristic variable after Z-score normalization treatment has the characteristic of standard normal distribution, the mean value is 0, and the variance is 1.
Step S400, performing modeling prediction on the standardized data by using a Machine learning algorithm, wherein the Machine learning algorithm is a multiple linear Regression (Multiple Linear Regression, MLR), a partial Least squares Regression (PARTIAL LEAST-square Regression, PLSR), a kernel-ridge Regression (KERNEL RIDGE Regression, KRR), a LASSO algorithm (LASSO solution SHRINKAGE AND Selection Operator, LASSO), a K nearest neighbor Regression (K-nearest Neighbors, KNN), a support vector Regression (Support Vector Regression, SVR), a Random Forest (RF), a limiting gradient lifting (eXtreme Gradient Boosting, XGB), a lightweight gradient lifting (LIGHT GRADIENT Boosting Machine, LGB), and an artificial neural network (ARTIFICIAL NEURAL NETWORK, ANN), respectively; parameter optimization of the four enhanced data sets by executing different machine learning algorithms specifically comprises:
The parameter optimization space for each machine learning algorithm was determined as shown in table 2:
table 2 super-parametric optimization space for different machine learning algorithms
Grid search is executed by combining ten-fold cross validation, R 2 values under each group of super parameters are calculated according to the following formula, and a group of super parameters corresponding to the maximum R 2 value are selected as optimal super parameters
Where n is the number of samples, y i is the experimental value,Is the average of experimental values,/>Is a model predictive value.
Comparing the performances of various machine learning algorithms, selecting the algorithm with the best effect to construct a final model, and specifically:
Comparing the effect of different enhancement times on model construction, the result is that the performance of the friction sensitivity and impact sensitivity prediction model is improved as the data enhancement times are increased as shown in fig. 3. Under the condition of 50 times of data enhancement, R 2 of most models is below 0.5, when the enhancement times are increased to 100 times and 150 times, the performances of the models are improved to different degrees, the maximum R 2 of the friction sensitivity prediction model can be above 0.75, and the maximum R 2 of the impact sensitivity prediction model can be above 0.8; when the number of enhancements reaches 200, the performance improvement of each model becomes slow, probably because too many data enhancements may burden the sample space with data redundancy when the original data sample is limited, and the distribution of samples may deviate from the actual situation of the original data, so 200 is the most suitable number of enhancements.
After determining the optimal number of enhancements, the performance of each machine learning algorithm is further compared. FIG. 4 shows the R 2, RMSE and MAE results of each machine learning algorithm cross-fold cross-validation under the RDX-CMDB propellant friction and impact sensitivity prediction model 200 times enhancement. From the results, it can be seen that the SVR model with "rbf" as the kernel function is the optimal RDX-CMDB propellant friction sensitivity prediction model, R 2, RMSE and MAE are 0.7950, 2.5077 and 1.4490, respectively; the ANN model is the optimal RDX-CMDB propellant crash sensitivity prediction model with R 2, RMSE, and MAE 0.8932, 0.1600, and 0.0888, respectively. The super parameters corresponding to the optimal model are shown in table 3.
The formulas for RMSE and MAE are as follows:
Where n is the number of samples, y i is the experimental value, Is a model predictive value.
TABLE 3 optimal superparameter settings and Performance for Security Performance prediction models
Step S500, the SVR model and the ANN model are verified by using 4 external samples, the relative prediction error is calculated according to the following formula, and the analysis result is shown in Table 4.
TABLE 4 RDX-external test results for CMDB propellant safety Performance prediction model
It can be seen that the relative error of the friction sensitivity and the impact sensitivity of the constructed prediction model to 4 external samples is within 12%, which proves good robustness and prediction capability of the model
And S600, storing the normalizer as a. Dat file by using a jackle. Dump method, storing the final model as a. Pkl file, and loading by using a jackle. Load method when predicting a new sample.
In summary, the invention realizes:
The AL-Mixup data enhancement technique is proposed to effectively amplify the propellant samples.
The propellant sensitivity measurement experiment difficulty is high, the cost is high, different experiment mode standards are not uniform, currently available propellant mechanical sensitivity data are few, machine learning is a science driven by data, the quality and the quantity of the data determine the extrapolation capability of a prediction model on an unknown sample to a great extent, and the development of the machine learning model is limited by the existing propellant safety performance data. According to the invention, mixup mixed interpolation technology is transferred from the image field to a propellant system, and is combined with the idea of active learning, and an adaptive AL-Mixup method is provided to realize higher-quality data enhancement. After the samples are processed by using the linear interpolation technology, the discrete samples can be more continuous in potential space, and meanwhile, the interpolated samples also have higher smoothness in the neighborhood, so that the defect of uneven distribution of the discrete samples is effectively overcome, and the robustness of the model to samples outside the distribution of the training samples is further improved. However, the method also has certain limitation, for example, the generated new data completely does not accord with the actual condition of the propellant formulation, so that the sample deviates from the actual and reasonable data distribution, the data enhancement at the moment is invalid, extra burden is possibly caused to the training of the model, the prediction performance of the model is reduced, and the problem can be solved just by introducing active learning. Active learning is one of typical small sample learning frameworks, and adopts an adaptive design strategy to combine the prediction result and uncertainty of a machine learning model through an optimization algorithm to calculate a ranking, so as to recommend the most valuable sample for the next test. The AL-Mixup method skillfully combines the active learning and Mixup mixed interpolation technology, monitors the quality of a new sample generated by Mixup by utilizing the idea of active learning, and simultaneously Mixup technology avoids experimental labeling required by active learning in a conventional sense, has complementary advantages and realizes high-quality data enhancement by resultant force.
And secondly, developing a reasonable propellant characterization mode, and constructing a safety performance (friction sensitivity and impact sensitivity) prediction model with excellent performance.
The propellant is a multi-component mixture composed of an oxidant, a combustible agent, a binder, a plasticizer and other functional aids, has complex components and cannot characterize a sample by using molecular structural properties like an elemental explosive. The most obvious characteristic of the composite propellant is the component proportion, and the type and content of the main body energetic material, the granularity of the sample and even the content of different additives can have great influence on the mechanical sensitivity of the propellant. The invention comprehensively analyzes the characteristics of the propellant sample, extracts the components, content and granularity information of the propellant and constructs the feature vector. Specifically, the formulation was represented by 13 components of N1, N2, R1, D2, C2, al2O3, BP, BC, pbO, D1, carbon, OC, and V, each component content was represented by mass percent, the particle size was represented by a logarithmic value of particle size, and thus each sample was characterized as a 14-dimensional feature vector.
The invention performs standardization treatment on the characteristic variable Z-score, so that the content of each component has the same weight influence on the model objective function, and the comparability between the characteristics is improved.
On the basis of AL-Mixup data enhancement and Z-score standardization processing, the optimal model of friction sensitivity and impact sensitivity is determined by utilizing grid searching and ten-fold cross validation, and the model has good robustness and predictive capability.
Although the application has been described herein with reference to the above-described illustrative embodiments thereof, the foregoing embodiments are merely preferred embodiments of the present application, and it should be understood that the embodiments of the present application are not limited to the above-described embodiments, and that numerous other modifications and embodiments can be devised by those skilled in the art that will fall within the scope and spirit of the principles of this disclosure.
Claims (5)
1. A method for predicting the safety performance of an RDX modified dual-based propellant, comprising:
Step S100, analyzing propellant sample data, characterizing the sample by using component proportion information and granularity information, and constructing an original data set D 0;
Step S200, amplifying the original data set D 0 for a plurality of times by adopting a data enhancement technology to obtain corresponding enhancement data sets D 1、D2、D3 and D 4;
step S300, performing Z-score standardization processing on the amplified data in the data set by adopting a standardization device to obtain standardized data;
step S400, modeling and predicting standardized data by using a machine learning algorithm, obtaining optimal super parameters of the machine learning algorithm by combining ten-fold cross verification with grid parameter searching, executing ten-fold cross verification by each machine learning algorithm under the optimal super parameters, selecting an optimal machine learning model and an optimal enhancement data set, and constructing an RBF core SVR model for RBF core RDX friction sensitivity prediction and an ANN model for impact sensitivity prediction based on the optimal enhancement data set and the optimal machine learning model;
Step S500, verifying the SVR model and the ANN model by using an external sample, and storing a normalizer, the SVR model and the ANN model if the prediction errors are within a set threshold;
And S600, respectively inputting the propellant component proportion and the granularity information into the SVR model and the ANN model to obtain predicted values of friction sensitivity and impact sensitivity.
2. The method for predicting the safety performance of an RDX modified dual-based propellant according to claim 1, wherein said step S100 comprises:
Step S110, the components of different formulas of the propellant are combined to realize characteristic unification, and finally 13 components are contained, wherein each component corresponds to one characteristic, the component proportion is represented by mass percent, and if the original formula does not contain a certain component, the component proportion is represented by 0;
step S120, taking the logarithmic value of granularity as a characteristic value and adding the logarithmic value into a characteristic vector;
In step S130, each group of samples represents the safety performance by using the friction sensitivity and the impact sensitivity, each sample is characterized as a 14-dimensional feature vector, and contains 13-dimensional component proportion and 1-dimensional granularity information, and the original data set is marked as D 0 according to two label values of the friction sensitivity and the impact sensitivity.
3. The method for predicting the safety performance of an RDX modified dual-based propellant according to claim 2, wherein said step S200 specifically comprises:
step S210, constructing a proxy model C f,0 and C i,0,Cf,0 as a friction sensitivity prediction model by using an original data set D 0, and C i,0 as an impact sensitivity prediction model;
step S220, randomly operating the Mixup mixed difference method for N times to generate N new data, so as to form a sample pool U 0 to be selected;
Step S230, predicting a sample pool U 0 to be selected by using C f,0 and C i,0, calculating an expected lifting value EI of each sample to be selected, sorting EI values of each sample to be selected in a descending order, adding samples M before the selection into an original dataset D 0 to form an enhanced dataset D 1, and taking N-M samples to be selected as sample pools U 1, wherein N and M are integers and M is smaller than N;
Step S240, retraining by using the enhanced dataset D 1 to obtain updated proxy models C f,1 and C i,1, predicting a sample pool U 1 by using C f,1 and C i,1, calculating an expected lifting value EI, sorting in descending order, selecting samples M before ranking, adding the samples M before ranking into the dataset D 1 to form the enhanced dataset D 2, and taking N-2*M samples to be selected as a sample pool U 2;
Step S250, retraining by using the enhancement data set D 2 to obtain updated proxy models C f,2 and C i,2, predicting a sample pool U 2, calculating an expected lifting value EI, sorting in a descending order, selecting samples M before ranking, adding the samples M to the enhancement data set D 2 to form an enhancement data set D 3, and taking N-3*M samples to be selected as the sample pool U 3;
Step S260, retraining with the enhanced data set D 3 to obtain updated proxy models C f,3 and C i,3, predicting the sample pool U 3, calculating the expected lifting value EI, sorting in descending order, selecting the samples M before ranking, adding them to the enhanced data set D 3 to form the enhanced data set D 4, and leaving N-4*M samples to be selected.
4. A method for predicting the safety performance of an RDX modified dual-based propellant as claimed in claim 3, wherein said step S400 comprises:
Step S410, determining parameter optimization spaces of ten machine learning algorithms; the machine learning algorithm is a multiple linear regression MLR, a partial least squares regression PLSR, a kernel ridge regression KRR, a LASSO algorithm LASSO, a K nearest neighbor regression KNN, a support vector regression SVR, a random forest RF, a limit gradient lifting XGB, a lightweight gradient lifting LGB and an artificial neural network ANN respectively;
Step S420, randomly dividing the D 1 reinforced data set into 10 parts according to friction/impact sensitivity value distribution, ensuring uniform distribution of each part of data, wherein 9 parts of data are used as training sets, and the rest 1 part of data are used as test sets;
Step S430, executing a parameter optimization process: for each algorithm, respectively:
For each group of super parameters, constructing a corresponding algorithm model on a training set, and obtaining an evaluation index R 2 on a testing set; cycling for 10 times to ensure that each data is used as a training set and a testing set, correspondingly, 10R 2 values are generated, and the average value is taken as the performance of the model under the super-parameters of the group;
After traversing each group of super parameters, obtaining R 2 values under all super parameter combinations, and selecting a group of super parameters corresponding to the maximum R 2 value as the optimal super parameters of a machine learning algorithm;
Step S440: step S420 and step S430 are sequentially performed on the enhanced data set D 2、D3、D4, so as to obtain the optimal super parameters of each machine learning algorithm under different data sets.
5. The method for predicting the safety performance of an RDX modified dual-based propellant as claimed in claim 4, wherein said step S500 comprises:
Step S510, based on the enhanced data set D 1, performing ten-fold cross validation on each machine learning algorithm according to the optimal super parameters in sequence to obtain ten-time average evaluation indexes R 2, RMSE and MAE, and selecting an optimal model according to the standards of the maximum R 2, the minimum RMSE and the minimum MAE;
Step S520, step S510 is sequentially executed on the enhanced data set D 2、D3、D4, and an optimal model corresponding to each data set is selected;
step S530, comparing the model performances under different enhancement data sets obtained in step S510 and step S520, and selecting an optimal enhancement data set and an optimal machine learning algorithm;
And S540, determining an optimal enhancement data set, determining an optimal friction sensitivity prediction algorithm as an rbf core SVR and an optimal impact sensitivity prediction algorithm as an ANN, and constructing a final rbf core SVR friction sensitivity prediction model and an ANN impact sensitivity prediction model by using optimal super parameters based on the optimal enhancement data set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311821308.8A CN117766058B (en) | 2023-12-27 | 2023-12-27 | Safety performance prediction method of RDX modified double-base propellant |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311821308.8A CN117766058B (en) | 2023-12-27 | 2023-12-27 | Safety performance prediction method of RDX modified double-base propellant |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117766058A CN117766058A (en) | 2024-03-26 |
| CN117766058B true CN117766058B (en) | 2024-06-21 |
Family
ID=90321816
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311821308.8A Active CN117766058B (en) | 2023-12-27 | 2023-12-27 | Safety performance prediction method of RDX modified double-base propellant |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117766058B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112382350A (en) * | 2020-11-20 | 2021-02-19 | 四川大学 | Machine learning estimation method for sensitivity, mechanical property and relation of energetic substances |
| CN116312890A (en) * | 2023-03-23 | 2023-06-23 | 西安邮电大学 | Method for screening high-hardness high-entropy alloy by aid of particle swarm optimization algorithm and machine learning |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2022063341A1 (en) * | 2020-09-27 | 2022-03-31 | 深圳太力生物技术有限责任公司 | Basal culture medium development method, basal culture medium formulation and development, and system thereof |
-
2023
- 2023-12-27 CN CN202311821308.8A patent/CN117766058B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112382350A (en) * | 2020-11-20 | 2021-02-19 | 四川大学 | Machine learning estimation method for sensitivity, mechanical property and relation of energetic substances |
| CN116312890A (en) * | 2023-03-23 | 2023-06-23 | 西安邮电大学 | Method for screening high-hardness high-entropy alloy by aid of particle swarm optimization algorithm and machine learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117766058A (en) | 2024-03-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Moein et al. | Predictive models for concrete properties using machine learning and deep learning approaches: A review | |
| Raychaudhuri et al. | Associating genes with gene ontology codes using a maximum entropy analysis of biomedical literature | |
| CN112382350B (en) | Machine learning estimation method for sensitivity and mechanical property of energetic substance and relation of energetic substance | |
| Mock et al. | Taxonomic classification of DNA sequences beyond sequence similarity using deep neural networks | |
| Nguyen et al. | Filter based backward elimination in wrapper based PSO for feature selection in classification | |
| Baghaei et al. | Durability assessment of FRP-to-concrete bonded connections under moisture condition using data-driven machine learning-based approaches | |
| CN113657661A (en) | Enterprise carbon emission prediction method and device, computer equipment and storage medium | |
| Zhu et al. | Single-cell clustering based on shared nearest neighbor and graph partitioning | |
| CN117766058B (en) | Safety performance prediction method of RDX modified double-base propellant | |
| Ouyang et al. | Predicting concrete’s strength by machine learning: Balance between accuracy and complexity of algorithms | |
| Bouras et al. | Prediction of high-temperature creep in concrete using supervised machine learning algorithms | |
| Zhu et al. | Scalable soft sensor for nonlinear industrial big data via bagging stochastic variational Gaussian processes | |
| Li et al. | PLPMpro: Enhancing promoter sequence prediction with prompt-learning based pre-trained language model | |
| Zhang et al. | Multi-fidelity surrogate model-assisted fatigue analysis of welded joints | |
| Mu et al. | iPseU-Layer: identifying RNA pseudouridine sites using layered ensemble model | |
| Yoo et al. | Discovery of gene-regulation pathways using local causal search. | |
| CN117079744A (en) | Artificial intelligent design method for energetic molecule | |
| Doan et al. | A hybrid model for predicting missile impact damages based on k-nearest neighbors and Bayesian optimization | |
| Khodaparasti et al. | Prediction of the concrete compressive strength using improved random forest algorithm | |
| Rao et al. | Motif discovery in tissue-specific regulatory sequences using directed information | |
| CN114741963A (en) | Lithium battery state-of-charge prediction method based on multi-scale attention mechanism | |
| Priyadharshini et al. | Materials property prediction using feature selection based machine learning technique | |
| Liu et al. | A machine learning method for buckling design of internally pressurized torispherical heads considering geometric imperfection | |
| Yin et al. | A new state‐of‐health estimation method for Li‐ion batteries based on interpretable belief rule base with expert knowledge credibility | |
| Müller et al. | Vector space classification of DNA sequences |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |