End-to-end high-dynamic target recognition method based on self-attention mechanism

Abstract

The invention discloses an end-to-end high-dynamic target recognition method based on a self-attention mechanism, and relates to the field of high-dynamic target recognition, in particular to a high-dynamic target recognition algorithm which can effectively extract time sequence change features and is high in reasoning speed, and robust recognition performance can still be achieved in a multi-target and high-noise complex scene. According to the technical scheme, the method comprises the following steps of 1, obtaining radar cross section data of a high-dynamic target, preprocessing the radar cross section data, and dividing the radar cross section data into time sequence fragments of different scales to serve as multi-scale samples; 2, performing deep feature extraction on the multi-scale time series data to obtain feature codes of different scales of the target RCS, and recording the feature codes as deep RCS features; and step 3, classifying the deep RCS features to obtain a target recognition result.

Description

Technical Field

[0001] The present invention relates to the field of high-dynamic target recognition, and in particular to an end-to-end recognition method based on a self-attention mechanism. Background Art

[0002] Correctly and quickly identifying effective targets and decoys of non-cooperative parties is of great significance for effectively intercepting effective targets and ensuring national security. The attack process of high-dynamic targets can usually be divided into boost phase, mid-stage, and reentry phase. The duration of the mid-stage can reach 80% to 90% of the entire flight, which is the longest and most easily identifiable phase. The radar cross-section RCS data of the target obtained by radar can directly reflect the physical characteristics of the target, including shape, size, material and other information, and can also provide rich target attitude information, including azimuth and pitch angle. Therefore, the RCS data of mid-stage targets are mostly used for high-dynamic target recognition. At present, high-dynamic target recognition is mainly divided into two categories: machine learning methods based on feature extraction and deep learning methods based on data-driven. Among them, the former uses artificially designed features to extract the classifiable features of the target, and then uses a machine learning classifier to identify the target; the latter constructs a neural network model and trains it on the electromagnetic data set of high-dynamic targets to directly obtain the target type through the target electromagnetic data.

[0003] Although these methods can achieve high-dynamic target classification to a certain extent, their recognition performance will decline when faced with complex scenes with multiple targets, high noise, and high nonlinearity. This is because the machine learning methods based on feature extraction rely on the original data distribution and expert experience when designing features that can effectively classify. When the original data is highly nonlinear or contains a lot of noise, complex features need to be designed to complete the classification, which brings great difficulties to the designer. In addition, since the method based on feature extraction needs to calculate the target features of the input data first, the reasoning speed still has room for improvement. In scenarios where the classification results are required to be real-time, its application is limited. As deep learning methods have been well applied in many fields, how to apply neural networks to the field of high-dynamic target classification has attracted the attention of scholars. Convolutional neural network (CNN) has been used for the problem of high-dynamic target classification based on RCS data. The CNN network structure is used to automatically extract high-dimensional features, without the need to manually design features based on the original data. The trained network can directly output the target category after inputting the target RCS data. Although the CNN network structure is introduced to improve the performance of the high-dynamic target recognition algorithm, the convolutional network does not consider the sequential nature of time series data. Therefore, when processing RCS time series, it cannot effectively capture the time series relationship contained in the data, and performs poorly when processing RCS data with strong time correlation. Therefore, it is urgent to design a high-dynamic target recognition algorithm that can effectively extract time series change features and has high inference speed, and can still achieve robust recognition performance in complex scenes with multiple targets and high noise. Summary of the invention

[0004] In view of this, the present invention provides an end-to-end high-dynamic target recognition method based on the self-attention mechanism, which can effectively extract time series change characteristics and a high-dynamic target recognition algorithm with high inference speed, and can still achieve robust recognition performance in complex scenes with multiple targets and high noise.

[0005] To achieve the above object, the technical solution of the present invention comprises the following steps:

[0006] Step 1: Obtain radar cross-section data of high-dynamic targets, perform preprocessing, and divide them into time series segments of different scales as multi-scale samples.

[0007] Step 2: Perform deep feature extraction on multi-scale time series data to obtain feature codes of target RCS at different scales, which are recorded as deep RCS features.

[0008] Step 3: Classify the deep RCS features to obtain target recognition results.

[0009] Furthermore, preprocessing includes removing outliers and normalizing data.

[0010] Furthermore, the time series segments are divided into different scales as multi-scale samples. The specific method is as follows:

[0011] First, determine the time windows of different scales, divide the preprocessed data by sliding windows, set the sliding window step to half the window length, and obtain a continuous data segment in each time window. After sliding window sampling, multi-scale samples are obtained.

[0012] Furthermore, step 2 is specifically as follows:

[0013] For the multi-scale samples, the longest sequence length of the multi-scale samples is used as a reference. For sequences shorter than the reference, a padding value is added at the end thereof to align them with the longest sequence length of the multi-scale samples, thereby obtaining an input sequence.

[0014] At the same time, a mask vector of the same length as the input sequence is created to record the valid data length of the input sample. The mask value of 1 indicates valid input and 0 indicates the padding part.

[0015] The input sequence is input into a feature encoder; and a plurality of feature encoders are stacked for use.

[0016] The feature encoder includes a self-attention module, a residual connection module and a layer normalization module.

[0017] The self-attention module adopts a multi-head self-attention mechanism, which is implemented by parallel calculation of multiple independent self-attentions; the number of heads is set to h, and each head has an independent linear transformation matrix of query, key and value. Perform h independent self-attention calculations on the sample data to obtain h groups of attention weights head1, head2, ..., head h , Finally, all the obtained weights are concatenated into a vector and passed through a linear transformation matrix W O The final attention weight Z is obtained as follows:

[0018] Z=Concat(head1,head2,...,head h )W O

[0019] After the residual connection module obtains the attention weight Z, it performs a residual connection with the original input X, that is, O=X+Z.

[0020] The matrix O is input to the layer normalization module. The function of layer normalization is to normalize the hidden layer in the neural network to a standard normal distribution. By normalizing the output of each layer, the normalized output has zero mean and unit variance.

[0021] The final output of the feature encoder is represented as X hidden .

[0022] Furthermore, multiple feature encoders are stacked and used. Specifically, 6 feature encoders are stacked, and the output of the previous encoder is used as the input of the next encoder, so as to finally obtain the deep RCS features of the high dynamic targets.

[0023] Furthermore, the classifier recognition module includes a fully connected layer, a linear activation layer and a classification layer.

[0024] The fully connected layer converts the deep RCS features into a fixed-length vector.

[0025] The linear activation layer performs a linear transformation on the output of the fully connected layer, adjusts the output dimension to the number of categories that high dynamic targets can be classified into, and activates it using the RuLU function.

[0026] The classification layer uses the Softmax function to output the probability of each category.

[0027] Finally, the classification result is the index of the element position with the largest value in the probability vector output by the classification layer.

[0028] Beneficial effects:

[0029] The present invention provides an end-to-end high-dynamic target recognition method based on a self-attention mechanism, including three modules: input data processing, deep feature extraction, and classifier recognition; a sliding window sampling strategy for high-dynamic target RCS data is designed to obtain multi-scale time series samples from the original data, and capture the feature information of the original RCS data at different time scales. Compared with the machine learning method based on feature extraction, this method uses the performance of deep neural networks to automatically extract high-dimensional features, and directly calculates features from radar cross-section RCS data using network parameters, avoiding the dependence of artificially designed features on experience and the problem of being unable to characterize target characteristics when the data is highly nonlinear, and has a faster reasoning speed. Compared with the deep learning method based on convolutional neural networks, this method uses a self-attention mechanism to capture data dependencies in RCS time series, can better focus on the intrinsic features in different high-dynamic target RCS sequences, and exhibits better high-dynamic target recognition performance in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is an end-to-end high dynamic target recognition method framework. DETAILED DESCRIPTION

[0031] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0032] The present invention provides a high-dynamic target recognition algorithm for complex scenes, which can still achieve robust recognition performance in complex scenes with multiple targets and high noise. The specific method framework is as follows Figure 1 shown.

[0033] A high-dynamic target is a target that moves much faster than the background target or other targets in the environment. The flight speed of a high-dynamic target varies depending on the target range, ranging from 1.5 to 7 km / s.

[0034] The end-to-end high-dynamic target recognition algorithm of the present invention consists of three modules: an input data processing module, a deep feature extraction module and a classifier recognition module. First, the input data is preprocessed to improve the data quality and divided into time series segments of different scales. Secondly, deep feature extraction is performed on the multi-scale time series data to obtain feature encodings of different scales of the target RCS. Finally, the category of the target is output through the classifier. In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in combination with specific implementation cases.

[0035] Step 1: Obtain radar cross-section data of highly dynamic targets, and perform data preprocessing and multi-scale data division. Data preprocessing includes removing outliers and normalizing data. Multi-scale data division requires first determining time windows of different scales, selecting shorter windows for fast-moving highly dynamic targets, and longer windows for rotating or oscillating targets. The original RCS data is divided by means of a sliding window, and the sliding window step is set to half the window length. Each time window will intercept a continuous piece of data. The multi-scale samples obtained after sliding window sampling are conducive to the method capturing the characteristic information of the original RCS data at different time scales.

[0036] Step 2: Input the RCS sequence samples after sliding window sampling in step 1 into the deep feature extraction module. Since the length of RCS sample data is inconsistent, for shorter sequences, a padding value (usually 0) is added to the end to align it with the longest sequence length. At the same time, a mask vector of the same length as the input sequence is created to record the effective data length of the input sample. The mask value of 1 indicates valid input and 0 indicates the padding part. When calculating self-attention, the mask is used to ignore the padding part to ensure that the data input to the network is a valid RCS sampling vector. Let the RCS sequence of the input model be X, with a shape of (l sample ,1), the deep feature dimension is l hidden . sample is the sequence length of the longest RCS sample.

[0037] The feature encoder includes main network layers such as self-attention module, residual connection, and normalization. In the self-attention module, for each moment data in the sequence, by calculating its correlation with other data in the sequence (attention weight), and generating a representation that integrates global information, it can capture the dependency between different positions in the sequence and better process long time series. First, the shape of (l sample ,1) becomes the shape of (l1,l2), satisfying l1*l2=l sample , and then through three independent linear transformation matrices W Q ,W K ,W V ∈(l2,l hidden ) respectively generate Q = XW Q Query, key K = XW K , value vector V = XW V ∈(l1,l hidden ). Each linear change transforms the last dimension l2 of the input data into the dimension l of the hidden layer feature hidden , the previous dimension remains unchanged. Next, the dot product of the query and the key is calculated and scaled, and then the attention weight Z is obtained through the Sotfmax function, as follows:

[0038]

[0039] where d k The vector dimensions for the query and key.

[0040] In order to capture more levels of sequence information, a multi-head self-attention mechanism is adopted, which is implemented by parallel calculation of multiple independent self-attentions. The number of heads is set to h, and each head has independent linear transformation matrices of query, key and value Perform h independent self-attention calculations on the sample data to obtain h groups of attention weights head1, head2, ..., head h Finally, concatenate all the obtained weights into a vector Concat(head1,head2,...,head h ), and obtain the final attention weight Z through a linear transformation matrix, as follows:

[0041] MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O

[0042] in, W O It is a fully connected layer that transforms the multi-head attention concatenation result to be consistent with the input dimension.

[0043] Residual connections introduce "skip connections" to allow gradients to be transmitted more directly, thereby alleviating the common gradient vanishing and gradient exploding problems in deep networks. At the same time, residual connections allow information to be directly transmitted from the previous layer to the next layer, making it easier for the model to learn the identity mapping and helping to avoid excessive distortion or loss of information in multi-layer nonlinear transformations. After obtaining the attention weight, it is residually connected to the original input X, that is, O = X + Z.

[0044] The role of layer normalization is to normalize the hidden layer in the neural network to a standard normal distribution, so as to speed up the training and convergence. By normalizing the output of each layer, the normalized output has zero mean and unit variance. First, calculate the mean of each column of the matrix O and variance o ij is the element in the i-th row and j-th column of matrix O, and m is the number of rows in matrix O. Then the output of the normalization layer can be expressed as Where ∈ is a very small number to prevent the denominator from being zero, for example, ∈ can be set to 10 -8 or 10 -6 The normalized output is forward propagated and the ReLU function is selected for activation. The expression of the ReLU activation function is ReLU(x)=max(0,x).

[0045] The final output of a feature encoder can be expressed as X hidden =Linear(ReLU(LayerNorm(O)));

[0046] X hidden Represents the high-dimensional features learned by the network. Feature encoders are usually stacked to increase the network depth, enhance the interaction between features, capture more complex patterns and features, and significantly enhance the expressiveness of the model. This method uses 6 feature encoders in a stack, with the output of the previous encoder used as the input of the next encoder, and finally obtains the deep RCS features of highly dynamic targets.

[0047] Step 3: Classify the deep RCS features. The classifier recognition module contains a fully connected layer, a linear activation layer, and a classification layer. The fully connected layer converts the deep features into a vector of fixed length. The linear activation layer performs a linear transformation on the output of the fully connected layer, adjusts the output dimension to the number of categories that high dynamic targets can be classified into, and activates it using the RuLU function. The classification layer uses the Softmax function to output the probability of each category. The public declaration of the Softmax function is where z i is the score of category i. Finally, the classification result is the index of the element position with the largest value in the probability vector output by the classification layer.

[0048] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An end-to-end high-dynamic target recognition method based on self-attention mechanism, characterized in that: include: Step 1: Obtain radar cross-section data of high-dynamic targets, perform preprocessing, and divide them into time series segments of different scales as multi-scale samples; Step 2: Perform deep feature extraction on multi-scale time series data to obtain feature codes of target RCS at different scales, recorded as deep RCS features; Step 3: Classify the deep RCS features to obtain target recognition results.

2. The end-to-end high-dynamic target recognition method based on the self-attention mechanism as claimed in claim 1, characterized in that: The preprocessing includes deleting outliers and normalizing data.

3. The end-to-end high-dynamic target recognition method based on the self-attention mechanism as claimed in claim 1, characterized in that: The time series segments divided into different scales are used as multi-scale samples in the following specific manner: First, determine the time windows of different scales, divide the preprocessed data by sliding windows, set the sliding window step to half the window length, and obtain a continuous data segment in each time window. After sliding window sampling, multi-scale samples are obtained.

4. The end-to-end high-dynamic target recognition method based on self-attention mechanism as claimed in claim 1, characterized in that: The step 2 is specifically as follows: For the multi-scale samples, taking the longest sequence length of the multi-scale samples as a reference, for sequences shorter than the reference, adding a padding value at the end thereof to align the sequence length with the longest sequence length of the multi-scale samples, thereby obtaining an input sequence; At the same time, a mask vector of the same length as the input sequence is created to record the effective data length of the input sample. The mask value is 1 for valid input and 0 for padding. Inputting the input sequence into a feature encoder; Multiple feature encoders are stacked for use; The feature encoder includes a self-attention module, a residual connection module and a layer normalization module; The self-attention module adopts a multi-head self-attention mechanism, which is implemented by parallel calculation of multiple independent self-attentions; the number of heads is set to h, and each head has an independent linear transformation matrix of query, key and value. Perform h independent self-attention calculations on the sample data to obtain h groups of attention weights head1, head2, ..., head h , Finally, all the obtained weights are concatenated into a vector and passed through a linear transformation matrix W O The final attention weight Z is obtained as follows: Z=Concat(head1,head2,...,head h )W O After the residual connection module obtains the attention weight Z, it performs a residual connection with the original input X, that is, O = X + Z; The matrix O is input to the layer normalization module. The function of layer normalization is to normalize the hidden layer in the neural network to a standard normal distribution. By normalizing the output of each layer, the normalized output has zero mean and unit variance. The final output of the feature encoder is represented as X hidden .

5. The end-to-end high-dynamic target recognition method based on self-attention mechanism as claimed in claim 4, characterized in that: The stacking of multiple feature encoders is specifically as follows: 6 feature encoders are stacked, the output of the previous encoder is used as the input of the next encoder, and finally the deep RCS features of the high dynamic target are obtained.

6. The end-to-end high-dynamic target recognition method based on self-attention mechanism as claimed in claim 5, characterized in that: The classifier recognition module includes a fully connected layer, a linear activation layer and a classification layer; The fully connected layer converts the deep RCS features into a fixed-length vector; The linear activation layer performs a linear transformation on the output of the fully connected layer, adjusts the output dimension to the number of categories that high-dynamic targets can be classified into, and uses the RuLU function for activation; The classification layer uses the Softmax function to output the probability of each category; Finally, the classification result is the index of the element position with the largest value in the probability vector output by the classification layer.

Images