Composite hydrogen storage cylinder bursting pressure prediction method based on small sample machine learning

Abstract

The invention discloses a composite material hydrogen storage cylinder bursting pressure prediction method based on small sample machine learning, and relates to the technical field of hydrogen storage cylinder bursting pressure prediction and machine learning, original data of a composite material hydrogen storage cylinder bursting test are obtained, and the original data comprise feature data and corresponding real bursting pressure; preprocessing the original data to form a real database; training the generative adversarial network based on the real database, generating virtual data which is several times of the original data by using the trained generative adversarial network, and establishing a virtual database; performing XGBoost extreme gradient lifting training based on the virtual database to preliminarily obtain a bursting pressure prediction model; and correcting the preliminarily obtained bursting pressure prediction model in combination with a real database to obtain a final bursting pressure prediction model. According to the method, the integrated learning model is trained based on a small amount of bursting test data, and rapid and accurate prediction of the bursting pressure of the gas cylinder is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrogen storage cylinder bursting pressure prediction and machine learning, and in particular to a composite hydrogen storage cylinder bursting pressure prediction method based on small sample machine learning. Background Art

[0002] Composite hydrogen storage cylinders (composite hydrogen storage cylinders) as key equipment for gaseous hydrogen storage have been widely researched and developed in recent years. It consists of an inner liner, a valve seat, and a carbon fiber resin-based composite winding layer. The internal pressure load is mainly borne by the composite layer. The bursting pressure of composite hydrogen storage cylinders is a key indicator of structural design and manufacturing, and is crucial to ensuring the safety and reliability of the cylinder. However, the design and manufacturing process of hydrogen storage cylinders involves multiple links such as carbon fiber winding and resin curing, and there are many influencing factors, which makes it very difficult to accurately predict the bursting pressure.

[0003] At present, the prediction of bursting pressure of composite hydrogen storage cylinders mainly relies on finite element simulation technology. The pressure value corresponding to the structural failure is calculated through simplified numerical models and idealized assumptions, which often ignores many random factors in the actual cylinder manufacturing process. As a result, the predicted value obtained through time-consuming modeling and calculation is significantly different from the experimental value, and repeated experimental iterations are required to find the optimal design and process. Summary of the invention

[0004] In order to overcome the defects in the above-mentioned prior art, the present invention provides a method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning, and trains an integrated learning model based on a small amount of burst test data to achieve rapid and accurate prediction of the bursting pressure of the cylinder.

[0005] To achieve the above object, the present invention adopts the following technical solutions, including: The method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning includes the following steps: S1, obtaining the original data of the explosion test of the composite material hydrogen storage cylinder, the original data including characteristic data and the corresponding real explosion pressure; S2, preprocessing the original data to form a real database; S3, training a generative adversarial network based on a real database, using the trained generative adversarial network to generate virtual data several times the amount of original data, and establishing a virtual database; S4, based on the virtual database, XGBoost extreme gradient boosting training was performed to preliminarily obtain the bursting pressure prediction model; S5, correcting the initially obtained bursting pressure prediction model in combination with the real database to obtain a final bursting pressure prediction model.

[0006] Preferably, in step S4, Lasso and Ridge regularization are introduced, and K-fold cross validation technology is combined to perform XGBoost extreme gradient boosting training of the model.

[0007] Preferably, the specific method of step S4 is as follows: Construct an XGBoost extreme gradient boosting model. The model is composed of multiple decision trees. The number of decision trees is used as a hyperparameter. The optimal number of decision trees is determined through grid search. The optimal number of decision trees is obtained by minimizing the mean square error under K-fold cross validation. After determining the optimal number of decision trees, set the search space for other hyperparameters, use the Bayesian optimization method to optimize various hyperparameter combinations, and obtain the optimal hyperparameter combination based on minimizing the mean square error under K-fold cross validation; Based on the determined optimal hyperparameter combination, the training set in the virtual database is learned, and the objective function Obj of the model training is: ; ; ; in, loss is the loss function; y i For the virtual database i The bursting pressure value of each sample, For the i The predicted burst pressure of samples; i , n are the sample sequence number and the total number of samples in the training set respectively; is a regularization term, including Lasso regularization and Ridge regularization. Lasso regularization adds weights to the objective function. w j Absolute value penalty coefficient α Realize feature selection and sparsity; Ridge regularization through penalty coefficient l Limit the weight of the entire model w j Size; Hyperparameters c Used to control the strength of the regularization term.

[0008] Preferably, the data in the virtual database is randomly divided into K subsets, and for each iterative training and testing, one of the subsets is selected as the test set, and the other K-1 subsets are combined as the training set for K-fold cross-validation.

[0009] Preferably, an early stopping technique is introduced during the XGBoost extreme gradient boosting training process of the model, specifically: setting the initial value of the parameter T, if the performance of the trained model on the test set does not improve, then adding 1 to the value of T and continuing the model training; if the performance improves, then resetting T and updating the model parameters; when T reaches the preset value, stopping the model training.

[0010] Preferably, in step S5, a radial basis function (RBF) method is used to perform adaptive proportional factor adjustment, and the initially obtained bursting pressure prediction model is corrected in combination with a real database.

[0011] Preferably, the specific method of step S5 is as follows: Calculate each sample in the real database x RBF kernel function value: ; in, F is the RBF kernel function; c represents the center point in the real database, which is the mean value of the real database; s is the width parameter; ; in, R is the data range; ; in, R k For the k The data range of the feature, that is, k The difference between the maximum and minimum values ​​of a feature; m k For the k The weight of each feature is obtained by normalizing the importance value of each feature, i.e., the SHAP value, as the corresponding weight. k , M are the feature number and the total number of features respectively; The objective function is the mean square error MSE : ; in, y i’ The real database i’ The true value of bursting pressure of samples; For the i’ The predicted burst pressure of samples, N is the number of samples in the real database; β is the scale factor; The objective function is minimized through the optimization algorithm to obtain the optimized proportional factor β, used to correct the preliminary burst pressure prediction model: ; in, is the bursting pressure prediction value of the bursting pressure prediction model obtained initially; is the corrected bursting pressure prediction value, that is, the bursting pressure prediction value of the final bursting pressure prediction model.

[0012] Preferably, in step S5, the SHAP value of each feature is calculated using the SHAP library in the Python platform.

[0013] Preferably, in step S1, the characteristic data of the composite hydrogen storage cylinder include one or more of the head ellipsoid ratio, winding tension, winding pressure, inner liner thickness, inner liner aspect ratio, maximum curing temperature, ratio of the thickness of the spiral winding layer to the annular winding layer, maximum winding angle of the spiral winding, total winding time, and fiber volume fraction of the composite layer.

[0014] Preferably, in step S2, robust normalization processing is performed on the feature data in the original data: ; Among them, X norm is the feature data after robust normalization; median( X ) is the characteristic data X median; IQR( X ) is the characteristic data X The interquartile range.

[0015] The advantages of the present invention are: (1) Considering that the explosion test of hydrogen storage cylinders is time-consuming, labor-intensive and expensive, the explosion pressure of the designed cylinder can be predicted efficiently and accurately by analyzing a small amount of test data through integrated machine learning methods, which accelerates the explosion pressure prediction and design iteration process of composite hydrogen storage cylinders. The method of the present invention trains the integrated learning model based on a small amount of explosion test data to achieve rapid and accurate prediction of the explosion pressure of the cylinder.

[0016] (2) Based on a small amount of composite hydrogen storage cylinder design and manufacturing related data and explosion test data, the association hidden behind the data is established to fully utilize the existing test data and quickly predict the explosion pressure, avoiding complex and distorted numerical simulation work.

[0017] (3) The key factors and data in the entire design and manufacturing process of the hydrogen storage cylinder are recorded and incorporated into the database of the machine learning model to achieve accurate prediction of the bursting pressure.

[0018] (4) By integrating machine learning and cross-validation methods, the robustness and generalization ability of the model based on small sample data can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning.

[0020] Figure 2 Comparison chart of SHAP values ​​for each feature. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] Depend on Figure 1 As shown, the method for predicting the bursting pressure of a composite hydrogen storage cylinder based on small sample machine learning of the present invention comprises the following steps: S1. Collect relevant data on the design and manufacturing of the liner of the composite hydrogen storage cylinder, the fiber winding curing process, and the cylinder burst test, namely the original data, including characteristic data and the corresponding actual burst pressure. Characteristic factors are factors that affect the burst strength of the cylinder, such as geometric characteristics, material parameters, layering scheme, curing system, etc.

[0023] S2. Normalize and preprocess the original data and establish a real database .

[0024] S3, based on real database Train Generative Adversarial Networks (GANs) to learn from real databases The spatial distribution characteristics of the original data are then used to generate new virtual data that is several times the amount of the original data using the trained GANs to establish a virtual database , and have similar data distribution characteristics and statistical properties, in particular, The virtual data in the model can have better diversity, making up for the problem of too little single feature information in the experimental data, thereby improving the effect of model training and prediction.

[0025] S4, based on virtual database XGBoost extreme gradient boosting training was performed, Lasso and Ridge regularization and early stopping techniques were introduced, and K-fold cross-validation technology was combined to effectively prevent model overfitting and improve model generalization ability, and a preliminary burst pressure prediction model for composite hydrogen storage cylinders was obtained. .

[0026] S5. Use radial basis function RBF method to adjust the adaptive scale factor, combined with real database The initial burst pressure prediction model After correction, the final burst pressure prediction model was obtained. , It has better prediction accuracy and generalization ability, and can quickly give accurate prediction values ​​of bursting pressure for parameter combinations within the design space of new composite hydrogen storage cylinders.

[0027] Example 1 The burst pressure prediction method of composite hydrogen storage cylinder based on small sample machine learning is as follows: S11, obtaining the original data of the explosion test of the composite hydrogen storage cylinder, the original data including characteristic data and the corresponding real explosion pressure. The characteristic data includes the head ellipsoid ratio k, winding tension F, winding pressure P, liner thickness h, liner aspect ratio L / D, maximum curing temperature T, spiral winding layer and circumferential winding layer thickness ratio n, spiral winding maximum winding angle θ m , total winding time t, fiber volume fraction V of the composite material layer f The characteristic data is obtained by monitoring the winding and curing process of the hydrogen storage cylinder. The corresponding label is the bursting pressure value of each composite hydrogen storage cylinder measured by the water pressure bursting test, that is, the actual bursting pressure. A total of 50 sample data were obtained.

[0028] S12, preprocess the raw data and establish a real database of small samples.

[0029] The feature data is normalized to eliminate the impact of dimension and improve the convergence speed of the algorithm. In order to ensure the robustness to outliers in the data, the feature data X Perform robust normalization: ; Among them, X norm is the normalized feature data; median( X ) is the characteristic data X median; IQR( X ) is the characteristic data XThe interquartile range refers to the difference between the third quartile and the first quartile after the data set is divided into four equal parts. It reflects the degree of dispersion of the middle 50% of the data set and is not sensitive to extreme values.

[0030] S13, based on the real database, training the generative adversarial network GANs, learning the spatial distribution characteristics of the original data in the real database, and then using the trained GANs to generate new virtual data that is several times the amount of original data to establish a virtual database.

[0031] Specifically, the number of virtual data generated by GANs is 10 times the number of original data, totaling 500 sample data, constituting a virtual database.

[0032] The data in the virtual database is randomly divided into K subsets. For each iteration of training and validation, one of the subsets is selected as the test set, and the other K-1 subsets are combined as the training set for subsequent K-fold cross-validation. Here K is 10. During the cross-validation process, the data set is divided into 10 equal parts, each including 50 samples. 9 of them are selected as the training set, and the remaining 1 part is used as the test set to evaluate the performance of the model.

[0033] S14, based on the virtual database, XGBoost extreme gradient boosting training was performed, Lasso and Ridge regularization and early stopping techniques were introduced, and K-fold cross-validation technology was combined to effectively prevent model overfitting and improve the model generalization ability. A preliminary burst pressure prediction model for composite hydrogen storage cylinders was obtained, which was recorded as the preliminary model.

[0034] Construct an XGBoost extreme gradient boosting model, which is specifically composed of multiple decision trees. The number of decision trees is a key hyperparameter. The optimal number of decision trees can be determined by grid search or other methods, or it can be determined in combination with the size of the data set to obtain the best model complexity and generalization ability. Here, we choose to use grid search to search the number of decision trees in the range of 50 to 400 with a step size of 50. The optimal number of decision trees obtained by minimizing the mean square error MSE under 10-fold cross validation is 100.

[0035] After determining the number of decision trees, set other hyperparameter search spaces, including the maximum depth of the tree (3-10), learning rate (0.01-0.3), minimum loss reduction (0-5), minimum weight of child nodes (1-10), subsampling ratio (0.5-1), column sampling ratio of each tree (0.5-1), etc. Use the Bayesian optimization tool on the Python platform to train models for various hyperparameter combinations, and determine the optimal hyperparameter combination based on minimizing the mean square error under 10-fold cross validation.

[0036] Use the hyperparameters determined above to carry out data learning on the training set in the virtual database, and the objective function Obj is: ; ; ; in, loss is the loss function. For regression problems, the mean square error is used here to measure the gap between the model prediction value and the actual value; y i For the virtual database i The bursting pressure value of each sample, For the i The predicted burst pressure of samples; i , n are the sample number and the total number of samples respectively. It is a regularization term, which includes Lasso regularization and Ridge regularization. Lasso regularization adds weights to the objective function. w j Absolute value penalty coefficient α Realize feature selection and sparsity to reduce the weight of some unimportant features; Ridge regularization uses the penalty coefficient l Limit the weight of the entire model w j Size, to prevent the model from being overly dependent on individual features; hyperparameters c Used to control the strength of the regularization term.

[0037] Because the amount of training data for small sample machine learning is small, it cannot fully represent the data distribution of the entire design range of composite hydrogen storage cylinders. The actual cylinder data set may contain some features that have little effect on the bursting pressure. In addition, the model is repeatedly iteratively trained on the small sample data set and learns the noise in the data. Many factors may cause the model to overfit and have poor generalization ability. That is, the prediction on the small sample data set is very accurate, but it cannot accurately predict the bursting pressure of the newly designed cylinder. Regularization term The introduction of controls the complexity of the model, prevents the model from overfitting, and helps the model to accurately predict the bursting pressure of newly designed gas cylinders.

[0038] The XGBoost extreme gradient boosting model is trained for regression through a virtual database to obtain a preliminary burst pressure prediction model (preliminary model). Two types of regularization and K-fold cross validation are used to effectively reduce the risk of model overfitting. In K-fold cross validation, the average mean square error of K-times validation is used to evaluate the overall performance of the model. The early stopping technique is used to further prevent the model from overfitting. The initial value of the parameter T (number of iterations) is set to 0. If the performance of the model training on the test set does not improve, T+1 is set and the model training continues; if the performance improves, T is reset to 0 and the model parameters are updated; when T reaches the preset value (such as 10), the training is stopped. Save the trained preliminary model.

[0039] Through the mean square error MSE, root mean square error RMSE, mean absolute error MAE and R 2 Four indicators are used to evaluate the prediction accuracy of the preliminary model from different perspectives. After training XGBoost extreme gradient boosting on a virtual database, the MSE and RMSE of the preliminary model are less than 0.01, the MAE is less than 0.03, and the R 2 When it is greater than 0.85, the model prediction accuracy is high.

[0040] The SHAP library is used in the Python platform to calculate the SHAP value (importance value) of each feature. The SHAP value is used to quantify the contribution of each feature to the model prediction and visualized through SHAP summary diagrams, such as Figure 2 As shown in the figure, it is found that the influence degree on the bursting pressure of composite materials is V from high to low. f ,n,F,θ m , k, T, P, t, h and L / D, where V f , n, and F have a greater impact on the bursting pressure, while h and L / D have almost no effect on the bursting pressure.

[0041] S15, using radial basis function RBF method to adjust the adaptive scale factor, input sample x From the real database, calculate the RBF kernel function value of each sample F ( x ): ; in, F is the RBF kernel function; c represents the center point in the real database, which is the mean value of the real database; s is the width parameter; ; in, R The data range is calculated by normalizing the SHAP value of each feature as the weight: ; in, R k For the k The data range of the feature, that is, k The difference between the maximum and minimum values ​​of a feature; m k For the k The weight of each feature, which is normalized by the feature SHAP value; k , M are the feature number and the total number of features respectively.

[0042] The objective function is the mean square error MSE : ; in, N is the number of samples in the real database; β is the scale factor; y i’ The real database i’ The true value of bursting pressure of samples; For the i’ The predicted burst pressure of samples.

[0043] The objective function is minimized through optimization algorithms such as gradient descent to obtain the optimized proportional factor β The burst pressure predictions used to calibrate the preliminary model are: ; in, is the burst pressure prediction value of the preliminary model; is the corrected burst pressure prediction value, that is, the burst pressure prediction value of the final burst pressure prediction model; The initial burst pressure prediction model was corrected using a real database to obtain the final burst pressure prediction model, which has better prediction accuracy and generalization ability, and can quickly give accurate prediction values ​​of burst pressure for parameter combinations in the design space of new composite hydrogen storage cylinders. As the composite hydrogen storage cylinder test iterative process, new test data is added to the model for dynamic update to further improve the model prediction performance.

[0044] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning, characterized in that: The following steps are involved: S1, obtaining the original data of the explosion test of the composite material hydrogen storage cylinder, the original data including characteristic data and the corresponding real explosion pressure; S2, preprocessing the original data to form a real database; S3, training a generative adversarial network based on a real database, using the trained generative adversarial network to generate virtual data several times the amount of original data, and establishing a virtual database; S4, based on the virtual database, XGBoost extreme gradient boosting training was performed to preliminarily obtain the bursting pressure prediction model; S5, correcting the initially obtained bursting pressure prediction model in combination with the real database to obtain a final bursting pressure prediction model.

2. The method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning according to claim 1 is characterized in that: In step S4, Lasso and Ridge regularization are introduced, and K-fold cross-validation technology is combined to perform XGBoost extreme gradient boosting training of the model.

3. The method for predicting the bursting pressure of a composite hydrogen storage cylinder based on small sample machine learning according to claim 2 is characterized in that: The specific method of step S4 is as follows: Construct an XGBoost extreme gradient boosting model. The model is composed of multiple decision trees. The number of decision trees is used as a hyperparameter. The optimal number of decision trees is determined through grid search. The optimal number of decision trees is obtained by minimizing the mean square error under K-fold cross validation. After determining the optimal number of decision trees, set the search space for other hyperparameters, use the Bayesian optimization method to optimize various hyperparameter combinations, and obtain the optimal hyperparameter combination based on minimizing the mean square error under K-fold cross validation; Based on the determined optimal hyperparameter combination, the training set in the virtual database is learned, and the objective function Obj of the model training is: ； ； ； in, loss is the loss function; y i For the virtual database i The burst pressure value of each sample, For the i The predicted burst pressure of samples; i , n are the sample sequence number and the total number of samples in the training set respectively; is a regularization term, including Lasso regularization and Ridge regularization. Lasso regularization adds weights to the objective function. w j Absolute value penalty coefficient α Realize feature selection and sparsity; Ridge regularization through penalty coefficient λ Limit the weight of the entire model w j Size; Hyperparameters γ Used to control the strength of the regularization term.

4. The method for predicting the bursting pressure of a composite hydrogen storage cylinder based on small sample machine learning according to claim 2 or 3, characterized in that: The data in the virtual database is randomly divided into K subsets. For each iteration of training and testing, one of the subsets is selected as the test set, and the other K-1 subsets are combined as the training set for K-fold cross-validation.

5. The method for predicting the bursting pressure of a composite hydrogen storage cylinder based on small sample machine learning according to claim 2 or 3, characterized in that: During the XGBoost extreme gradient boosting training process of the model, the early stopping technology is also introduced. Specifically, the initial value of the parameter T is set. If the performance of the trained model on the test set does not improve, the value of T is increased by 1 and the model training continues; if the performance improves, T is reset and the model parameters are updated; when T reaches the preset value, the model training is stopped.

6. The method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning according to claim 1 is characterized in that: In step S5, the radial basis function (RBF) method is used to perform adaptive proportional factor adjustment, and the initially obtained bursting pressure prediction model is corrected in combination with the real database.

7. The method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning according to claim 6 is characterized in that: The specific method of step S5 is as follows: Calculate each sample in the real database x RBF kernel function value: ； in, Φ is the RBF kernel function; c represents the center point in the real database, which is the mean value of the real database; σ is the width parameter; ； in, R is the data range; ； in, R k For the k The data range of the feature, that is, k The difference between the maximum and minimum values ​​of a feature; m k For the k The weight of each feature is obtained by normalizing the importance value of each feature, i.e., the SHAP value, as the corresponding weight. k , M are the feature number and the total number of features respectively; The objective function is the mean square error MSE : ； in, y i’ The real database i’ The true value of bursting pressure of samples; For the i’ The predicted burst pressure of samples, N is the number of samples in the real database; β is the scale factor; The objective function is minimized through the optimization algorithm to obtain the optimized proportional factor β , used to correct the preliminary burst pressure prediction model: ； in, is the bursting pressure prediction value of the bursting pressure prediction model obtained initially; is the corrected bursting pressure prediction value, that is, the bursting pressure prediction value of the final bursting pressure prediction model.

8. The method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning according to claim 7 is characterized in that: In step S5, the SHAP value of each feature is calculated using the SHAP library in the Python platform.

9. The method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning according to claim 1 is characterized in that: In step S1, the characteristic data of the composite hydrogen storage cylinder include one or more of the head ellipsoid ratio, winding tension, winding pressure, inner liner thickness, inner liner aspect ratio, maximum curing temperature, ratio of the thickness of the spiral winding layer to the circumferential winding layer, maximum winding angle of the spiral winding, total winding time, and fiber volume fraction of the composite material layer.

10. The method for predicting the bursting pressure of composite hydrogen storage cylinders based on small sample machine learning according to claim 1 is characterized in that: In step S2, robust normalization is performed on the feature data in the original data: ； Among them, X norm is the feature data after robust normalization; median( X ) is the characteristic data X median; IQR( X ) is the characteristic data X The interquartile range.

Images