Weapon combat effectiveness evaluation method based on small sample data generation based on CGAN

Abstract

This invention discloses a weapon combat effectiveness evaluation method based on small sample data generated by CGAN. The method includes obtaining a dataset, dividing it into a sample set and a test set, and training a CMGAN model using the test set; constructing a feature data model conversion network; building and training a CMGAN network model, including constructing a generator, constructing a discriminator, and training the CMGAN network; obtaining a virtual sample set under specific conditions; and obtaining effectiveness evaluation results. This invention can generate effective weapon combat effectiveness evaluation samples and improve the accuracy of weapon combat effectiveness evaluation under small sample conditions.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and in particular relates to a weapon combat effectiveness evaluation method based on small sample data generated by CGAN. Background Art

[0002] Generating virtual data samples is a mainstream approach to addressing the small sample size problem. This involves using specific methods to augment the dataset with small sample sizes. Traditional data augmentation methods primarily rely on oversampling. For example, the synthetic minority oversampling (SMOTE) method can be used to augment small sample sizes, improving detection accuracy. A comprehensive sampling method based on clustering structure properties can generate data samples with balanced sample distribution characteristics. These methods have achieved good results for small sample size evaluation and offer advantages such as strong interpretability and a simple mechanism. However, the features extracted from the generated samples are incomplete, which can easily lead to overfitting of the evaluation model. Generative adversarial networks (GANs) are generative models that have been widely used in data augmentation projects across various fields due to their powerful data generation capabilities. Using the original GAN ​​sample generation method and applying the generated samples to performance evaluation models has improved the accuracy of the performance evaluation models. However, the generated samples suffer from uncertain sample distribution and low sample quality, necessitating adaptive design based on the data distribution characteristics.

[0003] Traditional weapon performance evaluation relies on test data from weapon systems. Currently, the small sample size problem is primarily addressed through test data obtained through simulation. While this simulated test data can be used to increase the amount of test data and improve the performance of evaluation models, it contains a large amount of useless information, resulting in a waste of computing resources during the effectiveness evaluation phase. Redundant data also affects the accuracy of model evaluations. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a small sample data generation method and a weapon combat effectiveness evaluation method based on CGAN (conditional generative adversarial network). By utilizing the advantages of conditional generative network in convolution processing, the structure of conditional adversarial network is improved to achieve the purpose of generating effective virtual samples based on a small amount of real data.

[0005] The object of the present invention is achieved through the following technical solutions:

[0006] A weapon combat effectiveness evaluation method based on small sample data generation using CGAN, the method comprising:

[0007] Obtain a feature data set that affects combat effectiveness and perform data preprocessing to obtain a network input data set. Select several feature data as input variables, perform maximum and minimum value normalization on the data set, divide a portion of the data samples into a training set, and divide the remaining portion into a test set. Use the training set to train the CMGAN model.

[0008] Performing model conversion on the feature data in the training set to obtain a new data sample set after model conversion;

[0009] Build and train the CMGAN network model, including building the generator, building the discriminator, and training the CMGAN network; the CMGAN adversarial network consists of three parts: the generator G, the discriminator D, and the feature data model conversion network;

[0010] Obtain sparse categories in the performance evaluation samples, input the corresponding labels into the generator, and output a virtual sample set that meets the preset conditions;

[0011] A deep neural network is selected to train the model using the virtual sample set as network input, and the trained deep neural network uses the test set as input to obtain an efficiency evaluation result.

[0012] Furthermore, performing model conversion on the feature data in the training set to obtain a new data sample set after model conversion specifically includes:

[0013] Constructing a feature data model conversion network, wherein the feature data model conversion network includes a deconvolution layer, reconstructs a matrix shape, and outputs new sample data;

[0014] The feature data in the training set is converted into an n×n matrix form, where n is the number of feature data of the input variables. The network training process converts the feature data in the training set into a discriminator input data set.

[0015] Furthermore, the construction and training of the CMGAN network model includes constructing a generator, constructing a discriminator, and training the CMGAN network, specifically including:

[0016] Construct a 6-layer neural network of the generator, which is the first deconvolution layer, the connection layer, the second deconvolution layer, the third deconvolution layer, the fourth deconvolution layer and the fully connected layer in sequence;

[0017] Construct a 5-layer neural network of the discriminator, which is the first convolutional layer, the connection layer, the second convolutional layer, the third convolutional layer and the fourth convolutional layer in sequence;

[0018] Constructing a CGAN adversarial network, inputting the new sample data and the discriminator input data set into the discriminator of the CGAN adversarial network in sequence, training the discriminator, and updating the parameters of the generator to obtain a staged discriminator;

[0019] The label data and noise data in the training set are used as the input of the generator, and the output of the discriminator is used to update the parameters of the generator to obtain a staged generator;

[0020] The discriminator and generator are trained alternately to reach the number of training iterations, and finally a trained generator and a trained discriminator are obtained.

[0021] Furthermore, the construction generator specifically includes:

[0022] Construct the first deconvolution layer, the network input includes 100-dimensional noise data and conditional labels, and contains 256 filters;

[0023] Construct a connection layer to merge the output of the first deconvolution layer;

[0024] Construct the second deconvolution layer to deconvolve the input, which contains 256 filters;

[0025] Construct the third deconvolution layer to deconvolve the input, which contains 128 filters and uses ReLu as the activation function;

[0026] Construct the fourth deconvolution layer to deconvolve the input;

[0027] The activation function of the fully connected layer is constructed as Sigmoid, and the fake sample set is output.

[0028] Furthermore, the constructing of the discriminator specifically includes:

[0029] Construct the first convolutional layer. The network input consists of two parts: one is the new sample data and the discriminator input dataset, and the other is the conditional label. It contains 64 filters and performs deconvolution on the two parts of the input respectively.

[0030] Construct a connection layer that merges the two outputs of the first convolutional layer;

[0031] Construct the second convolutional layer to convolve the input with 256 filters;

[0032] Construct the third convolutional layer to convolve the input with 512 filters and ReLu activation function;

[0033] Construct the fourth convolutional layer, convolve the input, use the sigmoid activation function, and output the identification result.

[0034] Furthermore, the training of the CMGAN network specifically includes:

[0035] Construct the CMGAN adversarial network and set the maximum number of iterations and the initial value of the number of iterations for the adversarial network, discriminator, and generator;

[0036] Input the new sample data and the discriminator input data set into the discriminator of the adversarial network for forward propagation, obtain the adversarial loss, update the weight and bias of the discriminator, and obtain the discriminator after stage training when the number of iterations reaches the maximum value;

[0037] Input the noise data and label data into the generator of the adversarial network for forward propagation to obtain a supervised fake dataset;

[0038] The new sample data and the discriminator input data set are respectively input into the discriminator for forward propagation to obtain the adversarial loss, and the weights and biases of the generator are updated respectively until the number of iterations of the adversarial network reaches the maximum value.

[0039] Furthermore, the updating of the weights and biases of the discriminator specifically includes:

[0040]

[0041] Among them, θ D Represents the model parameters of the discriminator, L D represents the discriminator loss, λ D represents the discriminator learning rate, m is the sample size, and D is the discriminator.

[0042] Furthermore, the adversarial training loss of the adversarial network is:

[0043]

[0044] Among them, minmaxV(D,G) includes max V(D,G) and min V(D,G). Max V(D,G) is to optimize the discriminator D to distinguish the generated samples G(Z|Y), and min V(D,G) is to optimize the generator G so that the discriminator D cannot distinguish the generated samples G(Z|Y). is the expectation of the real data distribution, log D(X|Y) is the result of the discriminator’s discrimination of the real data, is the expectation of the noise data distribution, and log D(G(Z|Y)) is the discriminant result of the generated sample.

[0045] Furthermore, the weights and biases of the generator are updated as follows:

[0046]

[0047]

[0048] Among them, θG Represents the model parameters of the generator, L G represents the generator loss, λ G represents the generator learning rate, m is the sample size, and D is the discriminator.

[0049] Furthermore, the construction of the deep neural network specifically includes:

[0050] A deep neural network is built to encode the performance categories from 0 to 1. The deep neural network consists of an input layer, a hidden layer, an output layer, and a softmax function. The input layer consists of n neurons that affect the performance, which serves as the input vector. The hidden layer has two layers, each with 20 and 12 neurons respectively. The output layer consists of 4 neurons, corresponding to the number of performance categories. The output layer activation function uses the softmax function.

[0051] The neuron activation function of the network structure is the ReLU function, the loss function is the cross entropy, the iterative optimizer is Adam, and the connection weights and biases of each layer are initially randomly generated;

[0052] The evaluation index of the generalization performance of the model is the root mean square error:

[0053]

[0054] The beneficial effects of the present invention are:

[0055] (1) The present invention utilizes the advantages of the conditional generative network in convolution processing to improve the structure of the conditional adversarial network, thereby achieving the purpose of generating effective virtual samples based on a small amount of real data, and verifies in the subsequent performance evaluation model that the generated virtual samples also have a training effect on the model.

[0056] (2) The present invention performs well in the combat effectiveness evaluation task with small sample data. The generated virtual data set η value is close to 0, and the virtual sample quality is high. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flowchart of a method for evaluating weapon combat effectiveness based on small sample data generated by CGAN, provided by an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of the CMGAN network training process according to an embodiment of the present invention;

[0059] Figure 3 The embodiment of the present invention generates a characteristic graph of sample data under a specific label;

[0060] Figure 4 It is a JS scatter plot of the sample generated by the embodiment of the present invention and the test sample;

[0061] Figure 5 This is a result diagram of the real data training effectiveness evaluation model according to an embodiment of the present invention;

[0062] Figure 6 This is a result diagram of a data training effectiveness evaluation model generated by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0064] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0065] Traditional data sample generation methods have the problem that the features extracted from the generated samples are incomplete, which easily leads to problems such as overfitting of the evaluation model, uncertain sample distribution and low sample quality.

[0066] Traditional weapon performance evaluation methods mainly rely on test data obtained through simulation. These simulated test data can be used to increase the amount of test data and improve the performance of the evaluation model, but they contain a large amount of useless information, resulting in a waste of computing resources during the effectiveness evaluation stage. Redundant data also affects the accuracy of model evaluation.

[0067] In order to solve the above technical problems, the following embodiments of the weapon combat effectiveness evaluation method based on small sample data generated by CGAN are proposed.

[0068] Reference Figure 1 ,like Figure 1 The figure shows a flow chart of a weapon combat effectiveness evaluation method based on small sample data generated by CGAN provided in this embodiment. The specific steps of the method are as follows:

[0069] Step 1: Obtain the feature dataset D that affects combat effectiveness S1 , for D S1 Perform data preprocessing to obtain the network input data set D S2 ;

[0070] Select n feature data that will affect combat effectiveness as input variables, feature data set D S1 A total of m groups of sample data are used to normalize the data set to the maximum and minimum values. Take m1 data samples as training data DST2 , the remaining data points are used as test samples D STE2 . D ST2 Input the CMGAN model and train the model.

[0071] Step 2: The training data D ST2 Perform model conversion on the feature data in the model, and execute steps 2.1 to 2.2 respectively;

[0072] Step 2.1, constructing a feature data model conversion network;

[0073] The feature data model conversion network includes a deconvolution layer, which reconstructs the matrix shape and outputs the new sample data D SPV The deconvolution layer network does not have the ability to learn and is used to convert the generated virtual samples and real samples without any learning and training process.

[0074] Step 2.2, obtain the new data sample set DSP after model conversion;

[0075] D ST2 The feature data in the matrix is ​​converted into n×n matrix form. During the network training process, D ST2 The feature data is converted into dataset D SPR , and D in step 2.1 SPV as input to the discriminator.

[0076] Step 3: Build and train the CMGAN network model, and perform steps 3.1 to 3.3 in sequence.

[0077] The CMGAN adversarial network mainly consists of three parts: generator G, discriminator D and feature data model conversion network C.

[0078] Step 3.1. Build the generator

[0079] The generator G of the CGAN adversarial network mainly consists of 6 neural network layers, which are deconvolution layer DeConv1, connection layer ConCat2, deconvolution layer DeConv3, deconvolution layer DeConv4, deconvolution layer DeConv5 and fully connected layer Linear6.

[0080] Step 3.2: Build the Discriminator

[0081] The discriminator D of the CGAN adversarial network mainly has 5 neural network layers, which are convolutional layer Conv1, connection layer ConCat2, convolutional layer Conv3, convolutional layer Conv4 and convolutional layer Conv5.

[0082] Step 3.3: Train the CMGAN network

[0083] Construct CGAN adversarial network and transform D SPV and D SPR Input the discriminator D1 of the adversarial network in sequence, train the discriminator D1, update the parameters of the generator D1, and obtain the stage discriminator D1′; ST2 The label data and noise data Z in are used as the input of generator G1, and the parameters of generator G1 are updated with the output of the discriminator to obtain the staged generator G1′; D1 and G1 are trained alternately to reach the number of training iterations, and finally the generator G1″ and the discriminator D1″ are obtained.

[0084] The core operation of CMGAN training is to add conditional information to G and D. The input of generator G is the noise signal Z, and the category label comes from the dataset D. ST2 , can be combined with noise signals as latent space representation; the input of the discriminator D is the data after feature conversion (D SPR and D SPV ), it is also necessary to classify the label and D SP The datasets are concatenated as input to the discriminator.

[0085] Reference Figure 2 ,like Figure 2 The figure shows a characteristic graph of sample data under a specific label generated by this embodiment.

[0086] Step 4: Obtain a virtual sample set D under specific conditions VG ;

[0087] Obtain sparse categories in the performance evaluation samples, input the corresponding labels into the generator G1″, and output the virtual sample set D that meets the conditions through the fully connected layer VG This paper defines the indicator η to evaluate the distribution similarity between a small sample and the entire data. The distribution similarity of the datasets S1 and S2 is defined as follows:

[0088]

[0089] Where N represents the number of characteristic attributes of datasets S1 and S2; Respectively represent s1 i ∈S1 and s2 i ∈S2 probability distribution; JS represents the JS divergence of the attributes of the datasets S1 and S2, which is used here to measure the similarity of the probability distribution of the virtual sample and the real sample dataset. The smaller the value of η, the closer the two sample sets are, and the higher the overall quality of the sample. Figure 4 As shown, the sample data D generated this time VG The η value is close to 0, and the JS divergence value between each generated sample and the original test dataset is less than 0.0006, indicating that the generated virtual dataset is of high quality and can replace the original dataset in training.

[0090] Reference Figure 3 ,like Figure 3 The figure shows the JS scatter plot of the generated sample and the test sample in this embodiment, wherein the comprehensive evaluation coefficient of sample set similarity η is less than 0.00012.

[0091] Step 5: Obtain performance evaluation results;

[0092] Select the deep neural network (DNN) and use the original dataset D S1 As the network input, the model is trained. The trained DNN network DNN1 is tested with the sample D STE2 As output, the classification result R1 is obtained; the generated virtual sample D VG As the DNN network input, the model is trained. The trained DNN network DNN2 is tested with the sample D STE2 The performance evaluation result R2 is obtained as input, and the two performance evaluation results R1 and R2 are compared.

[0093] As an implementation method, step 3.1 of this embodiment is specifically as follows:

[0094] Step 3.1.1, construct the deconvolution layer DeConv1, the network input is two parts, one is 100-dimensional noise data Z, the other is the conditional label Y, containing 256 filters; construct the connection layer ConCat2, the connection layer merges the DeConv1 output; construct the deconvolution layer DeConv3, deconvolve the input, the filter is 256; construct the deconvolution layer DeConv4, deconvolve the input, the filter is 128, the activation function is ReLu; construct the deconvolution layer DeConv5, deconvolve the input; construct the fully connected layer Linear6, the activation function is Sigmoid, and output the false sample set D SV .

[0095] As an implementation method, step 3.2 of this embodiment is specifically as follows:

[0096] Step 3.2.1, construct the convolution layer Conv1, the network input is two parts, one part is the output data D of the feature data conversion network SPV 、D SPR, one part is the conditional label Y, which contains 64 filters and performs deconvolution on the two parts of the input respectively; constructs the connection layer ConCat2, which merges the two parts of the output in Conv1; constructs the convolution layer Conv3, which convolves the input with 256 filters; constructs the convolution layer Conv4, which convolves the input with 512 filters and the activation function is ReLu; constructs the convolution layer Conv5, which convolves the input with the activation function sigmoid and outputs the identification result.

[0097] As an implementation method, step 3.3 of this embodiment is specifically as follows:

[0098] Step 3.3.1. Build the CMGAN adversarial network and set the number of iterations of the adversarial network to K. 1 , set the number of iterations of the discriminator of the adversarial network to Set the number of iterations of the generator of the adversarial network to Set the number of iterations k of the adversarial network 1 The initial value k 1 =1, the number of iterations of the discriminator of the adversarial network Initial value of Number of generator iterations of the adversarial network Initial value of

[0099] Step 3.3.2, D SPR and D SPV Input them into the discriminator D1 of the adversarial network for forward propagation to obtain the adversarial loss and the weight W of the discriminator D1. D1 and bias b D1 Make updates;

[0100] Step 3.3.3, judgment Is it true? If so, get the discriminator D1′ after stage training and execute step 3.3.4; otherwise, let And return to step 3.3.2;

[0101] Step 3.3.4: Input the noise data Z and label data Y into the generator G1 of the adversarial network for forward propagation to obtain a supervised fake data set. D SPV and D SPR They are respectively input into the discriminator D1 for forward propagation to obtain the adversarial loss, and the weight W of the generator G1 is respectively G1 and bias b G1 Make updates; D SPV The i-th sample and its label in m Z D SPV Number of samples;

[0102] Step 3.3.5, judgment Is it true? If so, get the generator G1′ after stage training and execute step 3.3.6; otherwise, let And return to step 3.3.4;

[0103] Step 3.3.6, determine k 1 =K 1 Is it true? If so, get the generator G1″ and the discriminator D1″ and execute step 4; otherwise, let k 1 =k 1 +1, return to step 3.3.3.

[0104] As an implementation method, the parameter update in step 3.3.2 of this embodiment is:

[0105]

[0106] Among them, θ D Represents the model parameters of the discriminator, L D represents the discriminator loss, λ D represents the discriminator learning rate, m is the sample size, and D is the discriminator.

[0107] As an implementation method, in step 3.3.4 of this embodiment, the adversarial training loss of the adversarial network is:

[0108]

[0109] Where minmax V(D,G) includes max V(D,G) and min V(D,G). Max V(D,G) is to optimize the discriminator D to distinguish the generated sample G(z|Y), and min V(D,G) is to optimize the generator G so that the discriminator D cannot distinguish the generated sample G(Z|Y). The expectation of the real data distribution, log D(X|Y) is the result of the discriminator's discrimination of the real data, is the expectation of the noise data distribution, and log D(G(Z|Y)) is the discriminant result of the generated sample.

[0110] As an implementation method, step 3.3.4 of this embodiment updates the weights and biases of the generator as follows:

[0111]

[0112] Among them, θ G Represents the model parameters of the generator, L G represents the generator loss, λ G represents the generator learning rate, m is the sample size, and D is the discriminator.

[0113] As an implementation method, in step 3.3.1 of this embodiment, the number of iterations K 1 =1000,

[0114] As an implementation method, step 5 of this embodiment is specifically as follows:

[0115] The data set includes training samples and test samples. The training samples include the real data set D ST2 And the generated virtual sample D VG , the test samples include the real data set D STE2 . A deep neural network (DNN) is built to solve the problem of combat effectiveness evaluation, and the effectiveness categories are encoded in 0-1 (One-hot Encoding). The structure of the built DNN consists of an input layer, a hidden layer, an output layer, and a softmax function. The input layer consists of n neurons that affect the effectiveness as the input vector. The hidden layer has two layers, each with 20 and 12 neurons respectively. The output layer consists of 4 neurons, corresponding to the number of effectiveness categories. The output layer excitation function selects the softmax function to solve the multi-classification problem. The neuron activation function of the network structure is the ReLU function, the loss function is cross entropy, and the iterative optimizer selects Adam. Initially, the connection weights and biases of each layer are randomly generated. The root mean square error (RMSE) is used as the evaluation indicator of the generalization performance of the model, as follows:

[0116]

[0117] It is used here to evaluate the quality of the generated virtual sample sets in terms of quantity and generalization performance improvement.

[0118] Take the original data set D ST1 As the network input, the model is trained. The trained DNN network DNN1 is tested with the sample D STE2 As input, the classification result R1 is obtained, such as Figure 5 As shown. The generated virtual sample D VG As the DNN network input, the model is trained and the trained DNN network DNN2 is used as the test set sample D STE2 As input, the classification result R2 is obtained as Figure 6 As shown in Figure 2, comparing the two performance evaluation results R1 and R2, the model generalization performance of R2 is better.

[0119] The weapon combat effectiveness evaluation method based on small sample data generation using CGAN proposed in this embodiment is suitable for expanding combat data sample sets and plays a guiding role in reducing the cost of military weapon testing and improving the effectiveness evaluation of modern combat weapons.

[0120] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A weapon combat effectiveness evaluation method based on small sample data generation using CGAN, characterized by: The method comprises: Obtain a feature data set that affects combat effectiveness and perform data preprocessing to obtain a network input data set. Select several feature data as input variables, perform maximum and minimum value normalization on the data set, divide a portion of the data samples into a training set, and divide the remaining portion into a test set. Use the training set to train the CMGAN model. Performing model conversion on the feature data in the training set to obtain a new data sample set after model conversion; Build and train the CMGAN network model, including building the generator, building the discriminator, and training the CMGAN network; the CMGAN adversarial network consists of three parts: the generator G, the discriminator D, and the feature data model conversion network; Obtain sparse categories in the performance evaluation samples, input the corresponding labels into the generator, and output a virtual sample set that meets the preset conditions; Selecting a deep neural network to train the model using the virtual sample set as network input, and using the trained deep neural network with the test set as input to obtain an effectiveness evaluation result; The step of performing model conversion on the feature data in the training set to obtain a new data sample set after model conversion specifically includes: Constructing a feature data model conversion network, wherein the feature data model conversion network includes a deconvolution layer, reconstructs a matrix shape, and outputs new sample data; Converting the feature data in the training set into an n×n matrix, where n is the number of feature data of the input variables, and converting the feature data in the training set into a discriminator input data set during the network training process; The construction and training of the CMGAN network model includes constructing a generator, constructing a discriminator, and training the CMGAN network, specifically including: Construct a 6-layer neural network of the generator, which is the first deconvolution layer, the connection layer, the second deconvolution layer, the third deconvolution layer, the fourth deconvolution layer and the fully connected layer in sequence; Construct a 5-layer neural network of the discriminator, which is the first convolutional layer, the connection layer, the second convolutional layer, the third convolutional layer and the fourth convolutional layer in sequence; Constructing a CGAN adversarial network, inputting the new sample data and the discriminator input data set into the discriminator of the CGAN adversarial network in sequence, training the discriminator, and updating the parameters of the generator to obtain a staged discriminator; The label data and noise data in the training set are used as the input of the generator, and the output of the discriminator is used to update the parameters of the generator to obtain a staged generator; The discriminator and generator are trained alternately to reach the number of training iterations, and finally a trained generator and a trained discriminator are obtained.

2. The weapon combat effectiveness evaluation method based on small sample data generation using CGAN as claimed in claim 1, characterized in that: The construction generator specifically includes: Construct the first deconvolution layer, the network input includes 100-dimensional noise data and conditional labels, and contains 256 filters; Construct a connection layer to merge the output of the first deconvolution layer; Construct the second deconvolution layer to deconvolve the input, which contains 256 filters; Construct the third deconvolution layer to deconvolve the input, which contains 128 filters and uses ReLu as the activation function; Construct the fourth deconvolution layer to deconvolve the input; The activation function of the fully connected layer is constructed as Sigmoid, and the fake sample set is output.

3. The weapon combat effectiveness evaluation method based on small sample data generation using CGAN as claimed in claim 1, characterized in that: The constructing of the discriminator specifically includes: Construct the first convolutional layer. The network input consists of two parts: one is the new sample data and the discriminator input dataset, and the other is the conditional label. It contains 64 filters and performs deconvolution on the two parts of the input respectively. Construct a connection layer that merges the two outputs of the first convolutional layer; Construct the second convolutional layer to convolve the input with 256 filters; Construct the third convolutional layer to convolve the input with 512 filters and ReLu activation function; Construct the fourth convolutional layer, convolve the input, use the sigmoid activation function, and output the identification result.

4. The weapon combat effectiveness evaluation method based on small sample data generation using CGAN as claimed in claim 1, characterized in that: The training of the CMGAN network specifically includes: Construct the CMGAN adversarial network and set the maximum and initial iteration times of the adversarial network, discriminator, and generator. Input the new sample data and the discriminator input data set into the discriminator of the adversarial network for forward propagation, obtain the adversarial loss, update the weight and bias of the discriminator, and obtain the discriminator after stage training when the number of iterations reaches the maximum value; Input the noise data and label data into the generator of the adversarial network for forward propagation to obtain a supervised fake dataset; The new sample data and the discriminator input data set are respectively input into the discriminator for forward propagation to obtain the adversarial loss, and the weights and biases of the generator are updated respectively until the number of iterations of the adversarial network reaches the maximum value.

5. The weapon combat effectiveness evaluation method based on small sample data generation using CGAN as claimed in claim 4, characterized in that: The updating of the weights and biases of the discriminator specifically includes: Among them, θ D Represents the model parameters of the discriminator, L D represents the discriminator loss, λ D represents the discriminator learning rate, m is the sample size, and D is the discriminator.

6. The weapon combat effectiveness evaluation method based on small sample data generation using CGAN as claimed in claim 4, characterized in that: The adversarial training loss of the adversarial network is: Among them, minmaxV(D,G) includes maxV(D,G) and minV(D,G), max V(D,G) is to optimize the discriminator D to distinguish the generated samples G(Z|Y), and min V(D,G) is to optimize the generator G so that the discriminator D cannot distinguish the generated samples G(Z|Y); is the expectation of the real data distribution, logD(X|Y) is the result of the discriminator’s discrimination of the real data, is the expectation of the noise data distribution, and log D(G(Z|Y)) is the discriminant result of the generated sample.

7. The weapon combat effectiveness evaluation method based on small sample data generation using CGAN as claimed in claim 4, characterized in that: The weights and biases of the generator are updated as follows: Among them, θ G Represents the model parameters of the generator, L G represents the generator loss, λ G represents the generator learning rate, m is the sample size, and D is the discriminator.

8. The weapon combat effectiveness evaluation method based on small sample data generation using CGAN as claimed in claim 4, characterized in that: The construction of the deep neural network specifically includes: A deep neural network is built to encode the performance categories from 0 to 1. The deep neural network consists of an input layer, a hidden layer, an output layer, and a softmax function. The input layer consists of n neurons that affect the performance, which serves as the input vector. The hidden layer has two layers, each with 20 and 12 neurons respectively. The output layer consists of 4 neurons, corresponding to the number of performance categories. The output layer activation function uses the softmax function. The neuron activation function of the network structure is the ReLU function, the loss function is the cross entropy, the iterative optimizer is Adam, and the connection weights and biases of each layer are initially randomly generated; The evaluation index of the generalization performance of the model is the root mean square error:

Images