A stable target intention recognition method based on sample reweighting and Attention-Bi-GRU

Abstract

The present invention discloses a method for stable target intention recognition based on sample reweighting and Attention-Bi-GRU, comprising the following steps: Step 1 preprocesses target attribute data detected by sensors; Step 2 inputs the preprocessed target data into Attention-Bi-GRU to extract deep feature information; Step 3 inputs the deep feature information obtained in Step 2 into a sample reweighting module to obtain sample weights, and inputs the classification function layer, and calculates the sample weighted cross entropy loss; Step 4 iteratively trains a complete network model based on the loss function of Step 3; Step 5 preprocesses the target attribute data obtained in the actual environment according to Step 1, and then inputs the trained network model for testing to obtain the target intention recognition result. The present invention aims at the problem that the existing intention recognition method ignores the covariate shift between the measured data and the training data distribution, and considers the multidimensionality and temporal nature of the target attribute data, uses an attention mechanism to capture the moment information difference of the target sample, and adopts a bidirectional gating unit to mine the deep feature information of the target attribute data; In addition, a sample reweighting module is constructed to learn sample weights, weaken the distribution deviation of the training set and the measured data, so as to effectively perform target intention recognition in the actual environment, which is more practical and meaningful.

Description

Technical Field

[0001] The present invention belongs to the field of target intention recognition, and in particular relates to a target intention stable recognition method based on sample reweighting and a bidirectional gating unit. Background Art

[0002] With the increasing development of information technology, a large number of reconnaissance and sensor equipment are used on the battlefield, which greatly improves the ability to collect battlefield data. Identifying the enemy's target intention based on the situational characteristics related to the combat scenario is one of the keys to situational cognition and is also the premise and basis for intelligent decision-making. The commonly used intention recognition strategy is to perform inference recognition based on the target attribute data detected by the sensor, and then quickly and accurately identify the target intention from the rapidly changing battlefield situation. However, due to the multidimensionality of the target attribute data and the complex uncertainty of the battlefield environment, this strategy is a typical complex combination problem.

[0003] To solve this problem, several methods have been proposed, including template matching, expert systems, Bayesian networks, and deep learning. The template construction of the template matching method relies on the prior knowledge of domain experts, and it is difficult to ensure the objectivity and reliability of the template. Although the expert system has strong knowledge representation and reasoning capabilities, it has problems with insufficient fault tolerance and learning ability. Bayesian networks have attracted widespread attention due to their powerful causal probability reasoning capabilities and can be used to solve the problem of uncertainty reasoning of intentions, but it is difficult to determine the prior probability and conditional probability of each node event. With the rapid development of deep learning technology, the target intention recognition method based on deep learning has gradually become the mainstream, which to a certain extent avoids the shortcomings of the above methods. However, the deep learning algorithm is to learn the inherent laws and representation levels of sample data, and obtain feature information that is very helpful for target intention recognition during the learning process. This type of method often assumes that the distribution of the training set and the test set obeys independent and identical distribution, but this assumption is difficult to meet in the actual environment, that is, the distribution of the training set and the test set is difficult to guarantee the same distribution, which leads to a decrease in the generalization performance of the algorithm, making this type of method unsuitable in the actual battlefield environment.

[0004] Therefore, in order to solve the problem that existing intent recognition methods ignore the covariate shift between the measured data and the training data distribution, the present invention proposes a target intent intelligent recognition model based on sample reweighting and attention bidirectional gated unit (Attention-Bidirectional Gated Recurrent Unit, Attention-Bi-GRU) taking into account the multidimensionality and temporal nature of the target attribute data. The model performs self-attention weighting and de-weighting of sample features on the target information to achieve intelligent and stable recognition of target intent. Summary of the invention

[0005] The technical problem to be solved by the present invention is: how to realize the intelligent and stable recognition of target intention when the distribution of measured data and training data are inconsistent. Using this method to realize target intention recognition has important military value.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is a method for stable identification of target intention based on sample reweighting and bidirectional gating unit, which is characterized by comprising the following steps:

[0007] Step 1: preprocessing the target attribute data within a period of time t detected by the sensor;

[0008] Step 101: dividing the target attribute data into numerical attribute data and non-numerical attribute data, and performing normalization standard processing on them;

[0009] Step 102: performing intent encoding on the target intent space;

[0010] Step 2: Input the target attribute data within a period of time t after preprocessing into Attention-Bi-GRU to extract deep feature information;

[0011] Step 201: input the target attribute data normalized in step 101 into the attention mechanism module for moment weighting;

[0012] Step 202: Input the weighted target attribute data at step 201 into the Bi-GRU module to extract deep feature information;

[0013] Step 3: Input the deep feature information obtained in step 2 into the sample reweighting module to obtain the sample weight, input it into the Softmax classification layer, and calculate the sample weighted cross entropy loss;

[0014] Step 301: Use Random Fourier Feature (RFF) to map the deep feature information extracted by the network to the random feature space:

[0015]

[0016] Wherein, h represents the random feature space mapping function, and z represents the deep feature extracted in step 202;

[0017] Step 302: define the target sample weight as β, and perform a weighted independence test on the feature information of the random feature space based on the Frobenius norm: when When it approaches 0, the two variables tend to be independent; is the weighted cross-covariance between random features,

[0018]

[0019]

[0020]

[0021] Among them, n is a random feature The dimensions of p and q represent random features respectively. Dimension number of ;

[0022] Step 303: Combine the sample weights from step 302 with the cross entropy obtained from the Softmax classification layer to calculate the sample weighted cross entropy loss:

[0023]

[0024] Among them, f(·) is the deep feature representation function, g(·) represents the classification function, and I i represents the intent label encoding of the i-th target sample, β i represents the weight of the i-th target sample, L CE (·,·) represents the cross entropy loss function;

[0025] Step 4: Based on the loss function in step 3, iteratively train the complete network model. The iterative training process is as follows:

[0026]

[0027] in, k represents the training round;

[0028] Step 5: Preprocess the target attribute data obtained in the actual environment according to step 1, and then input it into the trained network model for testing to obtain the target intent recognition result.

[0029] The attention mechanism module described in step 2 is as follows: For a certain target attribute data X i ,

[0030]

[0031]

[0032] Among them, i represents the target number, X i is the attribute data of the i-th target, W Q ,W K ,W V are the weights of the query-key-value in the attention mechanism, Q i ,K i ,V iare the weighted query-key-value of the i-th target attribute data, d K Indicates the number of columns of the Q,K matrix. QK T The number of rows and columns in the obtained matrix is ​​the time length t in the target attribute data. Use softmax to calculate the attention coefficient of each moment to other moments, and then add it to the matrix V i Multiply them together to get the weighted input.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The present invention has simple steps, reasonable design, and is easy to implement and use;

[0035] 2. The present invention uses the attention mechanism to capture the moment information difference of the target samples, while considering the temporal and multidimensional nature of the target attribute data, and adopts the bidirectional gating unit to mine the deep feature information of the target attribute data;

[0036] 3. The present invention constructs a sample reweighting module to learn sample weights, weakening the distribution deviation between the training set and the measured data, thereby effectively identifying target intentions in a practical environment, which is more practical and of practical significance.

[0037] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flow chart of the method of the present invention

[0039] Figure 2 This is the network model training flow chart of the present invention

[0040] Figure 3 The effect of the first embodiment of the present invention is described as follows DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0042] Embodiment 1

[0043] Due to the complexity of the battlefield environment, there is inevitably a distribution deviation between the target attribute data in the actual environment and the target attribute data in the model training set. Therefore, it is necessary to improve the effectiveness of the algorithm model in the actual environment to make intent recognition more practical. To this end, this embodiment provides a target intent recognition method under the distribution deviation of training set and test data. Figure 1is a flow chart of the method of the present invention, Figure 2 : is a flow chart of the network model training of the present invention. As shown in the figure, the target intention recognition method of this embodiment includes:

[0044] Step 1: Obtain the target attribute dataset and corresponding intent labels.

[0045] The target attributes included in the data set include: X coordinate, Y coordinate, altitude, speed, heading angle, pitch angle, distance, pulse width, radar frequency, communication status, radar status. The target intention labels include: patrol, electronic reconnaissance, electronic interference, attack, strike, early warning detection, that is, the target intention space.

[0046] Step 2: Preprocessing of target attribute data.

[0047] like Figure 2 As shown, the preprocessing of target attribute data includes attribute data normalization processing and intention space encoding. Specifically, the target attribute data normalization processing is:

[0048]

[0049] in, represents the normalized value of the j-th dimension attribute of the ith sample at time t, represents the original value of the j-th dimension attribute of the i-th sample at time t, maxx i,j represents the maximum value of the j-th dimension attribute of the i-th sample at all times, minx i,j Represents the minimum value of the j-th dimension attribute of the i-th sample at all times.

[0050] The target intention space is coded as follows: 0-patrol, 1-electronic reconnaissance, 2-electronic jamming, 3-attack, 4-strike, 5-early warning detection.

[0051] Step 3: The preprocessed target data is input into the network model training to obtain a fully trained network.

[0052] like Figure 2 As shown, the network model training steps include:

[0053] S1: Input the target attribute data within a period of time t after preprocessing into Attention-Bi-GRU to extract deep feature information; specifically, S1 includes:

[0054] S11: Input the normalized target attribute data into the temporal attention mechanism layer for moment weighting;

[0055] In this embodiment, since the target attribute data is information in multiple consecutive moments, but not all moments should be treated equally, different weights need to be assigned to different moments. This embodiment weights different moments by practicing the attention mechanism layer.

[0056] S12: Input the moment-weighted target attribute data into the bidirectional gated unit layer to extract deep feature information.

[0057] S2: Input the deep feature information obtained by S1 into the sample reweighting layer to obtain the sample weight, input it into the classification function layer, and calculate the sample weighted cross entropy loss; specifically, S2 includes:

[0058] S21: Use random Fourier feature mapping to map the deep feature information obtained in S1 to the random feature space:

[0059]

[0060] Among them, h represents the random feature space mapping function, and z represents the deep features extracted by S12.

[0061] In this embodiment, since the deep feature information obtained in step S12 is a continuous value and has complex nonlinear dependencies, it is impossible to directly calculate the sample weights. Therefore, random Fourier features are used to map the features extracted by the network to the random feature space, converting the nonlinear dependencies into linear ones to reduce the computational complexity.

[0062] S22: Define the target sample weight as β, and perform a weighted independence test on the feature information of the random feature space based on the Frobenius norm: when When it approaches 0, the two variables tend to be independent; is the weighted cross-covariance between random features,

[0063]

[0064]

[0065]

[0066] Among them, n is a random feature The dimensions of p and q represent random features respectively. Dimension number of ;

[0067] In this embodiment, in order to eliminate the influence of the dependency relationship between features on the prediction performance and make the features as independent as possible, the Frobenius norm is used to perform independence test on the features as a criterion for random feature decorrelation.

[0068] S23: Combine the sample weights of step S22 with the cross entropy obtained by the classification function layer to calculate the sample weighted cross entropy loss:

[0069]

[0070] Among them, f(·) is the deep feature representation function, g(·) represents the classification function, and I i represents the intent label encoding of the i-th target sample, β i represents the weight of the i-th target sample, L CE (·,·) represents the cross entropy loss function;

[0071] S24: Iterative training network model:

[0072]

[0073] in, k represents the number of training rounds.

[0074] In this embodiment, by adjusting parameters and training the model, the network parameters finally determined are: the time series length is 500 frames, the number of network hidden layers is 3, the number of nodes in each hidden layer is 64, the optimizer is SGD, the learning rate is 0.001, the batch size is 256, and the number of iterations is 200.

[0075] Step 4: Use the trained network model to perform intent recognition on the measured target attribute data to obtain the target intent recognition result. Specifically, it includes the following steps:

[0076] S3: normalize the actual target attribute data, and the specific normalization is consistent with the target attribute data set;

[0077] S4: Input the normalized data into the trained network model for intent recognition to obtain the final recognition result.

[0078] Embodiment 2

[0079] This embodiment illustrates the effect of the method of embodiment 1 through simulation experiments.

[0080] Taking aerial targets as the research object, the target attribute data set was obtained through the simulation platform. Then the sample labels were corrected by experts. Finally, each target was sampled and 500 frames of time series data were extracted with equal length (each frame contains 11-dimensional attribute information). The final data set sample size is 3415, and the training set and test set are divided into 8:2, that is, the training set and test set contain 2731 samples and 684 samples respectively.

[0081] The distribution of the test set data is changed based on the inverse transformation and Box-Cox method, and is used as the measured target attribute data.

[0082] By using the method of Example 1, the target intention recognition results under different measured data are obtained, and compared with the deep neural network DBP, stacked autoencoder SAE, long short-term memory network LSTM, and bidirectional long short-term memory network Bi-LSTM methods. Figure 3 As shown. Figure 3 It can be found that the method proposed in the present invention has achieved the best results on multiple measured data sets, with an average recognition accuracy of 93.02%. The method of the present invention is 3.87% higher than the DBP method, 14.69% higher than the SAE method, 22.41% higher than the LSTM method, and 7.38% higher than the Bi-LSTM method. In addition, the recognition accuracy of the method of the present invention in the four measured data sets is more than 85%. The above results verify the effectiveness of the method of the present invention.

[0083] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A method for stable target intention recognition based on sample reweighting and Attention-Bi-GRU, characterized in that: The following steps are involved: Step 1: preprocessing the target attribute data within a period of time t detected by the sensor; Step 101: dividing the target attribute data into numerical attribute data and non-numerical attribute data, and performing normalization standard processing on them; Step 102: performing intent encoding on the target intent space; Step 2: Input the target attribute data within a period of time t after preprocessing into Attention-Bi-GRU to extract deep feature information; Step 201: input the target attribute data normalized in step 101 into the attention mechanism module for moment weighting; Step 202: Input the weighted target attribute data at step 201 into the Bi-GRU module to extract deep feature information; Step 3: Input the deep feature information obtained in step 2 into the sample reweighting module to obtain the sample weight, input it into the Softmax classification layer, and calculate the sample weighted cross entropy loss; Step 301: Use Random Fourier Feature (RFF) to map the deep feature information extracted by the network to the random feature space: Wherein, h represents the random feature space mapping function, and z represents the deep feature extracted in step 202; Step 302: define the target sample weight as β, and perform a weighted independence test on the feature information of the random feature space based on the Frobenius norm: when When it approaches 0, the two variables tend to be independent; is the weighted cross-covariance between random features, Among them, n is a random feature The dimensions of p and q represent random features respectively. Dimension number of ; Step 303: Combine the sample weights from step 302 with the cross entropy obtained from the Softmax classification layer to calculate the sample weighted cross entropy loss: Among them, f(·) is the deep feature representation function, g(·) represents the classification function, and I i represents the intent label encoding of the i-th target sample, β i represents the weight of the i-th target sample, L CE (·,·) represents the cross entropy loss function; Step 4: Based on the loss function in step 3, iteratively train the complete network model. The iterative training process is as follows: in, k represents the training round; Step 5: Preprocess the target attribute data obtained in the actual environment according to step 1, and then input it into the trained network model for testing to obtain the target intent recognition result.

2. The attention mechanism module described in step 2 is as follows: For a certain target attribute data X i , in, i represents the target number, X i is the attribute data of the i-th target, W Q ,W K ,W V are the weights of the query-key-value in the attention mechanism, Q i ,K i ,V i are the weighted query-key-value of the i-th target attribute data, d K Indicates the number of columns of the Q,K matrix; QK T The number of rows and columns in the obtained matrix is ​​the time length t in the target attribute data; use softmax to calculate the attention coefficient of each moment to other moments, and then add it to the matrix V i Multiply them together to get the weighted input.

Images