Lightweight intention recognition method based on diffusion model

Through the lightweight intention recognition method based on diffusion model, the problem of small amount of data and imbalance in tactical intention recognition is handled, which improves the recognition speed and accuracy of the model, and solves the problems of poor generalization and overfitting in the existing technology. , achieving better robustness and recognition effects.

CN119026033APending Publication Date: 2024-11-26AIR FORCE UNIV PLA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410530504.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-04-29
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

In the prior art, the tactical intention recognition model has poor generalization due to small data volume and unbalanced data sets, the model complexity increases, the recognition reaction speed is slow, it is easy to overfit, and the recognition results are biased towards the party with a larger sample size. .

Method used

Using a lightweight intention recognition method based on diffusion model, learn the long-term trend of time series data through the GRU module, and process the original samples in combination with the WDiffusion diffusion model, generate data sets with sufficient training samples and balanced categories, and pass through multi-layer GRU and fully connected layers. Processing, and finally applying softmax function to perform multi-category classification.

Benefits of technology

It improves the recognition speed and accuracy of the model, improves generalization and robustness, and the generated time series data is ahead of the adversarial generation network and its variants in indicators such as KL divergence, JS divergence, MSE, and MAE. Ability to mine time series data differences more accurately, and the identification results perform well in terms of diversity and fidelity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026033A_ABST
    Figure CN119026033A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of intention recognition, and provides a lightweight intention recognition method based on a diffusion model, which comprises the following steps of battlefield environment recognition, air battlefield environment key feature extraction, comparative analysis, association and reasoning. Time sequence characteristic acquisition: acquiring time sequence characteristic data; preprocessing the data and inputting the data into a GRU module; according to the method, the difference between two pieces of time series data can be more accurately mined through the diffusion model, and the generated time series data are ahead of the current generative adversarial network and variants thereof in four indexes of KL divergence, JS divergence MSE and MAE; the GRU is applied to a new data set, the parameter quantity of the model is far lower than that of a current intention recognition model, and the indexes such as the recognition speed, the accuracy and the F1 score exceed those of the current intention recognition model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of intention recognition, and specifically is a lightweight intention recognition method based on a diffusion model. Background Art

[0002] Target tactical intention recognition plays an important role in battlefield situation awareness. Its essence is pattern recognition under dynamic confrontation conditions. Intent recognition requires combining professional knowledge and relevant experience, considering key information such as battlefield environment, target attributes, and target status, and performing highly abstract and complex thinking activities such as key feature extraction, comparative analysis, and association reasoning to accurately identify target intentions. However, due to the lack of actual combat scenarios, intent recognition often faces problems such as small data volume and unbalanced data sets, resulting in poor generalization and lightweight models for tactical intent recognition.

[0003] In the existing technology, stacking different modules will not only increase the complexity of the model and slow down the recognition response speed, but also often cause the model training to overfit, with poor generalization and robustness, and the recognition results will be biased towards the side with a larger sample size.

[0004] To this end, those skilled in the art have proposed a lightweight intent recognition method based on a diffusion model to solve the problems raised by the background technology. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a lightweight intent recognition method based on a diffusion model to solve the problems in the prior art that stacking different modules not only increases the complexity of the model and slows down the recognition response speed, but also often causes overfitting of the model training, poor generalization and robustness, and the recognition results are biased towards the side with a larger sample size.

[0006] A lightweight intention recognition method based on a diffusion model includes the following steps: S1, battlefield environment recognition, extraction, comparative analysis, association and reasoning of key features of the air battlefield environment;

[0007] S2, time series feature collection, obtaining time series feature data;

[0008] S3, data preprocessing and input into the GRU module, the GRU module is responsible for learning the long-term trend of time series data and extracting features. The GRU module network is trained on the generated data set to learn the long-term change trend and internal features of the air combat intention time series data;

[0009] S301, inputting the original data set with unbalanced samples and small data volume into the WDiffusion diffusion model to obtain a data set with sufficient training samples and balanced categories;

[0010] S302, processing the time series data containing different air combat intention features through two GRU layers, with the number of neurons in each layer being 64;

[0011] S303, the output result of the GRU layer is connected through a fully connected layer to obtain an output value h;

[0012] S304, applying a softmax function to perform multi-category classification, wherein the number of softmax layers is equal to the number of output layers, that is, the total number of categories of air combat intention labels;

[0013] y = softmax(wh+b);

[0014] Where y is the intention result predicted by the model, w is the weight matrix to be learned, and b is the bias vector.

[0015] Preferably, S1 further comprises the following steps:

[0016] S101. For numerical features, since the units are different, in order to eliminate their influence, Min-Max normalization is used to process and map the data to the range of [0,1].

[0017] For the i-th numerical data X i =(x i1 ,x i2 ,…,x ik ,…,x in )(i=1,2,…,7), n is the number of data of the i-th feature;

[0018]

[0019] Where maxX i Represents a feature X i The maximum value, minX i Represents a feature X i The minimum value of x′ ik is the original data x ik The result of mapping to the interval [0,1];

[0020] S102. Encoding non-numeric data can convert it into numerical data to facilitate processing and analysis in the model;

[0021] S103, construct a combat intent feature set, label each sample, and use a deep learning network for training; thereby establishing a mapping relationship between the combat intent type space set and the time series feature data set; the mapping function can be expressed as:

[0022]

[0023] Where Ω is the combat intention output by the model, θ K From t1 to t k The time series feature set composed of K consecutive moments, t i Characteristics of the moment.

[0024] According to the expert knowledge in the field of air combat and the air combat experience of pilots, various combat intentions are mainly related to the state parameters such as the relative distance of the target, flight speed, flight altitude, azimuth, heading angle, flight acceleration, etc.; the target intention is finally realized through maneuvers, and the maneuver type can reflect the movement information of the target and is closely related to the combat intention; Tables 1 to 5 list the corresponding relationship between these characteristic state parameters and the relative distance of the target, maneuver type, altitude, speed, heading angle and combat intention; all characteristic spaces of the target are shown in Figure 2 ;

[0025]

[0026] Table 1 Relationship between target relative distance and its combat intention

[0027]

[0028] Table 2 Relationship between target maneuver type and its combat intent

[0029]

[0030] Table 3 Relationship between target altitude and its combat intention

[0031]

[0032]

[0033] Table 5 Relationship between target heading angle and its combat intention

[0034] Preferably, the WDiffusion diffusion model is a generative model that uses variational inference to train a parameterized Markov chain, gradually transforming the latent variables x from a simple distribution T Generate data x0 in complex distribution; where T represents the total diffusion step length; let is the result of each diffusion time step t∈{1,2,…,T}, x T and x0 have the same dimension d. The diffusion model includes the diffusion process and the reverse process. θ where θ represents the learnable components of the conditional noise predictor and low-resolution encoder in the diffusion model; the posterior q(x1,…,x T |x0) diffusion process transforms the distribution q(x0) into a latent variable q(x T) and fixed as a Markov chain, which is based on the variance table β1,…,β T Gradually add Gaussian noise ε to the data;

[0035]

[0036]

[0037] In the formula, β t is a small positive constant, which is regarded as a constant hyperparameter; let α t :=1-β t , The diffusion process allows sampling x at any time step t t , which is in closed form:

[0038]

[0039] Further reparameterization is:

[0040]

[0041] The reverse process transforms the latent variable distribution p θ (xT) is transformed into a data distribution p parameterized by θ θ (x0), which is formed by p(x T )=N(x T ; 0,I) is the definition of the Markov chain with Gaussian transition at the beginning;

[0042]

[0043] p θ (x t-1 ∣x t ):=N(x t-1 ;μ θ (x t ,t),σ θ (x t ,t) 2 I) (8);

[0044] During the training phase, the variational lower bound (ELBO) on the negative log-likelihood is maximized, and KL divergence and variance reduction are introduced;

[0045]

[0046] Transformation requires p θ (x t-1 ∣x t ) is directly compared with its corresponding diffusion process posteriors;

[0047] set up You can get:

[0048]

[0049] Equations (5), (8) and (10) guarantee all KL divergences in the equations, and equation (9) is a comparison between Gaussian functions;

[0050] When t>1, let C is a constant, we get:

[0051]

[0052] The training process minimizes a variant of ELBO with inputs x0 and t:

[0053]

[0054] In the formula, ε θ For the noise prediction module; sample an x T ~N(x T ; 0,I) Then sample x according to formula (8) t-1 ~p θ (x t-1 ∣x t );in:

[0055]

[0056]

[0057] The Wasserstein distance is defined as follows:

[0058]

[0059] In the formula, Π(P r ,P g ) is P r and P g The set of all possible joint distributions formed by the combination; for each possible joint distribution γ, sample the real sample x and generate the sample y from (x, y) ~ γ, and calculate the distance between the two samples ||xy||; calculate the expected value E of the distance between the two samples under all possible joint distributions (x,y)~γ [||xy||]; Wasserstein distance is defined as the lower bound of the expected value All possible joint distributions of .

[0060] Preferably, the GRU module has two inputs, which are the output state h at the previous moment and t-1 and the input sequence value x at the current moment t , the output is the current state ht ; It resets the gate r t and update gate z t To update the state of the model; reset the gate to control the degree to which historical state information is forgotten, so that the network can discard unimportant information; update the gate z t Control the ratio of the previous state information to the current state to help the network remember long-term information; the calculation formula is as follows:

[0061]

[0062] σ represents the sigmoid activation function, which is used to convert the intermediate state to the range of [0,1]; h t-1 and h t Respectively represent the output states at time t-1 and time t; x t Represents the input sequence value at time t; is a candidate output state; W r , W z , U r , U z and They represent the weight coefficient matrices corresponding to each part respectively; tanh represents the hyperbolic tangent function; “⊙” represents the Hadamard product of the matrix (element-by-element multiplication).

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. The present invention can more accurately mine the difference between two time series data through the diffusion model. The generated time series data is ahead of the current generative adversarial network and its variants in four indicators: KL divergence, JS divergence MSE, and MAE. When GRU is applied to the new data set, the number of model parameters is much lower than the current intent recognition model, and the recognition speed, accuracy, F1 score and other indicators exceed the current intent recognition model.

[0065] 2. The present invention applies the diffusion model to the multi-classification and multi-variable time series generation task and verifies its effectiveness in the generation task; the improved diffusion model can more accurately mine the difference between two time series data, and the generated time series data performs well in terms of realism and diversity; the results of the lightweight model exceed the current intent recognition model in recognition speed and accuracy, and the model can better identify and offset noise, thereby greatly improving robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is the overall framework diagram of the model of the present invention;

[0067] Figure 2 is the target feature space map;

[0068] Figure 3 It is a process diagram of the diffusion model;

[0069] Figure 4 This is the GRU module structure diagram;

[0070] Figure 5 This is the optimization comparison experiment diagram;

[0071] Figure 6 This is a graph of the learning rate comparison experiment;

[0072] Figure 7 This is a graph of the regularization parameter comparison experiment;

[0073] Figure 8 Comparison chart of accuracy and loss value of WDiffusion-GRU model;

[0074] Fig. 9 This is the confusion matrix diagram of WDiffusion-GRU recognition results;

[0075] Fig.10 Identify performance comparison graphs for different timing methods;

[0076] Fig.11 This is a comparison chart of the effects of different generation methods. DETAILED DESCRIPTION

[0077] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0078] Example: As attached Figure 1 To Attachment Fig.11 As shown: The present invention provides a lightweight intention recognition method based on a diffusion model. The original data comes from the Air Combat Maneuvering Generator (ACMG), from which 3520 air combat intention samples are obtained. Each sample collects 12 consecutive frames of information, including 12 dimensional feature information such as heading angle, flight altitude, interference status, radar status, etc., with a total of 12 time steps.

[0079] In order to generate intention labels, we rely on the intention recognition rules written by air combat experts based on their air combat experience to classify the sample data; the experts train the neural network and label the data by analyzing the simulation system data and simulating the recognition of actual combat intentions; although the combat process is subjective, the accuracy of the annotations is guaranteed by the experts' rich experience and practical operations; the model proposed in this paper strictly follows the experts' judgment results, can be applied in actual combat in theory, and provide accurate auxiliary information for the commander's command decision; the data set includes 7 types of target tactical intentions, namely attack intention, penetration intention, reconnaissance intention, surveillance intention, feint intention, interference intention and retreat intention; the proportion of these tactical intentions in the data set is retreat intention Figure 8 .71%, reconnaissance intention Figure 4 .95%, attack intention 14.3%, interference intention 28.4%, alert intention 28.3%, penetration intention Figure 8 .29%, feint Figure 6 .71%

[0080] We used the above dataset to train the diffusion model and generated 89,600 samples, with each intent accounting for 14.3%. The sample set was then divided into a training set and a test set in a ratio of 8:2, with the training set sample size being 71,680 and the test set sample size being 17,920. During the experiment, sample data with ambiguous intent classification were revised to ensure the accuracy and reliability of the dataset.

[0081] In order to evaluate the performance of the Dif-GRU aerial target combat intention recognition model proposed in this paper, five indicators are used to evaluate the classification situation, namely; Accuracy: the ratio of the number of correctly classified samples to the total number of samples; Loss: the loss function value of the model during training; Precision: the ratio of the number of samples correctly predicted as positive categories to the number of all samples predicted as positive categories; Recall: the ratio of the number of samples correctly predicted as positive categories to the number of all true positive category samples; F1-score: the harmonic mean of precision and recall, which comprehensively considers the performance of both; KL divergence (Kullback-Leibler divergence): measures the degree of difference between two distributions; JS divergence (Jensen–Shannon divergence): measures the similarity of two probability distributions; Mean Square Error (MSE): measures the prediction value The degree of matching with certain true values; Mean Absolute Error (MAE): reflects the actual situation of the predicted value error; their calculation formulas are:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091] In the above formula, N represents the total number of intent labels in the training set, M represents the total number of samples in the training set, TP represents the true positive class, FN represents the false negative class, FP represents the false positive class, and X represents the total number of samples in the training set. r Indicates that the true label in the dataset is r The number of samples that are correctly classified, Y r,l It indicates the number of samples that are misclassified as category l. i , P1, p(x) represents the real data distribution, P2, q(x) represents the generated data distribution, and n represents the total number of samples.

[0092] The hyperparameters in the WDiffusion diffusion model largely determine the quality of the generated data; therefore, the selection of hyperparameters is verified through experiments, and the indicators under different hyperparameters are shown in Tables 6 to 12 below;

[0093]

[0094] Table 6 Comparison of different learning rates

[0095] According to the results in Table 6, the diffusion model performs best when the learning rate is 1e-3. Compared with other learning rates, the parameters are the lowest, which means that the data generated at this time is most similar to the original data. Therefore, the learning rate is selected as 0.001.

[0096]

[0097] Table 7 Comparison of different gradient cumulative steps

[0098] Refers to the number of gradient accumulation steps, that is, the accumulation of gradients of multiple batches before updating model parameters. Gradient accumulation can reduce memory requirements, especially when the batch size is large, and can also improve the stability and generalization performance of the model; it is found that the indicator results are best when the number of steps is 2, so the number of gradient accumulation steps is set to 2.

[0099]

[0100] Table 8 Comparison of different iteration cycles

[0101] Regarding the iteration cycle, it was found that when the cycle exceeded 200 rounds, the evaluation indicators did not change significantly, so the number of iteration rounds was selected as 200.

[0102]

[0103] Table 9.1 Comparison of different exponential moving average decay rates

[0104] The exponential moving average decay value is used to calculate the average value of the model parameters; choosing an appropriate exponential moving average can smooth the parameter update process, reduce parameter oscillation, and thus improve the generalization performance of the model; as can be seen from Table 9.1, the various indicators decrease as the exponential moving average decay increases, until it reaches the minimum at around 0.995, so the exponential moving average decay value is selected as 0.995.

[0105]

[0106] Table 9.2 Comparison of sizes in different dimensions

[0107] The dimension size determines the dimension of the hidden layer in the model. A larger dimension can increase the representation ability of the model, but it will also increase the complexity and computational cost of the model. It is found that when the dimension exceeds 64, the reduction of various indicators is not obvious, so the dimension size is selected as 64.

[0108]

[0109] Table 10 Comparison of multiples of different dimensions

[0110] The dimension multiple can control the dimensionality change between different layers of the model, thereby affecting the complexity and representation ability of the model; a larger dimension multiple can increase the capacity of the model, but it will also increase the number of parameters and computational cost; a smaller dimension multiple can reduce the capacity of the model, but may limit the representation ability of the model.

[0111] It can be seen from Table 10 that the model effect is best when the dimension multiple is (1, 2, 4).

[0112]

[0113] Table 11 Comparison of different batch sizes

[0114] The batch size indicates the number of samples in each training batch. A larger batch size can speed up training but will also take up more memory. At the same time, a larger batch size may cause the model to converge to a local minimum rather than a global minimum. It can be seen from Table 11 that when the batch size is 32, all indicators are optimal, so the batch size is selected as 32.

[0115]

[0116] Table 12 Comparison of different beta scheduling methods

[0117] Beta scheduling is used to define the scheduling method of beta values ​​in the diffusion model. In the diffusion model, the beta value is used to control the intensity of the noise, thereby affecting the quality and diversity of the generated samples; beta_schedule = 'cosine' means using the cosine scheduling method to adjust the beta value;

[0118] In cosine scheduling, the beta value will gradually decrease from a larger initial value to a smaller final value, so that the model pays more attention to noise in the early stage of training and pays more attention to the accuracy of generated samples in the later stage of training. According to the conclusions of the above comparative experiments, the selected diffusion model hyperparameter configuration is shown in Table 13:

[0119]

[0120]

[0121] Table 13 Hyperparameters in the diffusion model

[0122] In order to verify the impact of parameter configuration on the model, we selected optimizers, learning rates, and weight decay regularization parameters for comparative experiments, and analyzed the convergence of the tests; Figure 5 - Figure 7 The effect of changing parameter configuration on model recognition accuracy is demonstrated.

[0123] Figure 5The results show that different optimizers have a significant impact on the recognition performance of MLSTM-FCN. It is found that optimizers such as SGD, Adagrad and Adadelta, which are more sensitive to learning rate parameters, fail to show ideal convergence effects. However, optimizers such as Adam, RMSprop and NAdam, which have the ability to adaptively adjust the learning rate, can make the convergence of the model smoother, especially the Adam optimizer, which has a faster convergence speed and recognition performance than RMSprop and NAdam. In other words, these optimizers that can adaptively adjust the learning rate are superior in the task of aerial target intention recognition and can achieve better and faster convergence effects.

[0124] according to Figure 6 We found that the adjustment of the learning rate is crucial. If the learning rate is set too high or too low, it will affect the convergence quality and accuracy of the model. The optimal learning rate should be around 1e-4, which can achieve the highest recognition efficiency while avoiding the problem of local optimal solution. Further observation Figure 7 From the displayed information, it is found that the influence of the regularization parameter weight_decay on the recognition accuracy shows a more complex trend. The accuracy starts to increase with the increase of the regularization parameter, and then begins to decrease;

[0125] Among them, when the regularization parameter is about 1e-4, the recognition result is optimal; therefore, it is very important to choose a suitable regularization parameter, which can effectively suppress overfitting and improve the accuracy of target recognition;

[0126] Finally, our selected optimal parameter configurations are listed in Exhibit 14.

[0127] Table 14 Hyperparameters in GRU

[0128]

[0129] The WDiffusion-GRU model has been trained and tested, and the experimental results are as follows Figure 8 As shown in the figure, with the iteration of the cycle, the model accuracy gradually increases, and the loss value gradually decreases. After 75 iterations, the model has entered a convergence state. Specifically, the accuracy of the training set can reach up to 99.99%, which means that the model performs extremely well in the training set; as for the test set, the accuracy is maintained at an average of 97%, and the highest can reach 97.09%, which fully demonstrates that the model has good generalization ability and can effectively identify the intention of aerial targets on new data sets.

[0130] In order to deeply explore the correlation between various intentions, this study selected the confusion matrix as an analysis tool. Fig. 9In the presented confusion matrix, the diagonal line represents the number of samples correctly identified by the model. Obviously, the model has shown high accuracy for all types of intentions, especially for the "retreat" intention, the recognition accuracy has reached 98.43%. However, in the confusion matrix, we can also see the misidentification between the attack intention and the feint intention. After analysis, it was found that this is mainly due to the high similarity of the combat intention characteristics of the air targets corresponding to these two intentions. In addition, the characteristics of the feint intention are highly deceptive, and it is difficult for the network to learn the obvious difference between the two intentions, resulting in misidentification. These findings are in line with the actual situation and also point out the direction for further optimization of the model.

[0131] In order to verify the recognition performance of WDiffusion-GRU, we selected the BiGRU-Attention-based battlefield enemy target tactical intention recognition model proposed in 2021; the BiGRU-TCN-based aerial target intention recognition model proposed in 2021; the BiLSTM-Attention aerial target combat intention recognition model proposed in 2023; the hierarchical aggregation model for identifying combat intentions proposed in 2023 and the BiLSTM-TCN proposed in 2023, a total of five time series models for comparative experiments; the experiments were carried out in the same environment and dataset.

[0132] according to Fig.10 From the results, it can be seen that the intent recognition method based on WDiffusion-GRU is significantly better than other methods. It not only has a fast convergence speed, stable recognition effect, but also a higher accuracy. Compared with BiGRU and BiLSTM that incorporate temporal convolutional networks, the accuracy of WDiffusion-GRU has increased by 3.62% and 3.06% respectively; compared with BiGRU and BiLSTM that add attention mechanisms to recurrent neural networks, the accuracy of WDiffusion-GRU has increased by 1.21% and 1.07% respectively; compared with the hierarchical aggregation model, the accuracy of WDiffusion-GRU has increased by more than 2.20%; this shows that under sufficient data training, WDiffusion-GRU is more effective and robust in extracting and fusing temporal features; in comparative experiments, the convergence speed of WDiffusion-GRU is significantly better than other temporal methods, indicating that lightweight GRU is more efficient in extracting features from temporal data.

[0133] In addition, WDiffusion-GRU not only has superior recognition performance, but also effectively avoids the gradient vanishing and gradient exploding problems that may occur when complex LSTM and GRU are stacked with deeper layers.

[0134] Table 15 Performance comparison of different timing methods

[0135]

[0136]

[0137] Deep features can be extracted by stacking recurrent neural network units, but stacked recurrent neural networks are limited by serial structures and have the disadvantages of large number of parameters and slow recognition speed. The WDiffusion-GRU proposed in this paper can effectively extract local and global temporal information under sufficient data training. Compared with the traditional method of increasing model complexity, WDiffusion-GRU effectively reduces the number of parameters and improves recognition speed.

[0138] As shown in Table 15, the number of parameters of the unidirectional network is small, and the number of parameters of the GRU is significantly smaller than that of other networks; the accuracy of the recurrent neural networks BiGRU and BiLSTM with the attention mechanism is higher than that of the BiGRU and BiLSTM with the temporal convolutional network, indicating that the attention mechanism can more effectively extract features in the intent recognition model.

[0139] However, the recognition performance of the bidirectional recurrent neural network with attention mechanism is worse than that of WDiffusion-GRU, because the bidirectional recurrent neural network is limited by the serial structure, and it is easy to experience memory decay when processing long sequence data, lose important sequence information, and have a slow recognition speed. The WDiffusion-GRU proposed in this paper can effectively improve the local and global multi-level features extracted by GRU by expanding the original data, achieving a higher recognition accuracy, and also has obvious advantages in recognition speed compared with other time series methods.

[0140] Table 16 Comparison of evaluation indicators of different timing methods (%)

[0141]

[0142]

[0143] In Table 16, ①, ②, ③, ④, ⑤, and ⑥ represent WDiffusion-GRU, BiGRU-Attention, BiGRU-TCN, BiLSTM-Attention, Hierarchical Aggregation model, and BiLSTM-TCN, respectively. It can be seen from the table that the WDiffusion-GRU model outperforms other models in the intent recognition task. For the most difficult feint and attack intentions, WDiffusion-GRU achieved the best performance in all four evaluation indicators. In the penetration intention recognition task, two evaluation indicators achieved the best performance. For the retreat, reconnaissance, interference, and alert recognition tasks, although the indicators did not reach the best, they also performed well.

[0144] In addition, the average recognition accuracy of models ②, ③, ④, ⑤, and ⑥ for interference intent is close to 100%, but the accuracy for other intents is relatively low. This is due to the imbalance of the data set during the training process, which causes the model to tend to predict the majority category and has poor prediction performance for the minority category.

[0145] In order to verify the effects of different generation methods, we compared the WDiffusion time series data generation model proposed in this paper and the GAN adversarial generation network proposed in 2014; the C_GAN conditional adversarial generation network proposed in 2014; the AC_GAN adversarial generation network with auxiliary classifier proposed in 2017; the DC_GAN deep convolutional generation adversarial network proposed in 2015 and the W_GAN proposed in 2017, a total of five generation models, and conducted comparative experiments under the same original data conditions.

[0146] We use visualization to show the detailed distribution patterns of each category in the synthetic data; the synthetic data constructed by the WDiffusion method and different distance generation strategies have t-SNE plots on the target intent recognition dataset. Fig. 9 As shown in the figure, different colored point sets represent the categories of various intentions: purple point sets represent retreat intentions, yellow point sets represent reconnaissance intentions, cyan point sets represent interference intentions, blue point sets represent attack intentions; green point sets represent feint intentions, black point sets represent penetration intentions, and red point sets represent alert intentions. Fig. 9From the observation, we can find that the category distribution of synthetic data generated by WDiffusion is better than that of GAN-based methods. In the t-SNE diagram of WDiffusion, data points with the same intent have tighter clustering, while data points with different intents are distributed at locations with a large distance from each other; this shows that WDiffusion has successfully learned the hidden temporal features in the real data, making the generated synthetic data with different intents more distinguishable. WDiffusion is able to generate multi-category, multi-variable time series data. In contrast, t-SNE diagrams using other methods often appear as a mixture of multiple color point sets. This shows that other methods cannot correctly learn the temporal features of data of different categories, resulting in unclear category features of the generated data and failing to meet the needs of multi-category generation. In addition, we also observed that the blue, green, and black point sets representing attack, feint, and penetration are close to each other in multiple t-SNE diagrams and are difficult to distinguish. This shows that the temporal features of the three elements of attack, feint, and penetration in the intent data are relatively similar, which also leads to a high similarity in the generated synthetic data.

[0147] Table 17 Performance comparison of different generation methods

[0148]

[0149]

[0150] As can be seen from Table 17, WDiffusion has the best effect on generating multi-classification and multi-variable time series data, and all four indicators have reached the optimal level; AC_GAN has achieved relatively good results due to the introduction of auxiliary classifiers and the increase of category information; while for GAN and other variants, the results obtained are not ideal.

[0151] In order to verify the improvement effect of the model robustness, we added white noise with different signal-to-noise ratios to the data set; the signal-to-noise ratio range is {-10, -5, 0, 5, 10, 20} dB; the specific formula of the signal-to-noise ratio is as follows:

[0152]

[0153]

[0154]

[0155] Among them, x is the original signal, n is the noise signal, N is the length of the original signal, and noise is the noise signal calculated by the signal-to-noise ratio.

[0156] Table 18 Recognition accuracy of robustness comparison experiment under different signal-to-noise ratio conditions (%)

[0157]

[0158] As shown in Table 18, as the signal-to-noise ratio increases, the accuracy also increases. Among the seven data sets with different signal-to-noise ratios, the WDiffusion-GRU model achieved the best performance in five data sets. This is because after the data set is diffused, it can cover more situations and changes, enabling the model to learn a wider range of features and patterns. When faced with noisy data, the model can better identify and offset the noise, thereby greatly improving robustness.

[0159] The embodiments of the present invention are provided for the purpose of illustration and description. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations of the present invention. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A lightweight intention recognition method based on a diffusion model, characterized in that: The following steps are involved: S1. Battlefield environment recognition, extraction, comparative analysis, association and reasoning of key features of the aerial battlefield environment; S2, time series feature collection, obtaining time series feature data; S3, data preprocessing and input into the GRU module, the GRU module is responsible for learning the long-term trend of time series data and extracting features. The GRU module network is trained on the generated data set to learn the long-term change trend and internal features of the air combat intention time series data; S301, inputting the original data set with unbalanced samples and small data volume into the WDiffusion diffusion model to obtain a data set with sufficient training samples and balanced categories; S302, processing the time series data containing different air combat intention features through two GRU layers, with the number of neurons in each layer being 64; S303, the output result of the GRU layer is connected through a fully connected layer to obtain an output value h; S304, applying a softmax function to perform multi-category classification, wherein the number of softmax layers is equal to the number of output layers, that is, the total number of categories of air combat intention labels; y = softmax(wh+b); Where y is the intention result predicted by the model, w is the weight matrix to be learned, and b is the bias vector.

2. A lightweight intention recognition method based on a diffusion model as claimed in claim 1, characterized in that: S1 also includes the following steps: S101. For numerical features, since the units are different, in order to eliminate their influence, Min-Max normalization is used to process and map the data to the range of [0,1]. For the i-th numerical data X i =(x i1 ,x i2 ,…,x ik ,…,x in )(i=1,2,…,7), n is the number of data of the i-th feature; Where maxX i Represents a feature X i The maximum value, minX i Represents a feature X i The minimum value of x′ ik is the original data x ik The result of mapping to the interval [0,1]; S102. Encoding non-numeric data can convert it into numerical data to facilitate processing and analysis in the model; S103, construct a combat intent feature set, label each sample, and use a deep learning network for training; thereby establishing a mapping relationship between the combat intent type space set and the time series feature data set; the mapping function can be expressed as: Where Ω is the combat intention output by the model, θ K is the time series feature set consisting of K consecutive moments from t1 to tk, t i Characteristics of the moment.

3. A lightweight intention recognition method based on a diffusion model as claimed in claim 2, characterized in that: The WDiffusion diffusion model is a generative model that uses variational inference to train a parameterized Markov chain, gradually transforming the latent variables x from a simple distribution. T Generate data x0 in complex distribution; where T represents the total diffusion step length; let is the result of each diffusion time step t∈{1,2,…,T}, x T and x0 have the same dimension d. The diffusion model includes the diffusion process and the reverse process. θ where θ represents the learnable components of the conditional noise predictor and low-resolution encoder in the diffusion model; the posterior q(x1,…,x T |x0) diffusion process transforms the distribution q(x0) into a latent variable q(x T ) and fixed as a Markov chain, which is based on the variance table β1,…,β T Gradually add Gaussian noise ε to the data; In the formula, β t is a small positive constant, which is regarded as a constant hyperparameter; let α t :=1-β t , The diffusion process allows sampling x at any time step t t , which is in closed form: Further reparameterization is: The reverse process transforms the latent variable distribution p θ (x T ) is transformed into a data distribution p parameterized by θ θ (x0), which is formed by p(x T )=N(x T ; 0,I) is the definition of the Markov chain with Gaussian transition at the beginning; p θ (x t-1 ∣x t ):=N(x t-1 ;μ θ (x t ,t),σ θ (x t ,t) 2 I) (8); During the training phase, the variational lower bound (ELBO) on the negative log-likelihood is maximized, and KL divergence and variance reduction are introduced; Transformation requires p θ (x t-1 ∣x t ) is directly compared with its corresponding diffusion process posteriors; set up You can get: Equations (5), (8) and (10) guarantee all KL divergences in the equations, and equation (9) is a comparison between Gaussian functions; When t>1, let C is a constant, we get: The training process minimizes a variant of ELBO with x0 and t as input: In the formula, ε θ For the noise prediction module; sample an x T ~N(x T ; 0,I), then sample x according to formula (8) t-1 ~p θ (x t-1 ∣x t );in: The Wasserstein distance is defined as follows: In the formula, Π(P r ,P g ) is P r and P g The set of all possible joint distributions formed by combining; for each possible joint distribution γ, sample the real sample x and generate the sample y from (x, y) ~ γ, and calculate the distance between the two samples ||xy||; calculate the expected value E of the distance between the two samples under all possible joint distributions (x,y)~γ [||xy||]; Wasserstein distance is defined as the lower bound of the expected value All possible joint distributions of .

4. A lightweight intention recognition method based on a diffusion model as claimed in claim 3, characterized in that: The GRU module has two inputs, namely the output state h of the previous moment t-1 and the input sequence value x at the current moment t , the output is the current state h t ; It resets the gate r t and update gate z t To update the state of the model; reset the gate to control the degree to which historical state information is forgotten, so that the network can discard unimportant information; update the gate z t Control the proportion of the previous state information brought into the current state to help the network remember long-term information; The calculation formula is as follows: σ represents the sigmoid activation function, which is used to convert the intermediate state to the range of [0,1]; h t-1 and h t Respectively represent the output states at time t-1 and time t; x t Represents the input sequence value at time t; is a candidate output state; W r , W z , U r , U z and Respectively represent the weight coefficient matrices corresponding to each part; tanh represents the hyperbolic tangent function; "⊙" represents the Hadamard product of the matrix (element-by-element multiplication).