A Method for Analyzing the Behavioral Intent of Air Combat Electromagnetic Countermeasures Based on Small-Sample Learning

Abstract

The present invention provides a method for analyzing the behavioral intention of air combat electromagnetic countermeasure based on few-shot learning. A GRU behavioral intention recognition model based on the attention mechanism is designed, which includes three parts: a fully connected layer, a bidirectional GRU layer, and an attention layer. And a training method of few-shot contrastive learning model based on data augmentation is adopted. The original data is augmented by RWGAN, and the contrastive learning framework is used to mine rich pattern information in multi-modal data to make up for the shortage of few-shot data scale, so as to improve the accuracy of the model in predicting the behavioral intention of electromagnetic targets.

Description

Technical Field

[0001] The present invention belongs to the technical field of intention recognition, and particularly relates to a method for analyzing the behavior intention of air combat electromagnetic countermeasure based on few-shot learning. Background Art

[0002] The method based on deep neural network has shown great advantages in predicting the behavior intention of electromagnetic countermeasure targets. However, the real air combat environment is complex and changeable, it is difficult to capture a large amount of countermeasure data of enemy targets for model training, and there is a serious problem of class imbalance among samples in the training data. For this reason, the Adaptive Synthetic (ADASYN) technology synthesizes inter-class imbalanced samples, enabling the neural network to be trained under relatively balanced conditions, effectively improving the accuracy of air combat target intention recognition. However, there is a certain blindness in the nearest neighbor selection of the ADASYN technology, which cannot overcome the data distribution problem of the imbalanced dataset, and is prone to distribution marginalization. Moreover, existing research has been carried out under the condition that the number of samples in each category of the training dataset is sufficient and relatively balanced. There is still little research on intention recognition under the conditions of few training samples and imbalance, and it is necessary to design an intelligent intention recognition model that can mine useful information from few training samples and imbalanced data, combine the potential advantages of human reasoning mode and cognitive experience, and realize the intelligent, real-time analysis, reasoning and judgment of the target's tactical intention. Summary of the Invention

[0003] In order to overcome the deficiencies of the prior art, the present invention provides a method for analyzing the behavior intention of air combat electromagnetic countermeasure based on few-shot learning. A GRU behavior intention recognition model based on the attention mechanism is designed, which includes three parts: a fully connected layer, a bidirectional GRU layer and an attention layer, and a few-shot contrastive learning model training method based on data augmentation is adopted. The original data is augmented by RWGAN, and the contrastive learning framework is used to mine rich pattern information in multi-modal data to make up for the lack of few-shot data scale, thereby improving the accuracy of the model in predicting the behavior intention of electromagnetic targets.

[0004] A method for analyzing the behavior intention of air combat electromagnetic countermeasure based on few-shot learning, characterized by the following steps:

[0005] Step 1, Data cleaning: First, extract the feature vectors from the air combat confrontation electromagnetic behavior dataset that have no missing dimensions and do not contain extreme values, constituting the air situation feature set, including continuous-value type behavior information and discrete-value type electromagnetic state information; then, use the sliding window method to clip the behavior information to obtain several behavior sequences with a time length of 30s, and perform standardization processing using the mean-variance normalization method; finally, divide the behavior sequences representing the same air battlefield advantages and disadvantages after expert annotation into one category, constituting an electromagnetic target tactical intention space with 5 categories: attack, break away, occupy a position, evade, and approach the enemy. Concatenate the behavior sequences belonging to the same category along the time dimension to obtain a behavior pattern sequence. Represent the different categories in the tactical intention space in the form of one-hot vectors to obtain the pattern labels corresponding to the behavior pattern sequence, and randomly divide the behavior pattern sequence into a training set, a validation set, and a test set according to the ratio of 8:1:1;

[0006] Step 2, Generate fake training set behavior sequences: Replace all linear layers in the WGAN with GRUs to obtain the generative model RWGAN, which consists of a generator and a discriminator; use the training set data obtained in Step 1 as the real data, and train the generator and discriminator of the model RWGAN in an unsupervised manner. Use the trained generator of RWGAN to generate fake training set behavior sequences with the same distribution characteristics as the real data;

[0007] Step 3, Construct a GRU behavior intention recognition model based on the attention mechanism: It includes three parts: a fully connected layer, a bidirectional GRU layer, and an attention layer. Among them, the fully connected layer takes the behavior sequence as input, reduces the dimension and extracts features of the behavior sequence to obtain behavior sequence information; the bidirectional GRU layer consists of two bidirectional GRU units and a feature enhancement layer based on the attention mechanism. The behavior sequence information output by the fully connected layer learns the temporal context information at the signal level through the first bidirectional GRU unit to obtain the behavior sequence information at the signal level. At the same time, use the feature enhancement layer based on the attention mechanism to mine the feature information of the electromagnetic state information to obtain the electromagnetic state guidance information, and then concatenate the electromagnetic state guidance information with the behavior sequence information at the signal level to obtain the attention-enhanced mixed feature information. The attention-enhanced mixed feature information is input into the second GRU unit, and the output is the behavior sequence information at the semantic level; the attention layer includes a linear layer and a softmax layer. The behavior sequence information at the semantic level is input into the attention layer, and the output is the intention recognition result;

[0008] Step 4, Contrastive Learning Model Training: First, define the encoder \(E_q\) and the memory encoder \(E_k\). The encoder \(E_q\) and the memory encoder \(E_k\) have the same network structure, both using the GRU intent recognition model based on the attention mechanism constructed in Step 3. Remove its attention layer, randomly initialize the two encoders and keep their parameters consistent. At the same time, combine the behavior sequences of the training set obtained in Step 1 and the fake behavior sequences of the training set generated in Step 2 as the training dataset for contrastive learning. Then, freeze the parameters of the memory encoder \(E_k\), randomly extract a behavior sequence \(x_q\) from the training dataset for contrastive learning, and perform data augmentation on \(x_q\) to obtain the behavior sequence \(x_k\). Use \(x_k\) as the positive sample of \(x_q\), and use the other behavior sequences in the training set as the negative samples of \(x_q\). Input the positive and negative samples into the encoder \(E_q\) to output the feature \(q\), and input the positive and negative samples into the memory encoder \(E_k\) to output the feature \(k\). Set the loss function as follows:

[0009]

[0010] where \(t\) represents the temperature hyperparameter, which is a constant greater than 0, \(N\) represents the number of negative samples, \(k\) i represents the feature of the \(i\)-th negative sample, and \(\exp(\cdot)\) represents the cosine distance;

[0011] Finally, use the backpropagation algorithm to update the parameters of the encoder \(E_q\), and use the momentum update method to update the parameters of the memory encoder \(E_k\). The specific update expressions are as follows:

[0012] \(\theta\) k \(\leftarrow m\theta\) k +(1 - m)\(\theta\) q (2)

[0013] where \(m\in[0, 1]\) is the momentum coefficient, \(\theta\) k represents the parameters of the memory encoder \(E_k\), and \(\theta\) q represents the parameters of the encoder \(E_q\);

[0014] Step 5, Model Parameter Tuning: First, use the transfer learning method to transfer the parameters of the fully connected layer and the bidirectional GRU layer in the model trained in Step 4 to the GRU behavior intent recognition model based on the attention mechanism, and randomly initialize the weights of the attention layer. Then, use the behavior sequences and pattern labels of the training set obtained in Step 1 as the input data and supervision information respectively to train the GRU behavior intent recognition model based on the attention mechanism. During training, use the cross-entropy loss function, and use the data of the validation set obtained in Step 1 to evaluate the effect of the model. Furthermore, adjust the hyperparameters of the model, including the dimension of the fully connected layer, the number of GRU layers in the bidirectional GRU layer, and the dimension size of the output feature vector, to obtain the optimal model;

[0015] Step 6, model application: Input the behavior pattern sequence of the test set into the optimal attention mechanism-based GRU behavior intention recognition model obtained in step 5, and output the intention recognition result.

[0016] The beneficial effects of the present invention are as follows: since the constructed GRU intention recognition model based on the attention mechanism adopts a bidirectional GRU structure, it is possible to mine the global temporal correlation of feature data, thereby enhancing the anti-interference ability of the intention recognition network and improving the effectiveness and accuracy of predicting enemy behavior intentions in real battlefields and various unknown scenarios; since a sample data expansion strategy based on RWGAN is adopted, it is possible to overcome the problems of certain blindness and easy distribution marginalization in the existing data expansion methods when expanding data, and has the advantage of being able to learn a wider range of features between samples, thereby being able to generate diverse data with sample distribution characteristics; since the network is first trained using a contrastive learning framework, by discriminating the difference between positive and negative samples in the input sample pair, the network is able to learn the discriminative features of the samples, and then the model parameters are migrated and the model is retrained to fine-tune the parameters, so that the model has better generalization performance, avoids the model from falling into a local optimum, and can effectively improve the accuracy of the model's intention recognition classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the air combat electromagnetic confrontation behavior intention analysis method based on small sample learning of the present invention. DETAILED DESCRIPTION

[0018] The present invention is further described below in conjunction with the accompanying drawings and embodiments. The present invention includes but is not limited to the following embodiments.

[0019] like Figure 1As shown, the present invention provides a method for analyzing the behavioral intention of air combat electromagnetic countermeasure based on few-shot learning, aiming to accurately identify the behavioral intention of air electromagnetic countermeasure targets and obtain the temporal state data of the targets at multiple moments. First, data cleaning is used to generate a behavioral pattern sequence, and a sample augmentation strategy based on RWGAN (Recurrent Wasserstein Generative Adversarial Network) is used to augment the unbalanced air combat countermeasure data between classes, generating high-quality fake data with the same distribution as the original data. Then, the generated fake data and the training set of real countermeasure data are combined into the training data for contrastive learning, and a GRU intention recognition model (Gated Recurrent Unit) based on the attention mechanism is trained using the few-shot contrastive learning framework, enabling the model to learn the differences between different class samples and saving the pre-trained weights. Next, using the method of transfer learning, the pre-trained weights are passed to the GRU behavioral intention recognition network based on the attention mechanism, and the network is retrained using the training set data of the original data, fine-tuning the network weights, and finding the optimal hyperparameters of the model through the validation set to find the optimal model weights and hyperparameter settings. Finally, the input test set behavioral pattern sequence is input into the optimal model, and the result of intention recognition can be quickly output.

[0020] The specific implementation process of the present invention is as follows:

[0021] Step 1, data cleaning: First, feature vectors that do not have missing dimensions and do not contain extreme sizes are extracted from the air combat countermeasure electromagnetic behavior dataset to form an air situation feature set, including continuous-value type behavioral information and discrete-value type electromagnetic state information. In this embodiment, the air combat electromagnetic countermeasure behavior dataset is selected as the experimental dataset, which is provided by the Key Laboratory of Electromagnetic Space Operations and Applications of the 29th Research Institute of China Electronics Technology Group Corporation. This dataset contains multi-element temporal data of 6513s in 5 behavioral categories, namely attack, evasion, approaching, breaking away, and occupying positions. The labels of all data are marked by experts. The data of each behavioral category has a high dimension, small samples, and unbalanced distribution, posing challenges to fully mining the effective information in the data. The proportion of each behavioral category in the data is as follows: attack: 4%, evasion: 10%, breaking away: 15%, approaching: 32%, occupying positions: 39%.

[0022] Then, the sliding window method is used to clip the behavior information to obtain several behavior sequences with a time length of 30s, and the mean-variance normalization method is used for normalization processing; finally, the behavior sequences representing the same advantages and disadvantages of the empty battlefield after expert annotation are grouped into one category, constituting an electromagnetic target tactical intention space containing 5 categories: attack, break away, occupy a position, avoid, and engage the enemy. The behavior sequences belonging to the same category are concatenated along the time dimension to obtain a behavior pattern sequence. The different categories of the tactical intention space are represented in the form of one-hot vectors to obtain the pattern labels corresponding to the behavior pattern sequence, and the behavior pattern sequence is randomly divided into a training set, a validation set, and a test set according to the ratio of 8:1:1.

[0023] Step 2, generate fake training set behavior sequences: Replace all linear layers in the WGAN with GRUs to obtain the generative model RWGAN, which consists of a generator and a discriminator; use the training set data obtained in Step 1 as the real data, and train the generator and discriminator of the model RWGAN in an unsupervised manner. Use the trained generator of RWGAN to generate fake training set behavior sequences with the same distribution characteristics as the real data.

[0024] Step 3, construct a GRU behavior intention recognition model based on the attention mechanism: It includes three parts: a fully connected layer, a bidirectional GRU layer, and an attention layer. Among them, the fully connected layer takes the behavior sequence as input, reduces the dimension and extracts features of the behavior sequence to obtain behavior sequence information; the bidirectional GRU layer consists of two bidirectional GRU units and a feature enhancement layer based on the attention mechanism. The behavior sequence information output by the fully connected layer learns the temporal context information at the signal level through the first bidirectional GRU unit to obtain the behavior sequence information at the signal level. At the same time, use the feature enhancement layer based on the attention mechanism to mine the feature information of the electromagnetic state information to obtain the electromagnetic state guidance information, and then concatenate the electromagnetic state guidance information with the behavior sequence information at the signal level to obtain the attention-enhanced mixed feature information. The attention-enhanced mixed feature information is input into the second GRU unit, and the output obtains the behavior sequence information at the semantic level; the attention layer includes a linear layer and a softmax layer. The behavior sequence information at the semantic level is input into the attention layer, and the output obtains the intention recognition result.

[0025] Step 4, Contrastive Learning Model Training: First, define the encoder \(E_q\) and the storage encoder \(E_k\). The encoder \(E_q\) and the storage encoder \(E_k\) have the same network structure, both using the GRU intent recognition model based on the attention mechanism constructed in Step 3. Remove its attention layer, randomly initialize the two encoders and keep their parameters consistent. At the same time, merge the behavior sequences of the training set obtained in Step 1 and the fake training set behavior sequences generated in Step 2 as the training dataset for contrastive learning. Then, freeze the parameters of the storage encoder \(E_k\). Randomly extract a behavior sequence \(x_q\) from the training dataset for contrastive learning, and perform data augmentation on \(x_q\) to obtain the behavior sequence \(x_k\). Use \(x_k\) as the positive sample of \(x_q\), and use the other behavior sequences in the training set as the negative samples of \(x_q\). Input the positive and negative samples into the encoder \(E_q\) to output the feature \(q\), and input the positive and negative samples into the storage encoder \(E_k\) to output the feature \(k\). Set the loss function as follows:

[0026]

[0027] where \(t\) represents the temperature hyperparameter, which is a constant greater than 0, \(N\) represents the number of negative samples, and \(k\) i represents the feature of the \(i\)-th negative sample, and \(\exp(\cdot)\) represents the cosine distance.

[0028] Finally, use the backpropagation algorithm to update the parameters of the encoder \(E_q\), and use the momentum update method to update the parameters of the storage encoder \(E_k\). The specific update expressions are as follows:

[0029] \(\theta\) k \(\leftarrow m\theta\) k +(1 - m)\(\theta\) q (4)

[0030] where \(m\in[0, 1]\) is the momentum coefficient, \(\theta\) k represents the parameters of the storage encoder \(E_k\), and \(\theta\) q represents the parameters of the encoder \(E_q\).

[0031] Step 5, Model Parameter Adjustment: First, use the method of transfer learning to transfer the parameters of the fully connected layer and the bidirectional GRU layer in the model trained in Step 4 to the GRU behavior intent recognition model based on the attention mechanism, and randomly initialize the weights of the attention layer. Then, use the behavior sequences and pattern labels of the training set obtained in Step 1 as the input data and supervision information respectively to train the GRU behavior intent recognition model based on the attention mechanism. During training, use the cross-entropy loss function, and use the data of the validation set obtained in Step 1 to evaluate the effect of the model, and then adjust the hyperparameters of the model, including the dimension of the fully connected layer, the number of GRU layers in the bidirectional GRU layer, and the dimension size of the output feature vector, to obtain the optimal model.

[0032] Step 6, model application: Input the behavior pattern sequence of the test set into the optimal attention mechanism GRU behavior intention recognition model obtained in Step 5, and output the intention recognition result.

Claims

1. A method for analyzing the behavioral intention of air combat electromagnetic countermeasure based on few-shot learning, characterized in that The steps are as follows: Step 1, data cleaning: First, extract the feature vectors that do not have dimension missing and do not contain extreme values from the air combat confrontation electromagnetic behavior dataset to form the air situation feature set, including continuous-value type behavior information and discrete-value type electromagnetic state information; then, use the sliding window method to clip the behavior information to obtain several behavior sequences with a time length of 30s, and perform standardization processing using the mean-variance standardization method; finally, divide the behavior sequences representing the same air battlefield advantages and disadvantages after expert annotation into one category to form an electromagnetic target tactical intention space with 5 categories: attack, break away, occupy a position, avoid, and engage the enemy. Concatenate the behavior sequences belonging to the same category along the time dimension to obtain the behavior pattern sequence, represent the different categories of the tactical intention space in the form of one-hot vectors to obtain the pattern labels corresponding to the behavior pattern sequence, and randomly divide the behavior pattern sequence into a training set, a validation set, and a test set according to the ratio of 8:1:1; Step 2, generate false training set behavior sequences: Replace all linear layers in the WGAN with GRUs to obtain the generative model RWGAN, which consists of a generator and a discriminator; use the training set data obtained in Step 1 as the real data, and train the generator and discriminator of the model RWGAN in an unsupervised form, and use the trained generator of RWGAN to generate false training set behavior sequences with the same distribution characteristics as the real data; Step 3, construct a GRU behavior intention recognition model based on the attention mechanism: It includes three parts: a fully connected layer, a bidirectional GRU layer, and an attention layer. Among them, the fully connected layer takes the behavior sequence as the input, reduces the dimension and extracts features of the behavior sequence to obtain the behavior sequence information; the bidirectional GRU layer consists of two bidirectional GRU units and a feature enhancement layer based on the attention mechanism. The behavior sequence information output by the fully connected layer learns the time-series context information at the signal level through the first bidirectional GRU unit to obtain the behavior sequence information at the signal level. At the same time, use the feature enhancement layer based on the attention mechanism to mine the feature information of the electromagnetic state information to obtain the electromagnetic state guidance information, and then concatenate the electromagnetic state guidance information with the behavior sequence information at the signal level to obtain the attention-enhanced mixed feature information. The attention-enhanced mixed feature information is input into the second GRU unit, and the behavior sequence information at the semantic level is output; the attention layer includes a linear layer and a softmax layer. The behavior sequence information at the semantic level is input into the attention layer, and the intention recognition result is output; Step 4, Contrastive Learning Model Training: First, define the encoder \(E_q\) and the storage encoder \(E_k\). The encoder \(E_q\) and the storage encoder \(E_k\) have the same network structure, both using the attention mechanism-based GRU intent recognition model constructed in Step 3. Remove its attention layer, randomly initialize the two encoders and keep their parameters consistent. At the same time, merge the behavior sequences of the training set obtained in Step 1 and the fake training set behavior sequences generated in Step 2 as the training dataset for contrastive learning. Then, freeze the parameters of the storage encoder \(E_k\), randomly extract a behavior sequence \(x_q\) from the training dataset for contrastive learning, and perform data augmentation on \(x_q\) to obtain the behavior sequence \(x_k\). Use \(x_k\) as the positive sample of \(x_q\), and use the other behavior sequences in the training set as the negative samples of \(x_q\). Input the positive and negative samples into the encoder \(E_q\) to output the feature \(q\), and input the positive and negative samples into the storage encoder \(E_k\) to output the feature \(k\). Set the loss function as follows: where t represents the temperature hyperparameter, which is a constant greater than 0, N represents the number of negative samples, and k i represents the feature of the i-th negative sample, and exp(·) represents the cosine distance; Finally, use the backpropagation algorithm to update the parameters of the encoder \(E_q\), and use the momentum update method to update the parameters of the storage encoder \(E_k\). The specific update expressions are as follows: θ k ←mθ k +(1 - m)θ a (2) where \(m\in[0,1]\) is the momentum coefficient, \(\theta\) k represents the parameters of the storage encoder \(E_k\), \(\theta\) q represents the parameters of the encoder \(E_q\); Step 5, Model Parameter Adjustment: First, use the method of transfer learning to transfer the parameters of the fully connected layer and the bidirectional GRU layer in the model trained in Step 4 to the attention mechanism-based GRU behavior intent recognition model, and randomly initialize the weights of the attention layer. Then, use the behavior sequences and pattern labels of the training set obtained in Step 1 as input data and supervision information respectively to train the attention mechanism-based GRU behavior intent recognition model. During training, use the cross-entropy loss function, and use the data of the validation set obtained in Step 1 to evaluate the effect of the model, and then adjust the hyperparameters of the model, including the dimension of the fully connected layer, the number of GRUs in the bidirectional GRU layer, and the dimension size of the output feature vector, to obtain the optimal model; Step 6, Model Application: Input the behavior pattern sequences of the test set into the optimal attention mechanism-based GRU behavior intent recognition model obtained in Step 5, and output the intent recognition result.

Images