A method and device for intelligently predicting structural strength and stiffness characteristics of a reduced-order model

Abstract

The present application belongs to the field of aircraft design technology, and particularly relates to a method and device for intelligently predicting the structural strength and stiffness characteristics of a reduced-order model. The method comprises: step S1, obtaining a first prediction model that has been trained for intelligently predicting the structural strength and stiffness characteristics; step S2, changing the parameter value of each input variable of the first prediction model, and determining the contribution rate of each input variable according to the output value of the first prediction model; step S3, selecting multiple input variables as reduced-order input variables based on the contribution rate; step S4, training multiple second prediction models with reduced-order input variables, calculating the prediction accuracy of each second prediction model with root mean square error or regression coefficient, and selecting the optimal second prediction model; step S5, predicting the strength and stiffness characteristics of the aircraft structure through the optimal second prediction model. The present application can improve aircraft safety and design efficiency.

Description

Technical Field

[0001] The present application belongs to the technical field of aircraft design, and in particular relates to a method and device for intelligently predicting structural strength and stiffness characteristics of a reduced-order model. Background Art

[0002] The two important goals of current aircraft design are rapid development and lightweight design. Traditional strength / stiffness simulation using CAE finite element analysis can achieve relatively accurate performance predictions through meshing, material settings, working condition simulation, etc. However, the process of CAE finite element analysis is relatively complicated, and iterative calculations take a long time, resulting in low solution efficiency and poor prediction real-time performance. Summary of the invention

[0003] In order to solve the above problems, the present application provides a method and device for intelligently predicting the structural strength and stiffness characteristics of a reduced-order model, so as to realize rapid prediction of the structural strength and stiffness response by constructing a high-dimensional parameter reduced-order model.

[0004] The first aspect of the present application provides an intelligent prediction method for structural strength and stiffness characteristics of a reduced-order model, which mainly includes:

[0005] Step S1, obtaining a first prediction model that has been trained for intelligent prediction of structural strength and stiffness characteristics;

[0006] Step S2: changing the parameter value of each input variable of the first prediction model respectively, so as to determine the contribution rate of each input variable based on the change of the output value of the first prediction model;

[0007] Step S3, ranking the contribution rates from large to small, and selecting a plurality of input variables with top rankings as reduced-order input variables;

[0008] Step S4: using the reduced-order input variables as input variables of multiple second prediction models, training the multiple second prediction models. During the training process, the prediction accuracy of each second prediction model is calculated using the root mean square error or the regression coefficient to select the optimal second prediction model.

[0009] Step S5: predict the strength and stiffness characteristics of the aircraft structure using the selected optimal second prediction model.

[0010] Preferably, step S2 further comprises performing the following steps on each input variable in the input variable combination of the first prediction model:

[0011] Step S21, respectively increasing or decreasing the input variables by a set percentage, and combining them with other unchanged input variables to form two new input variable combinations, and respectively calculating the two new input variable combinations through the first prediction model to obtain an incremental output combination and a decremental output combination;

[0012] Step S22, respectively calculating the normalized mean square error of the incremental output combination and the decremental output combination relative to the actual value;

[0013] Step S23, calculating the influence value of the input variable according to the regularized mean square error of the incremental output combination and the regularized mean square error of the decremental output combination;

[0014] Step S24: normalize the influence value to form the contribution rate of each input variable.

[0015] Preferably, the set percentage is 8%.

[0016] Preferably, step S3 further comprises:

[0017] Select a specified number or percentage of the top ranked input variables as reduced-order input variables; or

[0018] The contribution rate is accumulated from front to back until the contribution rate exceeds a set value, and each input variable involved in the accumulation is selected as a reduced-order input variable.

[0019] Preferably, in step S4, the plurality of second prediction models include a Dense model, a SVM model, a KNN model, a RF model and a GBDT model.

[0020] The second aspect of the present application provides an intelligent prediction device for structural strength and stiffness characteristics of a reduced-order model, which mainly includes:

[0021] A first prediction model acquisition module, used to acquire a trained first prediction model for intelligent prediction of structural strength and stiffness characteristics;

[0022] An input variable contribution rate calculation module, used to change the parameter value of each input variable of the first prediction model respectively, so as to determine the contribution rate of each input variable based on the change of the output value of the first prediction model;

[0023] A reduced-order input variable selection module is used to rank the contribution rates from large to small, and select a plurality of input variables with the highest ranking as reduced-order input variables;

[0024] A second prediction model selection module, used to train the plurality of second prediction models by using the reduced-order input variables as input variables of the plurality of second prediction models, and during the training process, calculating the prediction accuracy of each second prediction model by using the root mean square error or the regression coefficient to select the optimal second prediction model;

[0025] The strength and stiffness characteristics prediction module is used to predict the strength and stiffness characteristics of the aircraft structure by using the selected optimal second prediction model.

[0026] Preferably, the input variable contribution rate calculation module includes:

[0027] An input variable modification unit, used to increase or decrease the input variable by a set percentage, and form two new input variable combinations with other unchanged input variables, and respectively obtain an incremental output combination and a decremental output combination through calculation by the first prediction model;

[0028] A regularized mean square error calculation unit, used to calculate the regularized mean square error of the incremental output combination and the decremental output combination relative to the actual value;

[0029] An influence value calculation unit, used for calculating the influence value of the input variable according to the regularized mean square error of the incremental output combination and the regularized mean square error of the decremental output combination;

[0030] The contribution rate calculation unit is used to normalize the influence value to form the contribution rate of each input variable.

[0031] Preferably, the set percentage is 8%.

[0032] Preferably, the reduced-order input variable selection module includes:

[0033] Select by quantity unit, which is used to select the top-ranked input variables of a specified number or a specified percentage as reduced-order input variables;

[0034] The cumulative amount selection unit is used to accumulate the contribution rate from front to back until the contribution rate exceeds a set value, and each input variable involved in the accumulation is selected as a reduced-order input variable.

[0035] Preferably, the plurality of second prediction models include a Dense model, a SVM model, a KNN model, a RF model and a GBDT model.

[0036] By adopting the reduced-order model technology, this application can screen the characteristic parameters that affect the strength and stiffness of the aircraft, accurately predict the strength and stiffness response of the aircraft under specific environmental conditions, thereby quickly evaluating the strength and stiffness of the aircraft in real time, guiding the design and use process of the aircraft, and improving the safety and design efficiency of the aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of a preferred embodiment of the intelligent prediction method of structural strength and stiffness characteristics of the reduced-order model of the present application. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical scheme and advantages of the implementation of this application clearer, the technical scheme in the implementation of this application will be described in more detail in combination with the drawings in the implementation of this application. In the drawings, the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The described implementation is a part of the implementation of this application, not all of the implementations. The implementation described below with reference to the drawings is exemplary and is intended to be used to explain this application, and cannot be understood as a limitation on this application. Based on the implementation in this application, all other implementations obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The implementation of this application is described in detail below in combination with the drawings.

[0039] The first aspect of the present application provides an intelligent prediction method for structural strength and stiffness characteristics of a reduced-order model, such as Figure 1 As shown, it mainly includes:

[0040] Step S1, obtaining a first prediction model that has been trained for intelligent prediction of structural strength and stiffness characteristics;

[0041] Step S2: changing the parameter value of each input variable of the first prediction model respectively, so as to determine the contribution rate of each input variable based on the change of the output value of the first prediction model;

[0042] Step S3, ranking the contribution rates from large to small, and selecting a plurality of input variables with top rankings as reduced-order input variables;

[0043] Step S4: using the reduced-order input variables as input variables of multiple second prediction models, training the multiple second prediction models. During the training process, the prediction accuracy of each second prediction model is calculated using the root mean square error or the regression coefficient to select the optimal second prediction model.

[0044] Step S5: predict the strength and stiffness characteristics of the aircraft structure using the selected optimal second prediction model.

[0045] In response to the increasingly complex working condition simulation requirements, the prediction of structural strength and stiffness characteristics driven by reduced-order models will be an important step in aircraft design and development. The core is to achieve rapid prediction of structural strength and stiffness response by constructing a high-dimensional parameter reduced-order model. The high-dimensional parameter reduced-order model is to construct a simplified model based on a model with reasonable accuracy to approximate the original model. The core idea is to transform a large-scale complex system model into a smaller-scale approximate model under certain conditions, thereby improving the solution efficiency while meeting the prediction accuracy. Constructing a reduced-order model based on CAE finite element data or experimental data to fit the stress-strain response surface can make rapid predictions while ensuring accuracy.

[0046] Based on the above ideas, the intelligent prediction method for structural strength and stiffness characteristics of the reduced-order model provided in this application mainly reduces the order of the intelligent prediction model of structural strength and stiffness characteristics from two aspects. One is to reduce the input variables through steps S2 and S3, and the other is to select the optimal prediction model through step S4.

[0047] The aircraft body will obtain a large number of flight parameters during the mission, and the relationship between some flight parameters and the aircraft strength / stiffness response is weak. If all flight parameters are used as input parameters of the proxy model, the prediction accuracy of the proxy model will be reduced, affecting the actual use effect. Therefore, a data selection model based on an intelligent algorithm model can be used to screen the input data of the load prediction model, effectively identify the importance of various data, and realize the compression of input data.

[0048] Specifically, in step S2, input variables are selected according to sensitivity analysis, and the first prediction model can be described as: having X1, X2...X m There are m input characteristic condition quantities, i.e., multiple input variables, such as total mass, motion, altitude, Mach number, normal overload, pitch angle acceleration and other working condition parameters related to flight attitude, or design parameters such as strain gauge number and node number related to predicted position; model output Y1, Y2…Y r There are r decision quantities in total, that is, multiple output variables, such as the maximum, median, minimum or main direction values ​​of stress and strain in each direction such as E11 and S33. The purpose of this application is to use the least input variables to replace the original m input variables, while maintaining the regression or analysis ability of the original conditional feature set for the decision attributes. To this end, this application uses a genetic algorithm based on ELM to select and study the data, which can realize the sorting of the influence values ​​of different parameters. The size of the influence value reflects the importance of the parameter to the state parameter. The larger the influence value, the more important it is. That is to say, the first prediction model obtained in step S1 of this application is the ELM prediction model, and then in step S2, the input variables are continuously changed to check the degree of change of the output variables to determine which input variables are important and which input variables are not important.

[0049] In some optional implementations, step S2 further includes performing the following steps on each input variable in the input variable combination of the first prediction model:

[0050] Step S21, respectively increasing or decreasing the input variables by a set percentage, and combining them with other unchanged input variables to form two new input variable combinations, and respectively calculating the two new input variable combinations through the first prediction model to obtain an incremental output combination and a decremental output combination;

[0051] Step S22, respectively calculating the normalized mean square error of the incremental output combination and the decremental output combination relative to the actual value;

[0052] Step S23, calculating the influence value of the input variable according to the regularized mean square error of the incremental output combination and the regularized mean square error of the decremental output combination;

[0053] Step S24: normalize the influence value to form the contribution rate of each input variable.

[0054] In some optional embodiments, the set percentage is 8%.

[0055] In the above embodiment, in step S21, after a certain input variable is increased by 8%, the new input variable combination X in1ip For: X in1ip =(X1, X2…X i *1.08…X m ), after reducing the input variable by 8%, the new input variable combination X in1id For: X in1id =(X1, X2…X i *0.92…X m ), after calculation by the first prediction model, the incremental output combination and the decremental output combination are obtained respectively. Then in step S2, the corresponding regularized mean square error NMSE is calculated for the incremental output combination ip , calculate the corresponding regularized mean square error NMSE for the reduced output combination id , the formulas involved include:

[0056] ;

[0057] The above formula is a general formula, where N in the formula is the number of predicted values, which is consistent with the number r of output variables in the above embodiment. i represents the actual value of the i-th output variable, represents the predicted value of the i-th output variable given by the first prediction model, represents the mean of the actual values ​​of all output variables, As an intermediate variable, the above formula can calculate the normalized mean square error NMSE of the incremental output combination ip , Regularized mean square error NMSE of the reduced output combination id , and the regularized mean square error NMSE of the original combination i , corresponding to the predicted value of the output variable in the formula Parameters in the incremental output combination, decremental output combination, and original output combination should also be used instead.

[0058] Then, in step S23, the influence value of each input variable is calculated by the following formula:

[0059] ;

[0060] Among them, IV i is the influence value of the i-th input variable.

[0061] Finally, in step S24, the contribution rate of each input variable to the output variable is calculated by the following formula:

[0062] ;

[0063] Among them, CON i is the contribution rate of the ith input variable.

[0064] After calculating the contribution rate of each input variable as above, some input variables can be selected as reduced-order input variables in step S3 to participate in subsequent model prediction.

[0065] In some optional implementations, step S3 further includes:

[0066] Select a specified number or a specified percentage of input variables that are ranked high as reduced-order input variables; or accumulate the contribution rates from front to back until the contribution rates exceed a set value, and select each input variable involved in the accumulation as a reduced-order input variable.

[0067] In this embodiment, it is possible to directly specify how many input variables are selected as reduced-order input variables, for example, 20 are selected, or 80% of the total number of input variables are selected as reduced-order input variables. In addition, reduced-order input variables can also be selected by accumulation. For example, after the above-mentioned normalization calculation, the sum of the contribution rates of all input variables is 100%. The greater the contribution rate value, the higher the importance of the corresponding parameter. Therefore, after sorting the contribution rates from large to small, the cumulative sum is performed, and those input variables whose sum of contribution rates exceeds 90 are selected as reduced-order input variables. Generally speaking, the sum of the contribution rates of the top 20 input variables with the largest contribution rates will be greater than 90. This method can greatly reduce the number of input variables.

[0068] Step S4 is used to select the best second prediction model. In some optional implementations, in step S4, the plurality of second prediction models include a Dense model, a SVM model, a KNN model, a RF model and a GBDT model.

[0069] It is understandable that different reduced-order models all have hyperparameters that have a significant impact on the model prediction effect. The hyperparameters can be optimized using a grid search method to traverse possible parameter combinations and save the optimal combination.

[0070] The optimal hyperparameters for several typical models are as follows:

[0071] Dense model: The number of hidden layers is the number of middle layers, which is related to the amount of data. For general data sets, the number of hidden layers is 1-3. The number of nodes in each hidden layer is the number of neurons in each layer, which is related to the number of input and output features. The activation function is related to the nonlinear mapping capability and gradient back propagation.

[0072] SVM model: Its penalty coefficient C is related to the relaxation coefficient. When the penalty coefficient C is large, it indicates that the SVM model has a large error penalty for the training process, which may lead to overfitting; when the penalty coefficient C is small, it indicates that the SVM model has a small error penalty for the training process, which may lead to underfitting. Therefore, it is necessary to adjust it according to the training set and validation set data until the optimal value is found. The radial basis parameter gamma is only related to the radial basis kernel function RBF, which is used to characterize the influence of a single sample on the hyperplane; the polynomial parameter degree is only related to the polynomial kernel function Poly, which is used to characterize the degree of the polynomial; when the kernel function type is other, this parameter can be ignored.

[0073] KNN model: its weight is related to the weighting method of neighbor samples, uniform means that all neighbor samples are given the same weight, distance means that the weight is inversely proportional to the sample distance of the neighbor samples; the number of approximate samples n_neighbors means the value of K in K neighbor samples; the distance function p means the order of Minkowski distance.

[0074] RF model: tree n represents the number of trees in the forest model. More subtrees can make the model have better performance, but will reduce the computational efficiency; the number of features max_features represents the maximum number of features allowed for a single decision tree; the maximum depth max_depth represents the depth of the subtree.

[0075] GBDT model: The number of base learners n and the maximum depth max_depth have the same physical meanings as those in the RF model; the subsampling ratio subsample represents the sample fraction of each base learner.

[0076] The prediction model of this application is mainly for the prediction of stress and strain, which is a typical regression prediction. To this end, in step S4, RMSE (root mean square error), R 2 (Regression coefficient) is used as the training loss function and model optimization indicator for model training and multi-model comparison. Its calculation formula is as follows:

[0077] ;

[0078] ;

[0079] Where n is the number of output parameters of the model. RMSE is used to measure the deviation between the predicted value and the true value. The smaller the value, the more accurate the model prediction. 2 Used to evaluate the degree of deviation of the prediction error from the mean benchmark. The larger the value, the more accurate the prediction.

[0080] By adopting the reduced-order model technology, this application can screen the characteristic parameters that affect the strength and stiffness of the aircraft, accurately predict the strength and stiffness response of the aircraft under specific environmental conditions, thereby quickly evaluating the strength and stiffness of the aircraft in real time, guiding the design and use process of the aircraft, and improving the safety and design efficiency of the aircraft.

[0081] The second aspect of the present application provides an intelligent prediction device for structural strength and stiffness characteristics of a reduced-order model corresponding to the above method, mainly comprising:

[0082] A first prediction model acquisition module, used to acquire a trained first prediction model for intelligent prediction of structural strength and stiffness characteristics;

[0083] An input variable contribution rate calculation module, used to change the parameter value of each input variable of the first prediction model respectively, so as to determine the contribution rate of each input variable based on the change of the output value of the first prediction model;

[0084] A reduced-order input variable selection module is used to rank the contribution rates from large to small, and select a plurality of input variables with the highest ranking as reduced-order input variables;

[0085] A second prediction model selection module, used to train the plurality of second prediction models by using the reduced-order input variables as input variables of the plurality of second prediction models, and during the training process, calculating the prediction accuracy of each second prediction model by using the root mean square error or the regression coefficient to select the optimal second prediction model;

[0086] The strength and stiffness characteristics prediction module is used to predict the strength and stiffness characteristics of the aircraft structure by using the selected optimal second prediction model.

[0087] In some optional implementations, the input variable contribution rate calculation module includes:

[0088] An input variable modification unit, used to increase or decrease the input variable by a set percentage, and form two new input variable combinations with other unchanged input variables, and respectively obtain an incremental output combination and a decremental output combination through calculation by the first prediction model;

[0089] A regularized mean square error calculation unit, used to calculate the regularized mean square error of the incremental output combination and the decremental output combination relative to the actual value;

[0090] An influence value calculation unit, used for calculating the influence value of the input variable according to the regularized mean square error of the incremental output combination and the regularized mean square error of the decremental output combination;

[0091] The contribution rate calculation unit is used to normalize the influence value to form the contribution rate of each input variable.

[0092] In some optional embodiments, the set percentage is 8%.

[0093] In some optional implementations, the reduced-order input variable selection module includes:

[0094] Select by quantity unit, which is used to select the top-ranked input variables of a specified number or a specified percentage as reduced-order input variables;

[0095] The cumulative amount selection unit is used to accumulate the contribution rate from front to back until the contribution rate exceeds a set value, and each input variable involved in the accumulation is selected as a reduced-order input variable.

[0096] In some optional embodiments, the plurality of second prediction models include a Dense model, a SVM model, a KNN model, a RF model, and a GBDT model.

[0097] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. An intelligent prediction method for structural strength and stiffness characteristics of a reduced-order model, characterized in that: include: Step S1, obtaining a first prediction model that has been trained for intelligent prediction of structural strength and stiffness characteristics; Step S2: changing the parameter value of each input variable of the first prediction model respectively, so as to determine the contribution rate of each input variable based on the change of the output value of the first prediction model; Step S3, ranking the contribution rates from large to small, and selecting a plurality of input variables with top rankings as reduced-order input variables; Step S4: using the reduced-order input variables as input variables of multiple second prediction models, training the multiple second prediction models. During the training process, the prediction accuracy of each second prediction model is calculated using the root mean square error or the regression coefficient to select the optimal second prediction model. Step S5: predict the strength and stiffness characteristics of the aircraft structure using the selected optimal second prediction model.

2. The intelligent prediction method for structural strength and stiffness characteristics of a reduced-order model according to claim 1, characterized in that: Step S2 further includes performing the following steps on each input variable in the input variable combination of the first prediction model: Step S21, respectively increasing or decreasing the input variables by a set percentage, and combining them with other unchanged input variables to form two new input variable combinations, and respectively calculating the two new input variable combinations through the first prediction model to obtain an incremental output combination and a decremental output combination; Step S22, respectively calculating the normalized mean square error of the incremental output combination and the decremental output combination relative to the actual value; Step S23, calculating the influence value of the input variable according to the regularized mean square error of the incremental output combination and the regularized mean square error of the decremental output combination; Step S24: normalize the influence value to form the contribution rate of each input variable.

3. The intelligent prediction method for structural strength and stiffness characteristics of a reduced-order model according to claim 2, characterized in that: The set percentage is 8%.

4. The intelligent prediction method for structural strength and stiffness characteristics of a reduced-order model according to claim 1, characterized in that: Step S3 further comprises: Select a specified number or percentage of the top ranked input variables as reduced-order input variables; or The contribution rate is accumulated from front to back until the contribution rate exceeds a set value, and each input variable involved in the accumulation is selected as a reduced-order input variable.

5. The intelligent prediction method for structural strength and stiffness characteristics of a reduced-order model according to claim 1, characterized in that: In step S4, the plurality of second prediction models include a Dense model, a SVM model, a KNN model, a RF model and a GBDT model.

6. An intelligent prediction device for structural strength and stiffness characteristics of a reduced-order model, characterized in that: include: A first prediction model acquisition module, used to acquire a trained first prediction model for intelligent prediction of structural strength and stiffness characteristics; An input variable contribution rate calculation module, used to change the parameter value of each input variable of the first prediction model respectively, so as to determine the contribution rate of each input variable based on the change of the output value of the first prediction model; A reduced-order input variable selection module is used to rank the contribution rates from large to small, and select a plurality of input variables with the highest ranking as reduced-order input variables; A second prediction model selection module, used to train the plurality of second prediction models by using the reduced-order input variables as input variables of the plurality of second prediction models, and during the training process, calculating the prediction accuracy of each second prediction model by using the root mean square error or the regression coefficient to select the optimal second prediction model; The strength and stiffness characteristics prediction module is used to predict the strength and stiffness characteristics of the aircraft structure by using the selected optimal second prediction model.

7. The intelligent prediction device for structural strength and stiffness characteristics of a reduced-order model according to claim 6, characterized in that: The input variable contribution rate calculation module includes: An input variable modification unit, used to increase or decrease the input variable by a set percentage, and form two new input variable combinations with other unchanged input variables, and respectively obtain an incremental output combination and a decremental output combination through calculation by the first prediction model; A regularized mean square error calculation unit, used to calculate the regularized mean square error of the incremental output combination and the decremental output combination relative to the actual value; An influence value calculation unit, used for calculating the influence value of the input variable according to the regularized mean square error of the incremental output combination and the regularized mean square error of the decremental output combination; The contribution rate calculation unit is used to normalize the influence value to form the contribution rate of each input variable.

8. The intelligent prediction device for structural strength and stiffness characteristics of a reduced-order model according to claim 7, characterized in that: The set percentage is 8%.

9. The intelligent prediction device for structural strength and stiffness characteristics of a reduced-order model according to claim 6, characterized in that: The reduced-order input variable selection module includes: Select by quantity unit, which is used to select the top-ranked input variables of a specified number or a specified percentage as reduced-order input variables; The cumulative amount selection unit is used to accumulate the contribution rate from front to back until the contribution rate exceeds a set value, and each input variable involved in the accumulation is selected as a reduced-order input variable.

10. The intelligent prediction device for structural strength and stiffness characteristics of a reduced-order model according to claim 6, characterized in that: The multiple second prediction models include a Dense model, a SVM model, a KNN model, a RF model and a GBDT model.

Images