CN-122021320-A - Impact test method based on machine learning

CN122021320ACN 122021320 ACN122021320 ACN 122021320ACN-122021320-A

Abstract

The invention discloses an impact test method based on machine learning, which adopts machine learning to predict input and output parameters of the impact test, and concretely comprises the following steps of S1, ‌, S2, sampling the input parameters by using a Latin hypercube sampling method, S3, performing the impact test according to the sampling parameters to obtain the output parameters, S4, performing neural network training by using MATLAB to obtain a prediction model, S5, applying the model to the impact test, improving test precision and efficiency, and iterating the model. The invention realizes the intelligent prediction of impact test parameters by impact test and machine learning methods, and has the advantages of high efficiency and small error in solving the problem of impact test practical operation. By combining the Latin hypercube sampling method, the training efficiency of the machine learning model is improved while the sample size is reduced, the universality is better, and the constructed generalized architecture can be widely applied to various impact test scenes.

Inventors

ZHANG ZIYE
CAO DEHUA

Assignees

华中科技大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (3)

1. The impact test method based on machine learning is characterized by adopting the machine learning to predict the input and output parameters of the impact test, and specifically comprises the following steps: the method comprises the following steps of S1, confirming input parameters and output parameters, wherein ‌ is used for defining a parameter system of an impact test model, the input parameters comprise pulse width T, peak acceleration A and product weight m, and the output parameters comprise air pressure P, height H and elastic modulus E of a waveform generator; step S2, sampling input parameters by using a Latin hypercube sampling method, and generating data configuration for impact test to obtain an input parameter sampling set, wherein the method specifically comprises the following steps: step S21, defining a parameter space and establishing a three-dimensional design space: S22, space layering, namely equally dividing the whole parameter space into N three-dimensional cube grids, wherein N is the number of samples; step S23, carrying out correlated sampling, namely randomly selecting a sample point in each three-dimensional grid unit, and ensuring that a projection interval of each dimension only contains one sample point; Step S24, generating N groups of associated parameter combinations (T i ,A i ,m i ) in a combined mode, and meeting the condition of spatial uniform distribution: performing data sampling by using a Latin hypercube sampling method by executing the steps S21, S22, S23 and S24 to obtain an input parameter sampling set; and step S3, performing impact test according to the sampling parameters to obtain output parameters corresponding to the sampling input parameters, and constructing a data set, wherein the method specifically comprises the following steps: s31, configuring an impact table according to a parameter group (T i ,A i ,m i ); step S32, performing impact test to obtain a corresponding output data set (P i ,H i ,E i ); Step S4, training a neural network by using MATLAB to obtain a prediction model, wherein the neural network training is realized by using a Neural Network Fitting tool box of the MATLAB and is used for constructing an impact test model by adopting a machine learning method, specifically, according to a data set obtained by sampling and then performing an impact test, constructing the impact test model by adopting a back propagation neural network and adopting an objective function optimization test early-stage setting to obtain the impact test model and early-stage prediction data, constructing the impact test model by adopting the back propagation neural network and adopting the objective function optimization distillation simulation to obtain the impact test model and the test early-stage setting optimization, and the specific steps comprise: step S41, data import, namely importing a number set D= { (T, A, m) → (P, H, E) } into a Neural Network Fitting toolbox of MATLAB; step S42, constructing a neural network, selecting feedforwardnet functions to construct a feedforward network, wherein an input layer comprises 3 nodes corresponding to T, A and m, a hidden layer 1 layer, the number of the nodes is determined through an empirical formula, and an output layer comprises 3 nodes corresponding to P, H and E; step S43, training configuration and training functions, adopting a Levenberg-Marquardt algorithm, setting a learning rate to be a fixed value of 0.01, and setting a termination condition to be the maximum iteration number of 1000 or continuously rising an error of a verification set for 6 times; S44, generating an impact test prediction model after training is executed, and deriving a model structure and a weight matrix; The method comprises the steps of S5, applying a model to an impact test, improving test precision and efficiency, iterating the model, and performing intelligent optimization by combining an impact test prediction model obtained through training with an impact test task target, giving predicted test data and reference information, and adding input and output data obtained through each test to a training data set to iterate the model, wherein the method specifically comprises the following steps of: Step S51, when an impact test is carried out, given target input data is imported into an impact test prediction model to obtain predicted output data, the predicted output data is used for debugging, and the impact test is completed according to the steps; S52, comparing the predicted data with output data of an actual test, and evaluating the accuracy of a predicted model; Step S53, adding the input and output data corresponding to each actual test to a training data set; And S54, performing retraining iteration on the model to optimize the prediction effect.
2. The impact test method based on machine learning according to claim 1, wherein the step S31 comprises the following steps: step S311, connecting the computer, the control instrument and each connecting cable of the platform body, and connecting 220V power supply, connecting the three-phase power supply of the air compressor, and connecting the air compressor and the connecting air pipe of the platform body, wherein the output pressure of the air compressor is not less than 0.6MPa; Step S312, the damping air bag is filled with air; Step 313, placing a test piece and a sensor on the table top, and connecting a sensor wire to a measurement and control instrument; step S314, turning on a computer, turning on power switches of a controller and a tester, and running control software; step 315, confirming parameters such as input channels, setting sensitivity and the like in a hardware setting interface; Step S316, setting sampling parameters, test categories, test standards, trigger parameters and the like on a test target interface; Step S317, setting parameters such as filter type, high cut-off frequency and low cut-off frequency in a filter setting interface; and step S318, setting an operating pressure and a pressurizing pressure value according to actual conditions.
3. The impact test method based on machine learning according to claim 1, wherein the step S32 comprises the following steps: s321, firmly mounting a clamp and a test piece on the table surface of an impact table; Step S322, starting an impact test, prompting 'ready impact' after the table surface rises to a set height, and completing an impact test after 'confirmation'; Step S323, after the test is completed, the setting parameters, the experimental data and the waveform diagram can be output to obtain a corresponding output data set.

Description

Impact test method based on machine learning Technical Field The invention relates to the technical field of impact tests, in particular to an impact test method based on machine learning. Background The impact test is used as a core component of a mechanical environment test, is a key means for verifying the reliability of equipment under dynamic load, and belongs to the category of environment adaptability tests in general quality characteristics. With the development of modern weaponry to high precision and high reliability, the coverage range and precision requirements of fields such as aviation, aerospace and ships on impact tests are remarkably improved. Taking carrier-based electronic equipment as an example, the impact resistance of the carrier-based electronic equipment is directly related to the viability of a combat system and must be verified through a strict impact test. The current impact test technology faces three major core challenges, namely low test parameter debugging efficiency, repeated attempts are needed by the traditional method to obtain impact waveforms or response spectrums meeting the standard requirements of GJB150.18A and the like, the debugging time can be 3-5 times of effective test time, the control precision of the test process is insufficient, the test is easy to cause over-test phenomena (such as exceeding of peak acceleration) due to manual operation, unnecessary damage is caused to a test piece, and an intelligent test data application system is lacking, so that the construction requirement of a future unmanned test room is difficult to support. This is an urgent need to develop new test methods with intelligent prediction of parameters, whose technological breakthroughs will directly improve equipment development efficiency and reduce validation costs. Disclosure of Invention The invention aims to provide an impact test method based on machine learning, which can accurately predict parameters of an impact test bed according to test conditions and correct a prediction model according to test results so as to effectively improve impact test efficiency. To achieve the above object, the present invention provides an impact test method based on machine learning (taking a pneumatic impact bench half-sine impact test as an example), including: and S1, confirming input parameters and output parameters. ‌ A And S2, sampling the input parameters by using a Latin hypercube sampling method. And S3, configuring the weight of the tested object according to the sampling parameters, performing impact test to obtain a corresponding half-sine curve, and extracting output parameters including air pressure, height and elastic modulus of the waveform generator. And S4, training a neural network by using MATLAB to obtain a prediction model. And S5, applying the model to an impact test, improving test precision and efficiency, and iterating the model. Further, in step S1, the input parameters and the output parameters are confirmed, and the input parameters are used for defining a parameter system of the impact test model, the input parameters specifically include a pulse width T, a peak acceleration a, and a product weight m, and the output parameters specifically include an air pressure P, a height H, and an elastic modulus E of the waveform generator. Further, in step S2, the step of sampling the input parameters is used for generating a data configuration for performing an impact test, specifically, a latin hypercube sampling method is adopted to obtain an input parameter sampling set, and the method includes the following steps: step S21, defining a parameter space and establishing a three-dimensional design space: And S22, space layering, namely equally dividing the whole parameter space into N three-dimensional cube grids, wherein N is the number of samples. Step S23, carrying out correlated sampling, namely randomly selecting a sample point in each three-dimensional grid unit, and ensuring that a projection interval of each dimension only contains one sample point. Step S24, generating N groups of associated parameter combinations (T i,Ai,mi) in a combined mode, and meeting the condition of spatial uniform distribution: Further, in step S3, the impact test is performed, so as to obtain an output parameter corresponding to the sampled input parameter, and construct a data set, which specifically includes: And S31, configuring the impact table according to the parameter group (T i,Ai,mi). And S32, performing impact test to obtain a corresponding output data set (P i,Hi,Ei). Further, in step S4, the neural network training is implemented by using a Neural Network Fitting toolbox of MATLAB, and is used for constructing an impact test model by using a machine learning method, specifically, according to a data set obtained by performing an impact test after sampling, constructing the impact test model by using a counter-propagating neural network, and optimizing a test early-stage setting