CN111931404B - Self-optimization CNN-based bore complex surface contact collision response prediction method - Google Patents

Abstract

The invention relates to a bore complex surface contact collision response prediction method based on a self-optimization Convolutional Neural Network (CNN), which comprises the following steps: s1, establishing a wear barrel-projectile contact collision finite element model according to the barrel-projectile test platform based on artillery shooting test data; s2, constructing a sample set based on the finite element model and by combining the test result of the test platform; s3, based on the sample set, according to a cooperative working mechanism of a genetic algorithm-sequence quadratic programming algorithm combined optimization algorithm, self-optimization of the super-parameters is achieved, and therefore the optimal super-parameter set is obtained; meanwhile, training in the optimization of the hyper-parameters and obtaining an optimal barrel-projectile contact collision model based on self-optimization CNN; and S4, further researching the change rule of the contact collision response according to the obtained barrel-projectile contact collision model. According to the method, the influence of the complex surface of the worn inner bore and the contact collision energy loss are considered, the response precision is guaranteed, and meanwhile the prediction efficiency is greatly improved.

Description

Bore complex surface contact collision response prediction method based on self-optimization CNN

Technical Field

The invention relates to the technical field of contact collision analysis, in particular to a bore complex surface contact collision response prediction method based on self-optimization CNN.

Background

As a complex multi-body system, the artillery has shooting precision which is an important index of tactical performance. Under the influence of the gap between the cannonball and the shell, the cannonball continuously generates violent bullet plastic contact collision with the barrel in the high-speed movement process in the bore, so that the flexible vibration of the barrel is coupled with the movement of the cannon bore, the disturbance of the cannon mouth of the cannonball is caused, and the shooting precision of the cannon is reduced. During the service period of the artillery, the inner chamber is abraded due to ablation and washing of high-temperature and high-pressure gunpowder gas and repeated mechanical action of the projectile, and the barrel-projectile contact collision response is influenced.

Aiming at the problem of barrel-projectile complex surface elastoplasticity contact collision, the traditional knowledge-driven modeling method cannot provide response prediction with both precision and efficiency due to excessive assumption simplification or huge calculation memory requirements, and the existing research is less related to rifling abrasion in view of the complexity of the problem.

Disclosure of Invention

Technical problem to be solved

In view of the defects and shortcomings of the prior art, the invention provides a bore complex surface contact collision response prediction method based on self-optimization CNN, which solves the technical problems of efficiently and accurately predicting contact collision response and helping to clarify the stress state and motion rule in a projectile bore under the condition of considering the abrasion of the artillery bore.

(II) technical scheme

In order to achieve the purpose, the invention adopts the main technical scheme that:

s1, based on artillery shooting test data, establishing a wear barrel-projectile contact collision finite element model by taking a barrel-projectile test platform as a prototype;

s2, constructing a multi-working-condition contact collision sample set based on the finite element model and by combining a test result of the test platform;

s3, based on the sample set, according to a cooperative working mechanism of a convolutional neural network and genetic algorithm-sequence quadratic programming algorithm combined optimization algorithm, self-optimization of the convolutional neural network hyper-parameters is achieved, and therefore the optimal hyper-parameter set is obtained; training in the process of optimizing the hyper-parameters and obtaining an optimal barrel-projectile contact collision model based on the self-optimization convolutional neural network;

and S4, researching a contact collision response change rule in a variable parameter manner according to the self-optimization convolution neural network-based barrel-projectile contact collision model.

Optionally, step S1 includes:

s11, obtaining the corresponding relation between the distribution of the abrasion loss of the artillery inner bore and the shot firing number according to the collected artillery shooting test data;

s12, processing the test data by using a non-equal interval gray system model, and establishing a bore wear mathematical model by combining the corresponding relation between wear loss distribution and the shot number;

s13, constructing barrel-projectile contact collision prior finite element models with different wear degrees based on the bore wear mathematical model by taking the barrel-projectile test platform as a prototype;

s14, correcting relevant parameter settings of the prior finite element model by combining the stress analysis result and the test platform test data, wherein the relevant parameters comprise grid size and contact impact rigidity;

and S15, executing the step S14 for many times until the relative error between the corrected model simulation result and the test result of the test platform is smaller than a preset error, and obtaining the wear barrel-projectile contact collision finite element model.

Optionally, the bore wear mathematical model is:

Δd＝f(n,x)

where Δ d represents the bore wear amount, n is the number of shots, and x is the axial position.

Optionally, step S2 includes:

s21, determining the influence factors of the barrel-projectile contact collision response;

s22, generating an orthogonal table according to the physical quantity change range of each influence factor, and carrying out multi-working-condition elasto-plastic contact collision finite element numerical simulation according to the orthogonal table to obtain a simulation result;

and S23, analyzing the simulation result, determining the input physical quantity and the output physical quantity of the neural network model, and constructing a multi-working-condition contact collision sample set, wherein the sample set comprises a training set and a verification set.

Optionally, the influencing factors comprise the caliber of the barrel, the degree of wear and the initial contact collision motion state;

the input physical quantities comprise barrel-projectile material parameters, geometric dimensions and initial contact collision motion states;

the output physical quantity comprises contact impact force, rolling reduction, contact impact speed, contact area, plastic energy loss and deformation, maximum rifling number of contact impact, maximum contact pressure and Mises stress.

Optionally, in step S3, the implementing self-optimization of the hyper-parameters of the convolutional neural network according to a cooperative working mechanism of a convolutional neural network and a genetic algorithm-sequential quadratic programming algorithm combined optimization algorithm based on the sample set, so as to obtain the optimal hyper-parameter set includes:

s31, randomly generating convolutional neural network chromosomes with different sets of hyper-parameters, wherein each convolutional neural network chromosome comprises a plurality of hyper-parameter genes to be optimized, and the genes are used as initial populations of genetic algorithms; training to obtain the mean square error of each convolutional neural network;

s32, constructing an initial population individual fitness evaluation index based on a mean square error loss function of a convolutional neural network;

s33, selecting the parent chromosome and the parent chromosome by using a roulette mode, and generating child chromosomes continuously through crossing and mutation;

s34, judging whether a first convergence condition is met;

S35A, if not, returning to the step S32;

S35B, if yes, selecting the individuals with the highest individual fitness in the last generation of population as the optimal solution of the genetic algorithm;

s36, taking the optimal solution of the genetic algorithm as the input of a sequence quadratic programming algorithm;

s37, starting a local optimization process;

s38, judging whether a second convergence condition is met;

S39A, if not, returning to the step S37;

and S39B, if yes, using the combined optimized value as the optimal hyper-parameter group of the convolutional neural network.

Optionally, in step S3, the training and obtaining an optimal self-optimized convolutional neural network-based barrel-projectile contact collision model in the hyper-parametric optimization process includes:

a31, setting different sets of hyperparameters, wherein the steps A32 to A36 are all carried out in the same hyperparameter set;

a32, extracting the data features of the training set by adopting a forward propagation algorithm;

a33, measuring the estimated contact collision response and the actual value error of the convolutional neural network based on the mean square error loss function;

a34, judging whether a third convergence condition is met;

a35a, if the third convergence condition is not met, performing error back propagation by adopting an Adam gradient optimization algorithm, adjusting the network weight and bias, and repeating the step A33;

a35b, if a third convergence condition is met, terminating training to obtain a barrel-projectile contact collision prior model based on a convolutional neural network;

a36, evaluating the predicted performance of the prior model for new data by using the verification set;

a37, repeating the steps A32 to A36 for a plurality of times according to different super parameter groups to generate a plurality of barrel-projectile contact collision models based on different convolutional neural networks; the model with the optimal prediction performance on the verification set is the optimal barrel-projectile contact collision model based on the self-optimization convolutional neural network.

Optionally, the mean square error loss function is:

the individual fitness evaluation index is as follows:

wherein y is _i In order to achieve the target value,

n is the number of samples, and Index is the individual fitness evaluation Index.

Optionally, the first convergence condition is: setting a maximum evolution algebra, and multiplying the maximum evolution algebra by the number of variables through 100 to obtain the maximum evolution algebra; or when the fitness of the population does not rise any more in 50 generations, taking the fitness difference less than or equal to 1e-6 as a judgment basis; or setting the total iteration number to 10000;

the second convergence condition is as follows: step tolerance and function value tolerance are both smaller than a set threshold, and iteration is finished;

the third convergence condition is: stopping training when the MSE on the training set continuously decreases and the MSE on the verification set does not decrease and inversely increase; or when the MSE on the validation set no longer decreases for 20 consecutive times, stop training; or the maximum number of training 10000.

Optionally, the optimal hyper-parameter set includes a global hyper-parameter and a local hyper-parameter; the global hyper-parameter comprises: the method comprises the following steps of (1) learning rate, first moment estimation exponential decay rate, second moment estimation exponential decay rate, maximum iteration times, iteration stop threshold, batch processing scale and network weight initialization mode;

the local hyper-parameters comprise: convolution layer, batch normalization layer, nonlinear activation layer, pooling layer collocation mode and quantity, convolution filter size and number, filling type, pixel point number, convolution step, activation function type, pooling standard, pooling filter size and number, pooling filter step, full-connection layer activation function, full-connection layer number, neuron number of each layer and dropout probability.

(III) advantageous effects

The beneficial effects of the invention are: the invention establishes a self-optimization CNN-based bore complex surface contact collision response prediction method by utilizing a deep learning technology, considers the influence of the worn bore complex surface and the contact collision energy loss in the actual working condition, ensures the contact collision response precision and greatly improves the prediction efficiency. The prediction method provided by the invention can further define the stress state and the motion rule in the projectile bore and perfect the coupling theory of the projectile, is favorable for revealing the disturbance influence mechanism of the muzzle, and has important significance on the analysis of the firing precision and the structural optimization design of the artillery.

Drawings

Fig. 1 is a schematic flow chart of a bore complex surface contact collision response prediction method based on self-optimized CNN provided by the present invention;

fig. 2 is a detailed flowchart of step S1 of the method for predicting a bore complex surface contact collision response based on self-optimized CNN according to the present invention;

fig. 3a is a schematic diagram of the relative position of the barrel and the projectile of the bore complex surface contact collision response prediction method based on self-optimized CNN according to the present invention;

fig. 3b is a schematic diagram of a shot projectile before shooting based on a self-optimized CNN bore complex surface contact collision response prediction method provided by the present invention;

fig. 3c is a schematic diagram of a shooting forebarrel of the bore complex surface contact collision response prediction method based on self-optimized CNN provided by the present invention;

fig. 3d is a schematic diagram of a shot projectile according to the method for predicting contact collision response of a bore complex surface based on self-optimized CNN provided by the present invention;

fig. 3e is a schematic diagram of a fired barrel of the bore complex surface contact collision response prediction method based on self-optimized CNN provided by the present invention;

fig. 4 is a schematic diagram of a barrel-projectile test platform of the bore complex surface contact collision response prediction method based on self-optimization CNN provided by the invention;

fig. 5 is a diagram of a body tube-projectile test platform of the bore complex surface contact collision response prediction method based on self-optimized CNN provided by the present invention;

fig. 6 is a detailed flowchart of step S2 of the method for predicting a bore complex surface contact collision response based on self-optimized CNN according to the present invention;

fig. 7 is a specific flowchart of the automatic optimization of the hyper-parameters in step S3 of the method for predicting the contact collision response of the complex surface of the bore based on the self-optimized CNN according to the present invention;

FIG. 8 is a schematic diagram of automatic hyper-parameter optimization of a bore complex surface contact collision response prediction method based on self-optimized CNN provided by the present invention;

fig. 9 is a schematic flowchart of a specific process of model training in step S3 of the bore complex surface contact collision response prediction method based on self-optimized CNN according to the present invention;

fig. 10 is a schematic diagram of model training of the bore complex surface contact collision response prediction method based on self-optimized CNN provided by the present invention.

[ description of reference ]

1: a bearing; 2: an adapter; 3: swinging arms; 4: a projectile simulation sample section; 5: a barrel bore sample section; 6: a force sensor; 7: a base; 8-1: a first eddy current displacement sensor; 8-2: a second eddy current displacement sensor; 9: and a third eddy current displacement sensor.

Detailed Description

For the purpose of better explaining the present invention and to facilitate understanding, the present invention will be described in detail by way of specific embodiments with reference to the accompanying drawings.

Fig. 1 is a schematic flow chart of a bore complex surface contact collision response prediction method based on self-optimization CNN provided by the present invention, and as shown in fig. 1, the first premise of the prediction method provided by the present invention is to construct a high-precision diversity sample set, and therefore, a worn barrel-projectile contact collision finite element model is first established based on artillery fire test data and using a barrel-projectile test platform as a prototype. Secondly, on the basis of the finite element model, a high-precision multi-working-condition contact collision sample set is constructed by combining the test result of the test platform. Then, considering the advantages of the convolutional neural network in the network training efficiency and avoiding the risk of overfitting, in order to obtain the optimal convolutional neural network hyper-parameter, the invention develops research on a cooperative working mechanism of the convolutional neural network and genetic algorithm-sequence quadratic programming algorithm combined optimization algorithm based on a sample set, realizes the self-optimization of the convolutional neural network hyper-parameter, and thus obtains the optimal hyper-parameter set; meanwhile, training is carried out in the process of optimizing the hyper-parameters, and an optimal barrel-projectile contact collision model based on the self-optimization convolutional neural network is obtained. After the model is established, the contact collision response change rule can be predicted.

The invention establishes a self-optimization CNN-based bore complex surface contact collision response prediction method by utilizing a deep learning technology, considers the influence of the worn bore complex surface and the contact collision energy loss in the actual working condition, ensures the contact collision response precision and greatly improves the prediction efficiency. The prediction method provided by the invention can further define the stress state and the motion rule in the projectile bore and perfect the coupling theory of the projectile, is favorable for revealing the disturbance influence mechanism of the muzzle, and has important significance on the analysis of the firing precision of the artillery and the optimization design of the structure of the artillery.

For a better understanding of the above-described technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

CNN is the abbreviation of Convolutional Neural Networks (Convolutional Neural Networks), and GA-SQP is the abbreviation of Genetic Algorithm-Sequential Quadratic Programming (Genetic Algorithm-Sequential Quadratic Programming). The invention provides a self-optimization CNN-based bore complex surface contact collision response prediction method, which comprises the following steps:

and S1, establishing a wear barrel-projectile contact collision finite element model by taking the barrel-projectile test platform as a prototype based on artillery shooting test data.

Fig. 2 is a schematic flowchart of a step S1 of the method for predicting a bore complex surface contact collision response based on self-optimized CNN according to the present invention, and as shown in fig. 2, the following is a specific flowchart of the step S1:

and S11, obtaining the corresponding relation between the distribution of the abrasion loss of the artillery inner bore and the shot firing number according to the collected artillery shooting test data. Different axial positions have different wear, and the corresponding relation is that the number of shots corresponds to the wear amount. The corresponding relation between the distribution of the axial abrasion loss and the firing number is researched, so that the abrasion condition of the inner bore of the gun under different firing numbers can be determined, the ballistic performance of the gun can be calibrated, the service life of the gun can be predicted, and the adverse effect of the use of the gun can be avoided.

And S12, aiming at common defects (such as small samples, poor information, unequal interval acquisition and the like) of artillery shooting test data, processing the test data by utilizing an unequal interval gray system model, establishing an artillery bore wear mathematical model by combining the corresponding relation between wear loss distribution and shooting number, and researching the dynamic evolution rule of the rifling shape. The mathematical model of bore wear is:

Δd＝f(n,x)

where Δ d represents the bore wear amount, n is the number of shots, and x is the axial position. Specifically, the wear amount and the shot number show positive correlation change, and the relative change relationship with the axial position is more complex, and the barrel bore is divided into three regions along the axial direction according to the bore wear condition: severe wear areas, even wear areas, and muzzle wear areas. The severe abrasion area is positioned in the range from the initial position of the rifling to 10 times of the caliber length; the uniform wear zone is slightly worn compared to the severe wear zone, located in the middle section of the bore; and the muzzle abrasion area is located within 1.5-2 times of the caliber of the muzzle.

S13, as shown in fig. 3a, 3b, 3c, 3d and 3e, reveals barrel-projectile relative position and contact surface topography variations of the artillery bore. As can be seen from figure 3c, unlike conventional regular geometric shapes (e.g. spheres and cylinders), the number of rifling bores is numerous and the bore cross-section presents a profile-like irregular shape, complicating barrel-projectile contact collision surfaces.

Fig. 4 is a schematic diagram of a barrel-projectile test platform based on a self-optimized CNN bore complex surface contact collision response prediction method provided by the present invention, as shown in fig. 4, bearings 1 are disposed at two ends of an adaptor 2, a middle portion of the adaptor 2 is connected to a first end of a swing arm 3, and a projectile simulation sample segment 4 is connected to a second end of the swing arm 3. One end of the force sensor 6 is connected with the barrel inner bore sample section 5, and the other end is connected with the base 7. When the swing arm 3 is in the free-hanging position, the surface of the projectile simulation sample section 4 is in surface contact with the barrel bore sample section 5. The first eddy current displacement sensor 8-1 and the second eddy current displacement sensor 8-2 are positioned at two sides of the projectile simulation sample section 4 and are used for measuring the displacement of the projectile simulation sample section 4. A third eddy current displacement sensor 9 is mounted on the base 7 for monitoring the movement of the base 7. During the firing of the artillery, the contact collision between the barrel and the projectile is simulated in the following way: the swing arm 3 is pulled to a certain initial angle, the swing arm 3 is released freely, and when the swing arm 3 moves to the lowest point, the projectile simulation sample section 4 and the barrel bore sample section 5 are in contact collision. And fig. 5 is a real object diagram of a barrel-projectile test platform of the self-optimization CNN-based bore complex surface contact collision response prediction method, as shown in fig. 4 and 5, the barrel-projectile test platform is taken as a prototype, a bore wear mathematical model is used, a Python script file is used to realize secondary development of a finite element preprocessing module, and barrel-projectile contact collision prior finite element models with different wear degrees are constructed.

And S14, correcting the related parameter setting of the finite element model in advance by combining the stress analysis result and the test platform test data. Relevant parameters mainly include mesh size, impact contact stiffness and the like. Specifically, considering the problems of numerous model grids, large calculation amount and the like, firstly, simulation is carried out by adopting a coarse grid, then, a contact collision area is estimated by combining a stress analysis result, and grid encryption is implemented. And comparing test results of the test platform, and further adjusting the setting of parameters such as grid size, contact collision rigidity and the like.

And S15, executing the step S14 for multiple times until the relative error between the corrected model simulation result and the test result of the test platform is smaller than a preset error which is 5%, and obtaining the wear barrel-projectile contact collision finite element model.

And S2, constructing a multi-working-condition contact collision sample set based on the finite element model and combined with the test result of the test platform. The method has the advantages that the method for contact collision dynamics modeling based on self-optimization CNN is developed on the premise that a high-precision multi-working-condition contact collision sample set is constructed. On the basis of a finite element model of the contact collision of the worn barrel and the projectile, the numerical simulation research of the contact collision under multiple working conditions is carried out, and a sample set required by training is constructed. The training set is a data sample used for model fitting, and the validation set is used for evaluating the prediction capability of the model.

Fig. 6 is a schematic flowchart of a step S2 of the method for predicting a bore complex surface contact collision response based on self-optimized CNN according to the present invention, and as shown in fig. 6, the following is a specific flowchart of the step S2:

and S21, determining the influencing factors of the barrel-projectile contact collision response. The effect of various factors including the caliber of the barrel, the degree of wear of the barrel, and the initial contact collision motion state need to be considered before constructing the sample set.

And S22, in order to obtain a comprehensive, uniform and efficient simulation design scheme, generating an orthogonal table according to the physical quantity change range of each influence factor, and carrying out multi-working-condition elastoplastic contact collision finite element numerical simulation according to the orthogonal table. Table 1 shows a contact collision simulation orthogonal table of the bore complex surface contact collision response prediction method based on self-optimized CNN provided in the present invention, as shown in table 1, the contact collision simulation orthogonal table includes four influencing factors, which are respectively caliber a, projectile and cannon gap B, initial contact collision velocity C and angle D, and the horizontal number of each factor is 3, 8, 9, and 9. The data in the table represent the level number and the corresponding actual value, respectively.

A：105mm，122mm，155mm；

B：0.1mm，0.3mm，0.5mm，0.7mm，0.9mm，1.1mm，1.3mm，1.5mm；

C：1m/s，1.5m/s，2m/s，2.5m/s，3m/s，3.5m/s，4m/s，4.5m/s，5m/s；

D：1mil，2mil，3mil，4mil，5mil，6mil，7mil，8mil，9mil。

TABLE 1

And S23, analyzing the simulation result, determining the input physical quantity and the output physical quantity of the neural network model, and constructing a multi-working-condition contact collision sample set, wherein the sample set comprises a training set and a verification set. For a barrel-projectile contact collision system, the following three aspects are to be considered as input: barrel-projectile material parameters, geometry, and initial contact collision motion state. The material parameters mainly comprise density, elastic modulus, Poisson ratio, yield stress, hardening index and the like; the geometric dimensions mainly comprise a gun gap, a centering radius and width, a positive line radius and width, a negative line radius and width and a rifling height; the initial contact collision motion state mainly comprises an initial contact collision speed and an angle. The output mainly considers the contact collision force, the rolling reduction, the contact collision speed, the contact area, the plastic energy loss and the deformation, the maximum rifling number of the contact collision, the maximum contact pressure, the Mises stress and the like. In order to stabilize and accelerate the network training process, sample data needs to be normalized before training.

And S3, based on the sample set, realizing self-optimization of the hyper-parameters of the convolutional neural network according to a cooperative working mechanism of a convolutional neural network and genetic algorithm-sequence quadratic programming algorithm combined optimization algorithm, thereby obtaining the optimal hyper-parameter set. Meanwhile, training is carried out in the process of searching the hyper-parameters, and an optimal barrel-projectile contact collision model based on the self-optimization convolutional neural network is obtained. The hyper-parameters (including network architecture and related training parameters) are often referred to as "knobs" for neural networks, the settings of which severely impact network structure, performance, and training efficiency. The commonly used "trial and error" method is not only time consuming but also highly dependent on the debugging experience of the designer. And the GA-SQP combined optimization algorithm can greatly improve the solving precision and the convergence speed of parameter optimization. Based on the method, the CNN hyper-parameters are used as design variables to construct an optimization problem mathematical model, and the CNN and GA-SQP combined optimization algorithm is effectively combined, so that the CNN hyper-parameters are automatically optimized and obtained, and the optimal hyper-parameter set is obtained. Meanwhile, training is carried out in the process of optimizing the hyper-parameters, and an optimal barrel-projectile contact collision model based on the self-optimization convolutional neural network is obtained.

Fig. 7 is a schematic diagram of a detailed process of automatic optimization of the hyper-parameter in step S3 of the prediction method of bore complex surface contact collision response based on self-optimized CNN provided by the present invention, fig. 8 is a schematic diagram of automatic optimization of the hyper-parameter of the prediction method of bore complex surface contact collision response based on self-optimized CNN provided by the present invention, as shown in fig. 7 and 8, the following is a detailed process of automatic optimization of the hyper-parameter in step S3:

s31, randomly generating convolutional neural network chromosomes with different sets of hyper-parameters, wherein each convolutional neural network chromosome comprises a plurality of hyper-parameter genes to be optimized, and the genes are used as initial populations of genetic algorithms; and training to obtain the mean square error of each convolutional neural network.

S32, constructing an initial population individual fitness evaluation index based on the mean square error loss function of the convolutional neural network. The mean square error loss function is:

wherein y is _i In order to achieve the target value,

the Index is an individual fitness evaluation Index, and is known from a formula, and the smaller the value of the mean square error loss function MSE is, the larger the value of the individual fitness evaluation Index is.

S33, selecting the parent chromosome and the parent chromosome by roulette, and generating the child chromosomes continuously through crossing and mutation.

The roulette mode can enable individuals with high fitness evaluation indexes to have higher probability of being selected as parent individuals to generate new individuals, and the crossover refers to the exchange of partial genes of different chromosomes, and the mutation can prevent local convergence. The three operations of selection, crossing and variation not only ensure that the chromosomes with high individual fitness indexes are kept to the next generation of population as much as possible, but also ensure the change of the next generation of population and generate new chromosomes.

S34, it is determined whether the first convergence condition is satisfied. The first convergence condition is: setting a maximum evolution algebra, and multiplying the maximum evolution algebra by the number of variables through 100 to obtain the maximum evolution algebra; or when the population does not rise any more in the 50 generations, the fitness difference is less than or equal to 1e-6 as a judgment basis; or set the total number of iterations to 10000.

S35A, if not, returning to the step S32.

And S35B, if yes, selecting the individual with the highest individual fitness in the population of the last generation as the optimal solution of the genetic algorithm.

And S36, taking the optimal solution of the genetic algorithm as the input of the sequence quadratic programming algorithm.

And S37, starting a local optimizing process.

S38, see whether the second convergence condition is satisfied. The second convergence condition is: and step tolerance (1e-10) and function value tolerance (1e-6) are both smaller than a set threshold, and the iteration is finished.

S39A, if not, the flow returns to step S37.

And S39B, if yes, using the combined optimized value as the optimal hyper-parameter group of the convolutional neural network. The optimal hyper-parameter set comprises a global hyper-parameter and a local hyper-parameter; the global hyper-parameters include: the method comprises the following steps of (1) learning rate, first moment estimation exponential decay rate, second moment estimation exponential decay rate, maximum iteration times, iteration stop threshold, batch processing scale and network weight initialization mode; the local hyper-parameters include: convolution layer, batch normalization layer, nonlinear activation layer, pooling layer collocation manner and quantity, convolution filter size and number, filling type, pixel point number, convolution step, activation function type, pooling standard, pooling filter size and number, pooling filter step, full-connection layer activation function, full-connection layer number, neuron number of each layer and dropout probability. Learning rate: the method is mainly used for adjusting the update speed of the weight and the bias of the neural network. The learning rate is set to be too small, and the convergence speed is low; the arrangement is too large, which easily causes oscillation phenomenon. The first moment estimates the exponential decay rate, the second moment estimates the exponential decay rate: the Adam gradient optimization algorithm involves related parameters, and independent adaptive learning rates can be designed for different weights and bias learning. Maximum iteration number, stop iteration threshold: the parameters involved in the termination of the training are set. Batch size: the gradient descent method involves parameter setting. Network weight initialization mode: mainly refers to the weight initialization mode of the convolutional layer and the full connection layer in the established network.

Fig. 9 is a schematic diagram of a specific flow of model training in step S3 of the bore complex surface contact collision response prediction method based on self-optimized CNN provided by the present invention, and as shown in fig. 9, the following is a specific flow of model training in step S3:

and A31, setting different sets of hyperparameters, wherein the steps A32 to A36 are all carried out in the same hyperparameter set.

And A32, extracting the data characteristics of the training set by adopting a forward propagation algorithm.

A33, measuring the estimated contact collision response and the actual value error of the convolutional neural network based on the mean square error loss function.

And A34, judging whether the third convergence condition is satisfied.

And A35a, if the third convergence condition is not met, performing error back propagation by adopting an Adam gradient optimization algorithm, adjusting the network weight and the bias, and repeating the step A33.

And A35b, if a third convergence condition is met, terminating training to obtain a barrel-projectile contact collision prior model based on the convolutional neural network. The third convergence condition is: stopping training when the MSE on the training set continuously decreases and the MSE on the verification set does not decrease and inversely increase; or when the MSE on the verification set does not decrease any more for 20 consecutive times, stopping training; or a maximum number of training 10000.

And A36, evaluating the predicted performance of the prior model for the new data by using the verification set.

A37, repeating the steps A32 to A36 for a plurality of times according to different super parameter sets to generate a plurality of barrel-projectile contact collision models based on different convolutional neural networks. The model with the optimal prediction performance on the verification set is the optimal barrel-projectile contact collision model based on the self-optimization convolutional neural network.

Fig. 10 is a schematic diagram of model training of the bore complex surface contact collision response prediction method based on self-optimization CNN provided in the present invention, where the model training process is as shown in fig. 10, firstly, a forward propagation algorithm is used to extract data features of a training set, a mean square error regression loss function is used to measure a difference between a network predicted contact collision response and an actual value, and if a third convergence condition is not satisfied, an Adam gradient optimization algorithm is used to perform error back propagation, and network weight and bias are adjusted; if the third convergence condition is satisfied, the training is terminated.

And S4, according to the self-optimization convolutional neural network-based barrel-projectile contact collision model, researching the change rule of the contact collision response in a variable parameter manner. The barrel-projectile contact collision model based on the self-optimization convolutional neural network is used for carrying out variable parameter research on the worn barrel-projectile contact collision process, and mining the change rules of contact collision force, rolling reduction, contact collision speed and the like along with the gun gap, the contact collision motion state of the projectile and the barrel and the like.

In summary, the invention discloses a bore complex surface contact collision response prediction method based on self-optimization CNN, which specifically introduces: the gun with various calibers in service is taken as a research object, and a bore axial wear mathematical model is respectively established by utilizing the measured data of a target range, so that the dynamic evolution rule of the rifling shape is revealed. Then, a wear barrel-projectile contact collision finite element model is established. Based on the finite element model, carrying out multi-working-condition numerical simulation research, determining input and output characterization parameters of the neural network model, and constructing a multi-working-condition contact collision sample set.

Further, constructing an initial population, introducing the contact collision sample set, and improving the network prediction precision by adopting an error back propagation algorithm; and exploring a cooperative working mechanism of the CNN and a combined optimization algorithm, developing super-parameter self-optimization research, and establishing an optimal barrel-projectile contact collision model based on a self-optimization convolutional neural network. And verifying response prediction precision by using a test result of a test platform, developing variable parameter research in a contact collision process, and determining a dynamic response change rule.

The method utilizes a data-driven modeling framework, has simple principle and easy realization, and can realize the high-efficiency and high-precision characterization of the contact collision behavior of the complex surface without depending on the internal complex mechanism analysis and the early-stage related professional knowledge accumulation of the contact collision process.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions.

It should be noted that in the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the terms first, second, third, etc. are used for convenience only and do not denote any order. These words are to be understood as part of the name of the component.

Furthermore, it should be noted that in the description of the present specification, the description of the term "one embodiment", "some embodiments", "examples", "specific examples" or "some examples", etc., means that a specific feature, structure, material or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.

While preferred embodiments of the present invention have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, the claims should be construed to include preferred embodiments and all such variations and modifications as fall within the scope of the invention.

It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit or scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention should also include such modifications and variations.

Claims (1)

1. A self-optimization CNN-based bore complex surface contact collision response prediction method is characterized by comprising the following steps:

s1, based on artillery shooting test data, establishing a wear barrel-projectile contact collision finite element model by taking a barrel-projectile test platform as a prototype;

step S1 includes:

s11, obtaining the corresponding relation between the distribution of the abrasion loss of the artillery inner bore and the shot firing number according to the collected artillery shooting test data;

s12, processing the test data by using a non-equal interval gray system model, and establishing a bore wear mathematical model by combining the corresponding relation between wear loss distribution and the shot firing number;

the mathematical model of bore wear is:

Δd＝f(n,x)

wherein, deltad represents the wear amount of the inner bore, n is the number of shot shots, and x is the axial position;

s13, constructing barrel-projectile contact collision prior finite element models with different wear degrees by taking the barrel-projectile test platform as a prototype and based on the bore wear mathematical model;

s14, correcting the related parameter setting of the prior finite element model by combining the stress analysis result and the test platform test data, wherein the related parameters comprise grid size and contact collision rigidity;

s15, executing the step S14 for multiple times until the relative error between the corrected model simulation result and the test result of the test platform is smaller than the preset error, and obtaining a wear barrel-projectile contact collision finite element model;

s2, constructing a multi-working-condition contact collision sample set based on the finite element model and by combining a test result of the test platform;

step S2 includes:

s21, determining the influence factors of the barrel-projectile contact collision response;

s22, generating an orthogonal table according to the physical quantity change range of each influence factor, and carrying out multi-working-condition elasto-plastic contact collision finite element numerical simulation according to the orthogonal table to obtain a simulation result;

s23, analyzing the simulation result, determining input physical quantity and output physical quantity of the neural network model, and constructing a multi-working-condition contact collision sample set, wherein the sample set comprises a training set and a verification set;

the influencing factors comprise the caliber of the barrel, the abrasion degree and the initial contact collision motion state;

the input physical quantities comprise barrel-projectile material parameters, geometric dimensions and initial contact collision motion states;

the output physical quantity comprises contact collision force, rolling reduction, contact collision speed, contact area, plastic energy loss and deformation, maximum rifling number of contact collision, maximum contact pressure and Mises stress;

s3, based on the sample set, according to a cooperative working mechanism of a convolutional neural network and genetic algorithm-sequence quadratic programming algorithm combined optimization algorithm, self-optimization of the convolutional neural network hyper-parameters is achieved, and therefore the optimal hyper-parameter set is obtained; training in the process of optimizing the hyper-parameters and obtaining an optimal barrel-projectile contact collision model based on the self-optimization convolutional neural network;

in step S3, the implementing self-optimization of the convolutional neural network hyper-parameter according to the cooperative work mechanism of the convolutional neural network and the genetic algorithm-sequence quadratic programming algorithm combined optimization algorithm based on the sample set, so as to obtain the optimal hyper-parameter set includes:

s31, randomly generating convolutional neural network chromosomes with different sets of hyper-parameters, wherein each convolutional neural network chromosome comprises a plurality of hyper-parameter genes to be optimized, and the hyper-parameter genes are used as the initial population of the genetic algorithm; training to obtain the mean square error of each convolutional neural network;

s32, constructing an initial population individual fitness evaluation index based on a mean square error loss function of a convolutional neural network;

s33, selecting the parent chromosome and the parent chromosome by using a roulette mode, and generating child chromosomes continuously through crossing and mutation;

s34, judging whether a first convergence condition is met;

S35A, if not, returning to the step S32;

S35B, if yes, selecting the individuals with the highest individual fitness in the last generation of population as the optimal solution of the genetic algorithm;

s36, taking the optimal solution of the genetic algorithm as the input of a sequence quadratic programming algorithm;

s37, starting a local optimization process;

s38, judging whether a second convergence condition is met;

S39A, if not, returning to the step S37;

S39B, if yes, using the combined optimizing value as the optimal hyper-parameter set of the convolutional neural network;

the optimal hyper-parameter group comprises a global hyper-parameter and a local hyper-parameter;

the global hyper-parameter comprises: the method comprises the following steps of (1) learning rate, first moment estimation exponential decay rate, second moment estimation exponential decay rate, maximum iteration times, iteration stop threshold, batch processing scale and network weight initialization mode;

the local hyper-parameters include: convolution layer, batch normalization layer, nonlinear activation layer, pooling layer collocation mode and quantity, convolution filter size and number, filling type, pixel point number, convolution step, activation function type, pooling standard, pooling filter size and number, pooling filter step, full-connection layer activation function, full-connection layer number, neuron number of each layer and dropout probability;

in step S3, the training and obtaining the optimal self-optimized convolutional neural network-based barrel-projectile contact collision model in the optimization process of the hyper-parameters includes:

a31, setting different sets of hyperparameters, wherein the steps A32 to A36 are all carried out in the same hyperparameter set;

a32, extracting the data characteristics of the training set by adopting a forward propagation algorithm;

a33, measuring the estimated contact collision response and the actual value error of the convolutional neural network based on the mean square error loss function;

a34, judging whether a third convergence condition is met;

a35a, if the third convergence condition is not met, performing error back propagation by adopting an Adam gradient optimization algorithm, adjusting the network weight and bias, and repeating the step A33;

a35b, if a third convergence condition is met, terminating training to obtain a barrel-projectile contact collision prior model based on a convolutional neural network;

a36, evaluating the predicted performance of the prior model for new data by using the verification set;

a37, repeating the steps A32 to A36 for a plurality of times according to different super parameter groups to generate a plurality of barrel-projectile contact collision models based on different convolutional neural networks; the model with the optimal prediction performance on the verification set is the optimal barrel-projectile contact collision model based on the self-optimization convolutional neural network;

the mean square error loss function is:

the individual fitness evaluation index is as follows:

wherein y is _i Is a target value for the amount of time,

the Index is an individual fitness evaluation Index;

the first convergence condition is as follows: setting a maximum evolution algebra, and multiplying the maximum evolution algebra by the number of variables through 100 to obtain the maximum evolution algebra; or when the fitness of the population does not rise any more in 50 generations, taking the fitness difference less than or equal to 1e-6 as a judgment basis; or setting the total iteration number to 10000;

the second convergence condition is: step tolerance and function value tolerance are both smaller than a set threshold, and iteration is finished;

the third convergence condition is: stopping training when the MSE on the training set continuously decreases and the MSE on the verification set does not decrease and inversely increase; or when the MSE on the validation set no longer decreases for 20 consecutive times, stop training; or the maximum training frequency is 10000;

and S4, researching a contact collision response change rule in a variable parameter manner according to the self-optimization convolution neural network-based barrel-projectile contact collision model.

Images